EV and Renewable Energy Integration in Residential Buildings: A Global Perspective on Deep Learning, Strategies, and Challenges

Abstract

Charging electric vehicles (EVs) and integrating renewable energy sources (RESs) are becoming key aspects of residential energy systems. However, the variability of RES generation, combined with uncontrolled EV charging, poses challenges for reliability, power quality, and supply-demand balancing within communities. The challenges only grow when considering other electrified building loads as well. Accurate forecasting of power demand and renewable generation is essential for efficient and sustainable grid operation, optimal use of RESs, and effective energy trading within communities. Deep learning (DL), including supervised, unsupervised, and reinforcement learning (RL), has emerged as a promising solution for predicting consumer demand, renewable generation, and managing energy flows in residential environments. This paper provides a comprehensive review of the development and application of these methods for forecasting and energy management in residential communities. Evaluation metrics across studies indicate that supervised learning can achieve highly accurate forecasting results, especially when integrated with unsupervised K-means clustering and data decomposition. These methods help uncover patterns and relationships within the data while reducing noise, thereby enhancing prediction accuracy. RL shows significant potential in control applications, particularly for charging strategies. Similarly to how V2G-simulators model individual EV usage and simulate large fleets to generate grid-scale predictions, RL can be applied to various aspects of EV fleet management, including vehicle dispatching, smart scheduling, and charging coordination. Traditional methods are also used across different applications and help utilities with planning. However, these methods have limitations and may not always be completely accurate. Our review suggests that integrating hybrid supervised-unsupervised learning methods with RL can significantly improve the sustainability and resilience of energy systems. This approach can improve demand and generation forecasting while enabling smart charging coordination and scheduling for scalable EV fleets integrated with building electrification measures. Furthermore, the review introduces a unifying conceptual framework that links forecasting, optimization, and policy coupling through hierarchical deep learning layers, enabling scalable coordination of EV charging, renewable generation, and building energy management. Despite methodological advances, real-world deployment of hybrid and deep learning frameworks remains constrained by data-privacy restrictions, interoperability issues, and computational demands, highlighting the need for explainable, privacy-preserving, and standardized modeling approaches. To be effective in practice, these methods require robust data acquisition, optimized forecasting and control models, and integrated consideration of transport, building, and grid domains. Furthermore, deployment must account for data privacy regulations, cybersecurity safeguards, model interpretability, and economic feasibility to ensure resilient, scalable, and socially acceptable solutions.

1. Introduction

The global EV market is continuing on its strong growth trajectory. In 2024, over 17 million electric cars were sold worldwide, making up more than 20% of all new vehicle sales. Looking ahead, electric vehicle sales in 2025 are projected to surpass 20 million globally, representing over 25% of total new car sales across the world [1]. China remains at the forefront, accounting for 60% of global EV sales. However, the rapid growth of the EV industry introduces new challenges for power systems, including transformer overloading, voltage fluctuations, and peak demand exacerbation. Addressing these issues is crucial to facilitating a sustainable and efficient transition to electric mobility.

Renewable energy technologies are becoming vital in the modern residential energy systems, supporting the transition to low-carbon, electrified communities [2]. The increasing deployment of residential solar photovoltaic (PV) systems reduces environmental impacts while providing clean electricity. Combined with heat pumps and energy storage, these technologies enhance self-consumption, demand-side flexibility, and grid stability while supporting decarbonization goals.

The increasing digitalization of prosumers, supported by big data applications and machine learning, is poised to shape the future of power systems [3]. Systematic assessments of energy management that incorporate DL and advanced charging strategies, while accounting for emerging building technologies, are essential to develop a comprehensive understanding of the current state of knowledge. This review compiles insights from relevant studies on load demand, charging behavior, and charging strategies for scalable EV fleets at both regional and global levels. It also explores theoretical and technological scenarios of hybrid energy systems in residential buildings, while highlighting areas for further research and information transfer. The literature search was conducted using major scientific databases, including Scopus, Web of Science, IEEE Xplore, ScienceDirect, MDPI, and SpringerLink, complemented by Google Scholar. Keywords and Boolean combinations related to EVs charging, Deep Learning forecasting, energy management, V2G, and renewable integration were used to ensure comprehensive coverage (Figure 1). Studies were included if they reported quantitative performance metrics or validated models applicable to residential or community-scale systems. To enhance methodological transparency and reproducibility, a structured, stepwise workflow was adopted encompassing database selection, keyword formulation, filtering for relevance (2014–2025), removal of duplicates and non-quantitative studies, and categorization of the studies. The scope of this review is limited to light-duty passenger EVs and residential charging systems. It excludes large-battery fleets such as electric buses or heavy-duty vehicles that require extensive spatial planning and operational scheduling. However, their broader system-level implications are discussed throughout the review to contextualize future research and urban-scale energy planning.

2. Materials and Methods

Data for this review were collected from technical reports and peer-reviewed modeling studies addressing forecasting and energy management in residential communities and buildings. In total, 110 studies were examined: 36 from North America, 30 from Europe, 37 from Asia, 3 from Africa, and 4 from Oceania. Each study was compiled and, where possible, categorized by description and limitation, geographic context, methodological approach, and reported performance indicators, with quantitative success metrics (Table 1, Table 2, Table 3, Table 4, Table 5, Table 6 and Table 7). It is important to note that contributions from Africa, Latin America, and other low-income economies remain extremely limited due to restricted data availability, fragmented monitoring infrastructure, and constrained research funding. This imbalance hampers model validation, cross-regional benchmarking, and the development of context-specific forecasting and control frameworks, underscoring the urgent need for broader global representation, open data-sharing initiatives, and targeted investment in regional EV-renewable research to ensure methodological inclusivity and equitable policy development.

At the core of these studies is the use of quantitative metrics to evaluate forecasting accuracy, optimization performance, and energy management effectiveness across diverse methodological families, ranging from deep learning to traditional statistical approaches. For supervised learning, success metrics typically capture improvements in prediction accuracy, computational efficiency, and stability, while unsupervised approaches emphasize clustering of EV charging behaviors and enhancing load management strategies. RL studies often report reductions in load or cost variance, reflecting their focus on adaptive control under uncertainty. V2G simulators extend these analyses by modeling individual EVs’ driving and charging behaviors to generate scalable insights on aggregated grid impacts. Traditional approaches such as probabilistic models, Monte Carlo methods, and mixed-integer optimization highlight forecasting accuracy and efficiency gains in regional EV load estimation. Finally, research on overnight/workplace charging strategies and the integration of EV infrastructure with electrified building loads and RESs emphasizes outcomes such as enhanced energy management, peak reduction, cost savings, and household-level bill minimization. Together, these findings demonstrate how forecasting and management frameworks converge to support scalable, resilient, and context-aware energy solutions in residential settings. In addition to these categories, hybrid approaches have emerged as a particularly effective direction, consistently delivering superior forecasting accuracy and more resilient energy management than single-model methods. By integrating supervised, unsupervised, and reinforcement learning techniques—often combined with data decomposition methods such as VMD or CEEMDAN—these models can capture behavioral heterogeneity, reduce data noise, and adapt to uncertainty in real time. This convergence links predictive accuracy with adaptive control, enabling cost savings, peak shaving, and improved renewable integration, while challenges remain around scalability, behavioral variability, and regional generalizability. Together, these findings demonstrate how forecasting and management frameworks, especially when hybridized, converge to support scalable, flexible, and context-aware energy solutions in residential settings.

3. Modeling Techniques for EV Charging Load Forecasting, Charging Behaviors Analysis, and EV Fleet Management Using Deep Learning and Traditional Approaches

Over the past few decades, researchers have developed various methods to improve the accuracy of short-term load forecasting. These approaches use traditional techniques or, in recent years, have been increasingly incorporating DL. With the rapid progress of DL technology, intelligent algorithms have been successfully applied in various fields, especially in big data technologies. Despite the widespread use of DL models by power and energy utility companies to predict energy demand, challenges persist. Firstly, the complexity of the load series involves intricate seasonality patterns. Hourly load depends on the previous hour’s load and historical data from the same hour on previous days and weeks. Secondly, crucial exogenous variables such as weather conditions, holidays, and electricity prices impact the load demand. DL methods, which are divided into supervised, unsupervised, and RL, characterized by multiple hidden layers, effectively address these complexities and are widely employed in load forecasting, with charging strategies that surpass traditional shallow learning approaches. Figure 2 illustrates the utilization of three categories, including supervised, unsupervised, reinforcement learning, and V2G applications, based on total publications from 2015 to 2025. While this review aims to represent each learning paradigm comprehensively, the relative emphasis reflects the actual distribution of research activity, with supervised and reinforcement learning being more mature and extensively validated in EV load forecasting and control applications, whereas unsupervised learning remains comparatively nascent but promising for behavior clustering and data-driven pattern discovery; this imbalance is also influenced by the scope of the review, which focuses on light-duty passenger EVs and residential charging systems rather than large-battery fleets that require more complex spatial planning and operational scheduling frameworks. The figure clearly shows that the utilization of DL has increased gradually over the years. Supervised learning and reinforcement learning have been the most extensively used algorithms in the field of EV load forecasting and charging strategies. Unsupervised learning remains consistently low across all items, indicating a limited contribution compared to supervised learning and reinforcement learning.

Supervised learning models such as Gated Recurrent Units (GRU), Long Short-Term Memory (LSTM), and Bidirectional LSTM (BiLSTM) networks are widely used for forecasting electric loads, predicting EV charging behavior, and estimating battery State of Charge (SOC) [2,4,5,6,7,8]. These models are especially effective for time-series applications because they can capture complex temporal dependencies and nonlinear relationships [8]. GRU and LSTM are particularly effective at learning long-term dependencies in sequential data by mitigating the vanishing gradient problem, while BiLSTM processes information in both forward and backward directions to enhance prediction accuracy. A comparative study on day-ahead EV fleet load forecasting in Canada revealed that BiLSTM outperformed both GRU and LSTM models, delivering slightly higher accuracy in most cases [9]. Building upon recurrent sequence-learning architectures, emerging supervised frameworks such as Federated Learning (FL) and Physics-Informed Neural Networks (PINNs) extend model scalability and interpretability in energy forecasting. FL enables distributed model training across multiple charging stations or utilities without exchanging raw data, preserving user privacy while maintaining predictive performance across heterogeneous datasets. PINNs, in contrast, embed physical laws and energy-balance constraints directly into the learning process, ensuring physically consistent forecasts even under limited or noisy data conditions. When combined with recurrent or hybrid architectures, these methods bridge the gap between data-driven and physics-based modeling, providing greater robustness, transparency, and generalizability across diverse grid and climatic contexts. Table 1 provides an overview of recent research that employs supervised, federated, and physics-informed frameworks designed for regional load prediction or renewable energy generation. To improve usability for practitioners, detailed tabular summaries are complemented by conceptual figures (Figure 3, Figure 4, Figure 5 and Figure 6) that synthesize algorithmic roles, data flows, and decision hierarchies.

Table 1. Summary of research on supervised, federated, and physics-informed frameworks designed for regional energy development.

	Study Description and Limitations	Locations/Region	Models	Success Metrics/Results
[10]	This study establishes an EV charging load forecasting system based on data space theory planning and employs scenario analysis to assess the potential of EVs participating in future regional power grid demand responses. The limitation is that it relies on regional datasets that may not generalize well across different grid conditions or EV fleet compositions.	China	Backpropagation and Monte Carlo Method	The proposed model achieved a minimum prediction error of 0.001 in forecasting regional EV ownership and charging load. This result demonstrates high predictive accuracy and validates the model’s effectiveness as a foundation for analyzing EV participation in grid demand response.
[11]	This paper predicts electricity prices using a supervised learning approach to formulate the real-time EV charging and discharging scheduling problem. The drawback is that case studies were based on a single EV type and California ISO price data, which reduces their generalizability across different vehicle models, regions, and charging infrastructures.	China	LSTM and MDP	The LSTM method reduces charging costs by 38.59% compared to the uncontrolled solution. The hybrid method successfully learned cost-efficient charging and discharging schedules that adapt to dynamic electricity prices while satisfying user driving demands.
[12]	This work suggests a regional-based EV ultra-short-term load prediction approach. The method was only validated on datasets from Hubei province, China, which may not generalize to other regions with different EV adoption, traffic patterns, or grid conditions.	China	LSTM, BP, SVR and traditional method	The proposed LSTM-based method, which incorporates charging pile usage degree, reduces forecasting error compared to traditional methods, achieving an average MAPE of 28.9% versus 33.1% in small regions and 11.3% versus 13.1% in large regions.
[13]	This paper proposes a hybrid method for short-term load forecasting of an Electric Car Fleet by combining similar day (SD) selection, CEEMDAN decomposition, and deep learning models. The approach achieves high forecasting accuracy and demonstrates the benefits of hybridization over single-model techniques. However, its limitations include a narrow forecasting horizon restricted to one- to three-day predictions. The model has not been validated for medium- or long-term horizons, which are equally important for utility planning and strategic grid management.	Canada	Hybrid SD-CEEMDAN-Bi-LSTM	The proposed hybrid model outperformed both single models and other hybrid approaches, delivering the highest forecasting accuracy. It achieved the lowest MAPE of 2.63% across Canada, indicating superior stability and predictive performance for short-term EV fleet load forecasting.
[14]	This paper performs a comprehensive data analysis on EV charging stations. The drawback is that external factors, such as weather and heterogeneous user preferences, were not incorporated, which may affect prediction accuracy in real-world applications.	China	LSTM, historical average (HA), and the auto-regressive integrated moving average (ARIMA)	This paper predicts the total national charge energy consumption of EV charging stations using the temporal encoder–decoder plus LSTM model and compares it with HA and ARIMA. They found that the MAPE of the proposed approach is reduced by 1.2% on the National charge data.
[15]	This work introduces a hybrid method for short-term load forecasting of EV charging stations to improve accuracy problems. However, the approach is tested only on datasets from the U.S. (ANN-DATA) and a single Chinese station in Lanzhou, which may limit generalizability to broader grid or geographic conditions.	China	VMD-BiLSTM	The proposed model outperforms both single-model baselines and other optimized hybrid approaches, achieving a MAPE of 5.44% (Dataset 1) and 4.35% (Dataset 2).
[16]	This paper proposes a time-window electricity price forecasting method to optimize routes and charging schemes for EVs. However, the method is validated only in a limited case study with simulated scenarios, which may not fully capture real-world traffic dynamics, large-scale EV adoption, or communication uncertainties.	China	GRU and two algorithms	The time window price and intraday forecasting approaches achieved an MAPE of about 1.57% and 6.51%, respectively. Simulations demonstrated that the integrated algorithm provided charging routes with significantly lower costs compared to traditional path planning, while maintaining travel time efficiency.
[17]	This paper focuses on developing intelligent planning algorithms to manage public charging demand by predicting EV charging times. Its main limitation is the reliance on data from only two public charging stations, making the findings region-specific and less generalizable.	Morocco	RNN, LSTM, and GRU	The GRU regression model achieved the best performance with an MSE of 0.474%, demonstrating strong capability in capturing time-series and time-dependent patterns.
[18]	This paper assesses the energy supply/demand performance of a group of residential buildings in a Canadian community. Limitations include reliance on a small sample of 13 residential buildings, assumptions of consistent EV load and availability at charging times, and exclusion of other renewable and storage technologies, which restrict broader applicability.	Canada	GRU	The photovoltaic system supplied about 29.2% of the community’s annual EV load, and the GRU model achieved an accuracy of 88.6% in predicting the total load. Results show that coordinated charging strategies can improve solar use and lessen grid dependence.
[9]	This study employed and compared three DL algorithms to forecast the aggregate load for charging a fleet of EVs in Canada. While the models demonstrate strong performance, limitations include their reliance on a Canadian dataset that may not generalize to other regions.	Canada	LSTM, bi-LSTM, and GRU	Among the tested models, Bi-LSTM achieved the best results with a MAPE of 6.5% during testing, outperforming LSTM and GRU.
[19]	This study offers a comprehensive survey focused on load and RE forecasting. Its limitations include being restricted to works published up to 2020, and while it synthesizes methods and datasets, it does not provide a load forecasting model for EVs.	Cyprus	RNN, CNN, LSTM, GRU, and various Hybrid Models	The survey highlights that hybrid DL approaches (e.g., LSTM + optimization/decomposition) often achieve better accuracy than single models, particularly in handling intermittency and variability of renewable energy.
[20]	This paper uses a deep learning technique to study the optimal operation of electrical transportation systems and energy distribution resources. It depicts the uncertain charging behavior of EV users and their loads in smart cities. Limitations include reliance on linear modeling simplifications, assumptions of full infrastructure availability, and limited validation—using only Madrid Metro Line 3 as a case study—making broader applicability uncertain.	China	LSTM and Point Estimate Method,	The proposed framework demonstrated cost reductions by integrating regenerative braking energy, coordinating DERs, and applying demand response. The LSTM method was applied for forecasting hourly solar radiation and wind speed as part of the smart city model. However, the paper does not report explicit error metrics for the LSTM forecasting results.
[21]	This work provides an overview of DL applications for demand-side response. A limitation is that, while the review provides breadth, it is bounded by the literature up to 2019, missing more recent advances in scalable reinforcement learning for EV charging behaviors.	UK & USA	LSTM, Feed forward, and Convolutional NN	The survey finds that deep learning is particularly effective for short-term load forecasting and DR scheduling, with improvements commonly measured using MAE, RMSE, and MAPE.
[22]	This work analyzes charging demand on a session-by-session basis, aiming to facilitate various scheduling or V2G solutions that rely on predicting demand, often in real-time. Limitations include reliance on a single-state dataset and moderate predictive accuracy (with ~50% of demand variance unexplained).	USA	Linear, RF, SVM, and XGBoost methods	Among the tested models, XGBoost performed the best in predicting the charging demand, achieving an RMSE of 6.68 kWh, an MAE of 4.57 kWh, and an R2 of 51.9% on the test set.
[23]	This research proposes a comparative study of DL approaches to forecasting plug-in EVs’ super-short-term stochastic charging load. The main limitation is that the results are based on region-specific datasets with limited variability, which may affect generalizability to other locations or larger-scale EV fleets.	China	ANN, RNN, LSTM, GRU, SAEs, and Bi-LSTM	LSTM achieved an MAE of below 0.6 and an R2 of above 0.99 in charging station data, and an MAE as low as 0.29 with an R2 of above 0.97 in an official charging site aggregator case, confirming it as the most robust model for super-short-term EV charging load forecasting.
[24]	This work employs and compares four DL approaches in forecasting the EVs’ charging load from the charging station perspective. Its limitations include reliance on data from a single charging station with limited EV penetration, which restricts the generalizability of the results to larger networks or diverse charging infrastructures.		DNN, RNN, LSTM, GRU	The one hidden-layer GRU model achieved the best results, with NRMSE of 1.48% (train) and 2.89% (test), and NMAE of 0.47% (train) and 0.77% (test).
[7]	This paper proposes a convolutional recurrent unit network for estimating the state of charge (SOC) of lithium-ion batteries. Its limitations include reliance on laboratory test profiles that may not fully capture real-world driving conditions, as well as manual hyperparameter tuning, which can lead to overfitting.	China	RNN, GRU, CNN-GRU, ELM, and SVM	The CNN-GRU model outperformed benchmarks in various tests. MAE and RMSE were the lowest among all models at room temperature, with errors of less than 2%.
[25]	This work introduces a wind power prediction model that uses DL in smart grids. Limitations include the reliance on high-quality SCADA and NWP data, the need to select benchmark farms with strong correlation to regional output, and the absence of long-term forecasting capabilities.	China	BP, SVM, and LSTM	The LSTM-GPR model achieved high prediction accuracy with a normalized mean absolute error (eNMAE) of 0.57% and a normalized root mean square error (eNRMSE) of 1.07%, outperforming BP and SVM baselines by reducing errors by about 4–7%
[4]	This work forecasts the aggregated demand-side load over short- and medium-term monthly horizons. Limitations include the computational cost of GA tuning and dependence on a single-country dataset (France), which may impact generalizability.	UAE	LSTM. GA-LSTM-, Random Forest, Gradient Boosting, Extra Trees Regressor	The GA-optimized LSTM outperformed all machine learning benchmarks, achieving an RMSE of 288.32 MW, an MAE of 222.71 MW, and a CVRMSE of 0.57%. In comparison, the best machine learning baseline, the Extra Trees Regressor, recorded an RMSE of 428 MW, an MAE of 292 MW, and a CVRMSE of 0.78%.
[26]	This paper presents a comprehensive data-driven approach-based demand-side management for a solar-powered EV charging station connected to a microgrid. Limitations include reliance on a 24 h case study simulation without long-term validation, assumptions of full data availability (EV SoC, arrival/departure), and limited scalability testing under diverse EV penetration levels or stochastic grid events.	Qatar	LSTM, and Vector Autoregressive Moving Average (VARIMA)	The LSTM-based SoC estimator achieved higher accuracy compared to VARIMA, with an RMSE of 0.49% versus 0.87%. The DSM framework effectively reduced the peak demand supplied by the traditional generator from 182 kW to below 100 kW, demonstrating successful peak clipping and improved grid performance stability.
[27]	This study proposes a predict-and-optimize framework for scheduling charging events of electric buses to maximize the utilization of wind power. Limitations include the reliance on wind energy forecasts, which have inherent uncertainty, as prediction accuracy decreases with longer horizons, as well as assumptions such as fixed bus energy consumption (1 kWh/km) and simplified charging efficiency.	Ireland & Spain	LSTM, MIP, MPM, SPM	LSTM outperformed the baselines for 6 h forecasts, with a MAPE of 113.08 and RMSE of 313.56, compared to a MAPE of 230.14 and RMSE of 520.61 for the Smart Grid baseline, and it performed much better than SARIMA results. In scheduling, SPM cut non-clean energy use by ~2.2% and MPM by ~1.5%, with LSTM optimization within 6–12% of the oracle (perfect predictions).
[28]	This study proposes a horizontal FL framework for short-term EV charging-load forecasting while preserving data privacy across multiple urban charging stations. The limitation is that the model’s distributed implementation experiences slower convergence and reduced stability due to heterogeneous data distributions and unequal computational capacities across participating nodes.	China	PSO-VMD and LSTM under a horizontal FL architecture	The proposed model reduces MAE by ≈69–75% and RMSE by ≈63–66% compared to conventional deep learning baselines (LSTM, CNN, RNN).
[29]	This work develops a vertically FL model featuring a spatio-temporal hybrid attention mechanism to predict EV charging-station load. It incurs a significant computational burden due to homomorphic encryption and multi-module attention integration, which slows training and limits scalability for larger systems.	China	Edge-aggregation graph attention network combined with LSTM	The success metric is evaluated by MAPE, RMSE, and response speed, compared with local and non-federated baselines. The approach improved prediction performance by approximately 4%, achieving sub-second response time while preserving TN data privacy.
[30]	The paper proposes FMGCN, a federated meta-learning–based graph convolutional network that integrates spatio-temporal attention and distributed pre-training to forecast regional EV charging demand across six cities. The study’s evaluation is limited to a short-term dataset (35 days) from six cities in China’s Greater Bay Area, which may constrain generalizability to other regions and climatic conditions.	China, Singapore, & USA	Spatial-Temporal (ST) attention GCN backbone	FMGCN achieved up to 36% RMSE reduction and 62% faster convergence for EV charging demand forecasting across cities.
[31]	This research applies physics-informed neural networks (PINNs) for wind-power forecasting, embedding physical laws into loss functions to improve generalization under limited data. The model’s reliance on simulated and synthetic data, along with the absence of real-time SCADA integration, constrains its generalizability and real-world deployment potential across diverse climatic and technological conditions.	India	ML- PINN-Simulink framework	The Stacking Ensemble model achieved approximately a 2–4% lower prediction uncertainty compared to other conventional ML models and a 7% reduction relative to the PINN model, demonstrating the ensemble’s superior reliability and robustness.

Table 1. Summary of research on supervised, federated, and physics-informed frameworks designed for regional energy development.

	Study Description and Limitations	Locations/Region	Models	Success Metrics/Results
[10]	This study establishes an EV charging load forecasting system based on data space theory planning and employs scenario analysis to assess the potential of EVs participating in future regional power grid demand responses. The limitation is that it relies on regional datasets that may not generalize well across different grid conditions or EV fleet compositions.	China	Backpropagation and Monte Carlo Method	The proposed model achieved a minimum prediction error of 0.001 in forecasting regional EV ownership and charging load. This result demonstrates high predictive accuracy and validates the model’s effectiveness as a foundation for analyzing EV participation in grid demand response.
[11]	This paper predicts electricity prices using a supervised learning approach to formulate the real-time EV charging and discharging scheduling problem. The drawback is that case studies were based on a single EV type and California ISO price data, which reduces their generalizability across different vehicle models, regions, and charging infrastructures.	China	LSTM and MDP	The LSTM method reduces charging costs by 38.59% compared to the uncontrolled solution. The hybrid method successfully learned cost-efficient charging and discharging schedules that adapt to dynamic electricity prices while satisfying user driving demands.
[12]	This work suggests a regional-based EV ultra-short-term load prediction approach. The method was only validated on datasets from Hubei province, China, which may not generalize to other regions with different EV adoption, traffic patterns, or grid conditions.	China	LSTM, BP, SVR and traditional method	The proposed LSTM-based method, which incorporates charging pile usage degree, reduces forecasting error compared to traditional methods, achieving an average MAPE of 28.9% versus 33.1% in small regions and 11.3% versus 13.1% in large regions.
[13]	This paper proposes a hybrid method for short-term load forecasting of an Electric Car Fleet by combining similar day (SD) selection, CEEMDAN decomposition, and deep learning models. The approach achieves high forecasting accuracy and demonstrates the benefits of hybridization over single-model techniques. However, its limitations include a narrow forecasting horizon restricted to one- to three-day predictions. The model has not been validated for medium- or long-term horizons, which are equally important for utility planning and strategic grid management.	Canada	Hybrid SD-CEEMDAN-Bi-LSTM	The proposed hybrid model outperformed both single models and other hybrid approaches, delivering the highest forecasting accuracy. It achieved the lowest MAPE of 2.63% across Canada, indicating superior stability and predictive performance for short-term EV fleet load forecasting.
[14]	This paper performs a comprehensive data analysis on EV charging stations. The drawback is that external factors, such as weather and heterogeneous user preferences, were not incorporated, which may affect prediction accuracy in real-world applications.	China	LSTM, historical average (HA), and the auto-regressive integrated moving average (ARIMA)	This paper predicts the total national charge energy consumption of EV charging stations using the temporal encoder–decoder plus LSTM model and compares it with HA and ARIMA. They found that the MAPE of the proposed approach is reduced by 1.2% on the National charge data.
[15]	This work introduces a hybrid method for short-term load forecasting of EV charging stations to improve accuracy problems. However, the approach is tested only on datasets from the U.S. (ANN-DATA) and a single Chinese station in Lanzhou, which may limit generalizability to broader grid or geographic conditions.	China	VMD-BiLSTM	The proposed model outperforms both single-model baselines and other optimized hybrid approaches, achieving a MAPE of 5.44% (Dataset 1) and 4.35% (Dataset 2).
[16]	This paper proposes a time-window electricity price forecasting method to optimize routes and charging schemes for EVs. However, the method is validated only in a limited case study with simulated scenarios, which may not fully capture real-world traffic dynamics, large-scale EV adoption, or communication uncertainties.	China	GRU and two algorithms	The time window price and intraday forecasting approaches achieved an MAPE of about 1.57% and 6.51%, respectively. Simulations demonstrated that the integrated algorithm provided charging routes with significantly lower costs compared to traditional path planning, while maintaining travel time efficiency.
[17]	This paper focuses on developing intelligent planning algorithms to manage public charging demand by predicting EV charging times. Its main limitation is the reliance on data from only two public charging stations, making the findings region-specific and less generalizable.	Morocco	RNN, LSTM, and GRU	The GRU regression model achieved the best performance with an MSE of 0.474%, demonstrating strong capability in capturing time-series and time-dependent patterns.
[18]	This paper assesses the energy supply/demand performance of a group of residential buildings in a Canadian community. Limitations include reliance on a small sample of 13 residential buildings, assumptions of consistent EV load and availability at charging times, and exclusion of other renewable and storage technologies, which restrict broader applicability.	Canada	GRU	The photovoltaic system supplied about 29.2% of the community’s annual EV load, and the GRU model achieved an accuracy of 88.6% in predicting the total load. Results show that coordinated charging strategies can improve solar use and lessen grid dependence.
[9]	This study employed and compared three DL algorithms to forecast the aggregate load for charging a fleet of EVs in Canada. While the models demonstrate strong performance, limitations include their reliance on a Canadian dataset that may not generalize to other regions.	Canada	LSTM, bi-LSTM, and GRU	Among the tested models, Bi-LSTM achieved the best results with a MAPE of 6.5% during testing, outperforming LSTM and GRU.
[19]	This study offers a comprehensive survey focused on load and RE forecasting. Its limitations include being restricted to works published up to 2020, and while it synthesizes methods and datasets, it does not provide a load forecasting model for EVs.	Cyprus	RNN, CNN, LSTM, GRU, and various Hybrid Models	The survey highlights that hybrid DL approaches (e.g., LSTM + optimization/decomposition) often achieve better accuracy than single models, particularly in handling intermittency and variability of renewable energy.
[20]	This paper uses a deep learning technique to study the optimal operation of electrical transportation systems and energy distribution resources. It depicts the uncertain charging behavior of EV users and their loads in smart cities. Limitations include reliance on linear modeling simplifications, assumptions of full infrastructure availability, and limited validation—using only Madrid Metro Line 3 as a case study—making broader applicability uncertain.	China	LSTM and Point Estimate Method,	The proposed framework demonstrated cost reductions by integrating regenerative braking energy, coordinating DERs, and applying demand response. The LSTM method was applied for forecasting hourly solar radiation and wind speed as part of the smart city model. However, the paper does not report explicit error metrics for the LSTM forecasting results.
[21]	This work provides an overview of DL applications for demand-side response. A limitation is that, while the review provides breadth, it is bounded by the literature up to 2019, missing more recent advances in scalable reinforcement learning for EV charging behaviors.	UK & USA	LSTM, Feed forward, and Convolutional NN	The survey finds that deep learning is particularly effective for short-term load forecasting and DR scheduling, with improvements commonly measured using MAE, RMSE, and MAPE.
[22]	This work analyzes charging demand on a session-by-session basis, aiming to facilitate various scheduling or V2G solutions that rely on predicting demand, often in real-time. Limitations include reliance on a single-state dataset and moderate predictive accuracy (with ~50% of demand variance unexplained).	USA	Linear, RF, SVM, and XGBoost methods	Among the tested models, XGBoost performed the best in predicting the charging demand, achieving an RMSE of 6.68 kWh, an MAE of 4.57 kWh, and an R2 of 51.9% on the test set.
[23]	This research proposes a comparative study of DL approaches to forecasting plug-in EVs’ super-short-term stochastic charging load. The main limitation is that the results are based on region-specific datasets with limited variability, which may affect generalizability to other locations or larger-scale EV fleets.	China	ANN, RNN, LSTM, GRU, SAEs, and Bi-LSTM	LSTM achieved an MAE of below 0.6 and an R2 of above 0.99 in charging station data, and an MAE as low as 0.29 with an R2 of above 0.97 in an official charging site aggregator case, confirming it as the most robust model for super-short-term EV charging load forecasting.
[24]	This work employs and compares four DL approaches in forecasting the EVs’ charging load from the charging station perspective. Its limitations include reliance on data from a single charging station with limited EV penetration, which restricts the generalizability of the results to larger networks or diverse charging infrastructures.		DNN, RNN, LSTM, GRU	The one hidden-layer GRU model achieved the best results, with NRMSE of 1.48% (train) and 2.89% (test), and NMAE of 0.47% (train) and 0.77% (test).
[7]	This paper proposes a convolutional recurrent unit network for estimating the state of charge (SOC) of lithium-ion batteries. Its limitations include reliance on laboratory test profiles that may not fully capture real-world driving conditions, as well as manual hyperparameter tuning, which can lead to overfitting.	China	RNN, GRU, CNN-GRU, ELM, and SVM	The CNN-GRU model outperformed benchmarks in various tests. MAE and RMSE were the lowest among all models at room temperature, with errors of less than 2%.
[25]	This work introduces a wind power prediction model that uses DL in smart grids. Limitations include the reliance on high-quality SCADA and NWP data, the need to select benchmark farms with strong correlation to regional output, and the absence of long-term forecasting capabilities.	China	BP, SVM, and LSTM	The LSTM-GPR model achieved high prediction accuracy with a normalized mean absolute error (eNMAE) of 0.57% and a normalized root mean square error (eNRMSE) of 1.07%, outperforming BP and SVM baselines by reducing errors by about 4–7%
[4]	This work forecasts the aggregated demand-side load over short- and medium-term monthly horizons. Limitations include the computational cost of GA tuning and dependence on a single-country dataset (France), which may impact generalizability.	UAE	LSTM. GA-LSTM-, Random Forest, Gradient Boosting, Extra Trees Regressor	The GA-optimized LSTM outperformed all machine learning benchmarks, achieving an RMSE of 288.32 MW, an MAE of 222.71 MW, and a CVRMSE of 0.57%. In comparison, the best machine learning baseline, the Extra Trees Regressor, recorded an RMSE of 428 MW, an MAE of 292 MW, and a CVRMSE of 0.78%.
[26]	This paper presents a comprehensive data-driven approach-based demand-side management for a solar-powered EV charging station connected to a microgrid. Limitations include reliance on a 24 h case study simulation without long-term validation, assumptions of full data availability (EV SoC, arrival/departure), and limited scalability testing under diverse EV penetration levels or stochastic grid events.	Qatar	LSTM, and Vector Autoregressive Moving Average (VARIMA)	The LSTM-based SoC estimator achieved higher accuracy compared to VARIMA, with an RMSE of 0.49% versus 0.87%. The DSM framework effectively reduced the peak demand supplied by the traditional generator from 182 kW to below 100 kW, demonstrating successful peak clipping and improved grid performance stability.
[27]	This study proposes a predict-and-optimize framework for scheduling charging events of electric buses to maximize the utilization of wind power. Limitations include the reliance on wind energy forecasts, which have inherent uncertainty, as prediction accuracy decreases with longer horizons, as well as assumptions such as fixed bus energy consumption (1 kWh/km) and simplified charging efficiency.	Ireland & Spain	LSTM, MIP, MPM, SPM	LSTM outperformed the baselines for 6 h forecasts, with a MAPE of 113.08 and RMSE of 313.56, compared to a MAPE of 230.14 and RMSE of 520.61 for the Smart Grid baseline, and it performed much better than SARIMA results. In scheduling, SPM cut non-clean energy use by ~2.2% and MPM by ~1.5%, with LSTM optimization within 6–12% of the oracle (perfect predictions).
[28]	This study proposes a horizontal FL framework for short-term EV charging-load forecasting while preserving data privacy across multiple urban charging stations. The limitation is that the model’s distributed implementation experiences slower convergence and reduced stability due to heterogeneous data distributions and unequal computational capacities across participating nodes.	China	PSO-VMD and LSTM under a horizontal FL architecture	The proposed model reduces MAE by ≈69–75% and RMSE by ≈63–66% compared to conventional deep learning baselines (LSTM, CNN, RNN).
[29]	This work develops a vertically FL model featuring a spatio-temporal hybrid attention mechanism to predict EV charging-station load. It incurs a significant computational burden due to homomorphic encryption and multi-module attention integration, which slows training and limits scalability for larger systems.	China	Edge-aggregation graph attention network combined with LSTM	The success metric is evaluated by MAPE, RMSE, and response speed, compared with local and non-federated baselines. The approach improved prediction performance by approximately 4%, achieving sub-second response time while preserving TN data privacy.
[30]	The paper proposes FMGCN, a federated meta-learning–based graph convolutional network that integrates spatio-temporal attention and distributed pre-training to forecast regional EV charging demand across six cities. The study’s evaluation is limited to a short-term dataset (35 days) from six cities in China’s Greater Bay Area, which may constrain generalizability to other regions and climatic conditions.	China, Singapore, & USA	Spatial-Temporal (ST) attention GCN backbone	FMGCN achieved up to 36% RMSE reduction and 62% faster convergence for EV charging demand forecasting across cities.
[31]	This research applies physics-informed neural networks (PINNs) for wind-power forecasting, embedding physical laws into loss functions to improve generalization under limited data. The model’s reliance on simulated and synthetic data, along with the absence of real-time SCADA integration, constrains its generalizability and real-world deployment potential across diverse climatic and technological conditions.	India	ML- PINN-Simulink framework	The Stacking Ensemble model achieved approximately a 2–4% lower prediction uncertainty compared to other conventional ML models and a 7% reduction relative to the PINN model, demonstrating the ensemble’s superior reliability and robustness.

The conceptual flowchart of the unifying supervised learning framework for EV load forecasting illustrates a streamlined pipeline beginning with input and feature engineering, where diverse datasets (e.g., charging sessions, weather, tariffs, and fleet composition) are standardized and refined through supervised feature selection. These inputs are passed to supervised deep learning models such as LSTM, GRU, Bi-LSTM, CNN–LSTM hybrids, attention-based architectures, and transformer models, which form the methodological core of the framework. The flow then advances to training and evaluation, employing rolling horizon validation and standardized error metrics (MAE, RMSE, and MAPE) to ensure reproducibility and comparability. To extend applicability, the framework incorporates generalization and transferability, enabling cross-regional forecasting, transfer learning, and multi-horizon prediction. Finally, the flowchart concludes with integration into grid objectives, where supervised forecasts are linked to demand-side management, vehicle-to-grid (V2G) optimization, and scheduling, ensuring that methodological advances directly support practical energy system planning and sustainability goals (Figure 3).

Unsupervised learning is a type of machine learning in which the algorithm learns from unlabeled data without specific guidance or supervision. Unlike supervised learning, where the model is trained using labeled examples (input–output pairs), unsupervised learning operates on data without explicit target values. It helps to discover patterns, relationships, and structures within the data without relying on labeled outputs. There is limited literature reviewing the use of unsupervised learning to cluster EV charging behaviors and improve EV load management. Recent studies have also integrated explainable deep learning (XDL) techniques with unsupervised frameworks to improve interpretability—revealing which features or behavioral patterns most influence clustering outcomes, and supporting transparent, data-driven decision-making in energy performance. Table 2 summarizes research on unsupervised and explainable DL models developed for regional load development.

Table 2. Summary of research on unsupervised and explainable deep learning models for regional energy development.

Source	Study Description and Limitations	Locations /Region	Models	Success Metrics/Results
[8,13]	These papers used the Similar Day (SD) selection to capture the features of load using the XGBoost algorithm and cluster them using the k-means method. The limitations of both studies arise from their reliance on clustering quality and the representativeness of SDs. The effectiveness of SD and k-means greatly depends on feature selection (such as weather, day type, or EV characteristics), which can limit generalizability when applied to regions or datasets with different grid conditions, EV adoption levels, or user behaviors.	Canada	Clustering	SD selection and k-means clustering proved effective in both grid-level [8] and EV fleet forecasting [13], with improvements in the MAPE of approximately 3% and 2%, respectively. However, the results depend heavily on regional datasets and careful feature weighting.
[32]	This book chapter’s scope is to utilize ML techniques to investigate and analyze the charging behavior of EVs, and to cluster the charging patterns of these vehicles. The main limitation is the sensitivity of clustering results to feature selection and parameter tuning (e.g., number of clusters). This may lead to unstable or misleading groupings if applied to datasets from other regions or behavioral conditions	Iran & UK	Clustering and Classification	Results demonstrated that clustering enabled better recognition of regular versus irregular charging behaviors, improving the potential for smart charging scheduling and demand management.
[33]	This article proposes an ML-based EV profiling technique to better understand the information behind the random probability and irregularity of EV load. Limitations include the reliance on datasets with missing values, limited training data for some stations, and the focus on daily maximum load prediction rather than full temporal profiles.	South Korea	Clustering	The LSTM-based profiling method outperformed statistical baselines (ES, SARIMA, TBATS) in three of four datasets, achieving the lowest MAE values across most station groups. However, performance degraded in stations with high data frequency but few locations, due to insufficiently diverse training samples.
[34]	This paper employs a clustering algorithm and a multilayer perceptron on historical charging records to predict EV load for smart energy management. The main limitations include dependence on a region-specific dataset from Los Angeles, the sensitivity of clustering results to feature selection and parameter tuning, and limited validation with larger or more diverse EV populations, which might affect overall applicability.	USA	Clustering & Classification	The MLP achieved an average classification accuracy of approximately 85% on the training set and about 78% on the test set across 10-fold cross-validation. For load forecasting, the MAPE ranged from 0.145 to 0.351, confirming that the combined clustering–classification framework is suitable for charging control scheduling and online EV load forecasting.
[35]	This study developed an explainable deep learning (XDL) framework to classify building energy performance, an extensive dataset of buildings from the EPC Dataset Region Lombardy, Italy. The model’s generalization and scalability are constrained by reliance on region-specific energy datasets, the absence of multi-climate validation, and untested computational feasibility for real-time deployment.	Portugal and Italy	Clustering & Classifica-tion	The XDL achieved high classification accuracy (≈99.95%) and outperformed conventional ML baselines.
[36]	This study proposes a sustainable framework for locating EV charging stations to identify demand hot spots and guide EVCS siting. The main limitations include relying on coarse-resolution district-level data, which may overlook finer spatial details and heterogeneity.	Brazil	Kernel Density Estimation (KDE)	The analysis identified 21 districts as priority areas for expanding charging infrastructure, with the highest demand focused in central and western São Paulo. Validation against the 90 existing charging stations confirmed that the methodology correctly matched 9 of the top-ranked districts with real-world installations.
[37]	This study develops a data-driven framework to analyze and forecast EV charging energy consumption by constructing representative EV load profiles. The main limitation is the assumption of unidirectional charging reliance on EVSE data without including external variables such as traffic, travel history, or weather.	South Africa	Regression models and KDE	Combining a KDE-based charging-duration model with advanced supervised ML methods—and fine tree regression—provides highly accurate and scalable tools for modeling stochastic EV charging behavior and generating reliable load profiles for grid applications.
[38]	This study proposes a dynamic charging price strategy to optimally balance EV charging loads between residential and commercial/industrial distribution networks. The limitations include reliance on assumptions of perfect communication and user responsiveness to dynamic prices, use of simplified GMMs for load representation, and testing only under simulation conditions without field deployment.	South Africa	Gaussian mixture models (GMMs)	For residential networks, GMM-based forecasting reduced MAPE from 14.42% under TOU pricing to 6.48% with the dynamic strategy. In commercial/industrial networks, MAPE dropped from 29.7% to 11.13%, confirming improved accuracy and load alignment.

Table 2. Summary of research on unsupervised and explainable deep learning models for regional energy development.

Source	Study Description and Limitations	Locations /Region	Models	Success Metrics/Results
[8,13]	These papers used the Similar Day (SD) selection to capture the features of load using the XGBoost algorithm and cluster them using the k-means method. The limitations of both studies arise from their reliance on clustering quality and the representativeness of SDs. The effectiveness of SD and k-means greatly depends on feature selection (such as weather, day type, or EV characteristics), which can limit generalizability when applied to regions or datasets with different grid conditions, EV adoption levels, or user behaviors.	Canada	Clustering	SD selection and k-means clustering proved effective in both grid-level [8] and EV fleet forecasting [13], with improvements in the MAPE of approximately 3% and 2%, respectively. However, the results depend heavily on regional datasets and careful feature weighting.
[32]	This book chapter’s scope is to utilize ML techniques to investigate and analyze the charging behavior of EVs, and to cluster the charging patterns of these vehicles. The main limitation is the sensitivity of clustering results to feature selection and parameter tuning (e.g., number of clusters). This may lead to unstable or misleading groupings if applied to datasets from other regions or behavioral conditions	Iran & UK	Clustering and Classification	Results demonstrated that clustering enabled better recognition of regular versus irregular charging behaviors, improving the potential for smart charging scheduling and demand management.
[33]	This article proposes an ML-based EV profiling technique to better understand the information behind the random probability and irregularity of EV load. Limitations include the reliance on datasets with missing values, limited training data for some stations, and the focus on daily maximum load prediction rather than full temporal profiles.	South Korea	Clustering	The LSTM-based profiling method outperformed statistical baselines (ES, SARIMA, TBATS) in three of four datasets, achieving the lowest MAE values across most station groups. However, performance degraded in stations with high data frequency but few locations, due to insufficiently diverse training samples.
[34]	This paper employs a clustering algorithm and a multilayer perceptron on historical charging records to predict EV load for smart energy management. The main limitations include dependence on a region-specific dataset from Los Angeles, the sensitivity of clustering results to feature selection and parameter tuning, and limited validation with larger or more diverse EV populations, which might affect overall applicability.	USA	Clustering & Classification	The MLP achieved an average classification accuracy of approximately 85% on the training set and about 78% on the test set across 10-fold cross-validation. For load forecasting, the MAPE ranged from 0.145 to 0.351, confirming that the combined clustering–classification framework is suitable for charging control scheduling and online EV load forecasting.
[35]	This study developed an explainable deep learning (XDL) framework to classify building energy performance, an extensive dataset of buildings from the EPC Dataset Region Lombardy, Italy. The model’s generalization and scalability are constrained by reliance on region-specific energy datasets, the absence of multi-climate validation, and untested computational feasibility for real-time deployment.	Portugal and Italy	Clustering & Classifica-tion	The XDL achieved high classification accuracy (≈99.95%) and outperformed conventional ML baselines.
[36]	This study proposes a sustainable framework for locating EV charging stations to identify demand hot spots and guide EVCS siting. The main limitations include relying on coarse-resolution district-level data, which may overlook finer spatial details and heterogeneity.	Brazil	Kernel Density Estimation (KDE)	The analysis identified 21 districts as priority areas for expanding charging infrastructure, with the highest demand focused in central and western São Paulo. Validation against the 90 existing charging stations confirmed that the methodology correctly matched 9 of the top-ranked districts with real-world installations.
[37]	This study develops a data-driven framework to analyze and forecast EV charging energy consumption by constructing representative EV load profiles. The main limitation is the assumption of unidirectional charging reliance on EVSE data without including external variables such as traffic, travel history, or weather.	South Africa	Regression models and KDE	Combining a KDE-based charging-duration model with advanced supervised ML methods—and fine tree regression—provides highly accurate and scalable tools for modeling stochastic EV charging behavior and generating reliable load profiles for grid applications.
[38]	This study proposes a dynamic charging price strategy to optimally balance EV charging loads between residential and commercial/industrial distribution networks. The limitations include reliance on assumptions of perfect communication and user responsiveness to dynamic prices, use of simplified GMMs for load representation, and testing only under simulation conditions without field deployment.	South Africa	Gaussian mixture models (GMMs)	For residential networks, GMM-based forecasting reduced MAPE from 14.42% under TOU pricing to 6.48% with the dynamic strategy. In commercial/industrial networks, MAPE dropped from 29.7% to 11.13%, confirming improved accuracy and load alignment.

The unifying framework for unsupervised learning in EV load forecasting and charging behavior analysis emphasizes clustering and pattern discovery as a foundation for forecasting and control. Historical charging data are transformed into feature vectors that include session timing, stay duration, energy demand, similar day selection, and contextual variables such as weather or traffic. Clustering methods such as K-means (fixed partitions), Gaussian Mixture Models (GMMs) (probabilistic groupings), and Kernel Density Estimation (KDE) (continuous density estimation) are applied to uncover hidden patterns and group sessions into homogeneous categories. These clusters provide representative load profiles that enhance hybrid forecasting models, support demand response scheduling, and guide infrastructure planning. Evaluation relies on validity indices alongside improvements in downstream forecasting and scheduling. Despite sensitivities to feature choice, cluster numbers, and regional transferability, the framework offers a robust and scalable means of incorporating behavioral heterogeneity into EV-grid interaction studies (Figure 4).

Reinforcement Learning (RL) is a deep learning approach where an agent learns to make decisions by interacting with an environment and receiving feedback through rewards or penalties [39]. RL has been widely applied in EV fleet management for tasks such as vehicle dispatching, smart scheduling, and coordinated charging. In smart charging applications, Q-learning helps optimize decisions related to cost minimization, grid constraints, and user preferences, and has also been used to enhance user privacy [40]. However, its effectiveness diminishes in complex, high-dimensional scenarios. To address this, Deep Q-Networks (DQN) use deep neural networks to enable scalable and intelligent decision-making [41,42]. Double DQN further improves reliability by reducing the overestimation of Q-values, enhancing stability in uncertain environments [43]. Markov Decision Processes (MDP) offer a mathematical foundation for modeling sequential charging decisions under constraints, while Kernel Density Estimation (KDE) supports a non-parametric prediction of charging demand distributions, improving load forecasting and scheduling. Together, these methods support efficient, flexible, and user-centric charging strategies. Table 3 summarizes recent RL-based models for regional-scale EV fleet management.

Table 3. Summary of research on model developments using reinforcement learning for a scalable fleet of EVs on the regional level.

Source	Study Description and Limitations	Locations/Region	Models	Success Metrics/Results
[40]	This study proposes a multi-agent RL (MARL) approach for EV owners participating in the electricity market. Each EV is considered an agent, and all EVs have vehicle-to-grid capability. Limitations include assumptions of simplified trip behavior (e.g., one trip per day) and constant battery capacity (16 kWh) across all EVs.	Portuguese/ Finland	Q-learning	Performance was evaluated across three bidding scenarios, including constant price bidding, time-of-use, and dynamic hourly bidding. Results confirm that the Q-learning bidding strategy significantly reduces energy costs while preserving user autonomy and privacy in market participation.
[41]	This study developed a neural network that leverages historical data to identify the optimal charging station and time window for electric vehicle recharging. Limitations include the reliance on simulated rather than real-world data, which may not fully capture real charging behaviors or infrastructure constraints.	Romania	DQN	The DQN was trained on 400 cars in a simulated network of three interconnected cities. The results confirm that RL-based smart scheduling can effectively handle congestion and enable optimized trip planning, though validation on real-world datasets remains necessary.
[44]	This paper proposes a charging coordination system for a scalable fleet of EVs. Limitations include reliance on simulated EV user data rather than real-world datasets and testing only under residential overnight charging scenarios. The approach also assumes uniform battery capacities (30 kWh) and slow charging rates (3.7 kW), which may limit generalizability to diverse EV fleets and fast-charging contexts.	Germany	MDP	Performance was evaluated over 300 simulated days, where the RL strategy nearly matched the near-optimal optimization baseline. Compared with uncontrolled charging, it reduced load variance by 65% and demonstrated adaptability close to optimization with perfect foresight.
[45]	This study extracts information on previous energy prices to determine the current charging or discharging actions, thereby proposing an optimal charging control strategy. Limitations include reliance on simulation-based validation rather than real-world deployments, as well as simplified assumptions about user behavior. The model also assumes accurate price signals and stable communication, which may not hold in practice.	China	MDP & LSTM	Performance was measured using charging cost, user satisfaction, and price reward as metrics. Compared to benchmarks, the charging control deep deterministic policy gradient (DDPG) achieved the highest cumulative reward, improving by 56.4% over the DDPG and 68.4% over the DQN in convergence tests.
[46]	This paper proposes an optimal EV charging strategy in a distribution network to maximize the profit of distribution system operators (DSO) while satisfying all the physical constraints. Limitations include dependence on simulation-based validation using synthetic datasets, which may not accurately represent the variability of real-world EV behavior. The approach also presumes continuous charging rates, whereas in reality, charging technologies are discrete and may lead to infeasibility.	China, Denmark, & the USA	MDP	Performance was tested on a 33-bus distribution network with 400 EVs (40% penetration). Results showed that the proposed RL-based strategy increased the daily revenue of the DSO to US$4089, compared to US$2853 for the traditional method, representing a roughly 43% improvement.
[47]	This study proposes a model-free solution based on deep RL for EV charging scheduling. Limitations include reliance on synthetic commuting behavior and electricity price datasets, which may not capture full real-world variability. The model assumes EV users are price-takers and does not consider the feedback effect of widespread adoption on electricity prices.	USA,	Constrained MDP	Performance was compared with DQN, DDPG, Model Predictive Control (MPC), and baseline strategies. The proposed approach reduced the total charging cost by 63.14% compared to the baseline.
[48]	This study proposes a two-layer deep RL framework to optimize EV charging scheduling and voltage control in distribution networks. Limitations include reliance on real-time price signals derived from CAISO data without incorporating market feedback mechanisms, assumptions of centralized coordination with full observability and uniform EV characteristics, and validation limited to a single IEEE-33-bus test system rather than large-scale or geographically diverse networks.	USA, China	MDP	The proposed framework maintained node voltages within the safe range of 0.95–1.05 p.u., converging within approximately 500–1000 episodes, faster than benchmarks. It achieved notable cost reductions of 20.1–30.6% compared with conventional charging strategies while minimizing cumulative voltage violations, demonstrating an effective balance between economic efficiency and grid stability.
[49]	This paper proposes a daily decision-making problem for choosing the amount of energy to charge within a day. Limitations include the reliance on assumed known driving patterns, which may not hold in real-world settings where user behavior is more stochastic. The study also focuses on individual PEVs rather than large-scale fleets, limiting scalability.	Finland	MDP	For price forecasting, the proposed Bayesian Neural Network improved accuracy by reducing MAPE by 3.34% compared to ARIMA and 0.94% compared to VJ-BN. In charging optimization, the RL strategy lowered costs by 10–50% relative to uncontrolled charging.
[50]	This paper addresses the problem of defining a day-ahead consumption plan for charging a fleet of EVs. Limitations include assumptions of the aggregator as a price-taker, meaning it cannot influence market prices, and reliance on simulated fleet scenarios rather than real-world data.	USA	MDP	The RL approach produced near-optimal day-ahead charging schedules, with costs converging to the stochastic programming benchmark within about 20–30 learning days. In large-scale tests with 2500 EVs, it maintained scalability and stability, with costs averaging only ~13% higher under grid constraints.
[51]	This study suggests a data-driven RL method to tackle EV charging and discharging challenges. Limitations include assumptions that all EVs seek a full charge before departure, reliance on simulated charger data from limited sites (two residential locations in Korea), and simplification of charger/battery heterogeneity.	Republic of Korea	KDE	The proposed RL method reduced charging costs to 69.7% of the baseline at Site A and 91.1% at Site B, outperforming other RL benchmarks. It also improved the load factor from 0.64 to 0.71, demonstrating effective peak shaving and valley filling.

Table 3. Summary of research on model developments using reinforcement learning for a scalable fleet of EVs on the regional level.

Source	Study Description and Limitations	Locations/Region	Models	Success Metrics/Results
[40]	This study proposes a multi-agent RL (MARL) approach for EV owners participating in the electricity market. Each EV is considered an agent, and all EVs have vehicle-to-grid capability. Limitations include assumptions of simplified trip behavior (e.g., one trip per day) and constant battery capacity (16 kWh) across all EVs.	Portuguese/ Finland	Q-learning	Performance was evaluated across three bidding scenarios, including constant price bidding, time-of-use, and dynamic hourly bidding. Results confirm that the Q-learning bidding strategy significantly reduces energy costs while preserving user autonomy and privacy in market participation.
[41]	This study developed a neural network that leverages historical data to identify the optimal charging station and time window for electric vehicle recharging. Limitations include the reliance on simulated rather than real-world data, which may not fully capture real charging behaviors or infrastructure constraints.	Romania	DQN	The DQN was trained on 400 cars in a simulated network of three interconnected cities. The results confirm that RL-based smart scheduling can effectively handle congestion and enable optimized trip planning, though validation on real-world datasets remains necessary.
[44]	This paper proposes a charging coordination system for a scalable fleet of EVs. Limitations include reliance on simulated EV user data rather than real-world datasets and testing only under residential overnight charging scenarios. The approach also assumes uniform battery capacities (30 kWh) and slow charging rates (3.7 kW), which may limit generalizability to diverse EV fleets and fast-charging contexts.	Germany	MDP	Performance was evaluated over 300 simulated days, where the RL strategy nearly matched the near-optimal optimization baseline. Compared with uncontrolled charging, it reduced load variance by 65% and demonstrated adaptability close to optimization with perfect foresight.
[45]	This study extracts information on previous energy prices to determine the current charging or discharging actions, thereby proposing an optimal charging control strategy. Limitations include reliance on simulation-based validation rather than real-world deployments, as well as simplified assumptions about user behavior. The model also assumes accurate price signals and stable communication, which may not hold in practice.	China	MDP & LSTM	Performance was measured using charging cost, user satisfaction, and price reward as metrics. Compared to benchmarks, the charging control deep deterministic policy gradient (DDPG) achieved the highest cumulative reward, improving by 56.4% over the DDPG and 68.4% over the DQN in convergence tests.
[46]	This paper proposes an optimal EV charging strategy in a distribution network to maximize the profit of distribution system operators (DSO) while satisfying all the physical constraints. Limitations include dependence on simulation-based validation using synthetic datasets, which may not accurately represent the variability of real-world EV behavior. The approach also presumes continuous charging rates, whereas in reality, charging technologies are discrete and may lead to infeasibility.	China, Denmark, & the USA	MDP	Performance was tested on a 33-bus distribution network with 400 EVs (40% penetration). Results showed that the proposed RL-based strategy increased the daily revenue of the DSO to US$4089, compared to US$2853 for the traditional method, representing a roughly 43% improvement.
[47]	This study proposes a model-free solution based on deep RL for EV charging scheduling. Limitations include reliance on synthetic commuting behavior and electricity price datasets, which may not capture full real-world variability. The model assumes EV users are price-takers and does not consider the feedback effect of widespread adoption on electricity prices.	USA,	Constrained MDP	Performance was compared with DQN, DDPG, Model Predictive Control (MPC), and baseline strategies. The proposed approach reduced the total charging cost by 63.14% compared to the baseline.
[48]	This study proposes a two-layer deep RL framework to optimize EV charging scheduling and voltage control in distribution networks. Limitations include reliance on real-time price signals derived from CAISO data without incorporating market feedback mechanisms, assumptions of centralized coordination with full observability and uniform EV characteristics, and validation limited to a single IEEE-33-bus test system rather than large-scale or geographically diverse networks.	USA, China	MDP	The proposed framework maintained node voltages within the safe range of 0.95–1.05 p.u., converging within approximately 500–1000 episodes, faster than benchmarks. It achieved notable cost reductions of 20.1–30.6% compared with conventional charging strategies while minimizing cumulative voltage violations, demonstrating an effective balance between economic efficiency and grid stability.
[49]	This paper proposes a daily decision-making problem for choosing the amount of energy to charge within a day. Limitations include the reliance on assumed known driving patterns, which may not hold in real-world settings where user behavior is more stochastic. The study also focuses on individual PEVs rather than large-scale fleets, limiting scalability.	Finland	MDP	For price forecasting, the proposed Bayesian Neural Network improved accuracy by reducing MAPE by 3.34% compared to ARIMA and 0.94% compared to VJ-BN. In charging optimization, the RL strategy lowered costs by 10–50% relative to uncontrolled charging.
[50]	This paper addresses the problem of defining a day-ahead consumption plan for charging a fleet of EVs. Limitations include assumptions of the aggregator as a price-taker, meaning it cannot influence market prices, and reliance on simulated fleet scenarios rather than real-world data.	USA	MDP	The RL approach produced near-optimal day-ahead charging schedules, with costs converging to the stochastic programming benchmark within about 20–30 learning days. In large-scale tests with 2500 EVs, it maintained scalability and stability, with costs averaging only ~13% higher under grid constraints.
[51]	This study suggests a data-driven RL method to tackle EV charging and discharging challenges. Limitations include assumptions that all EVs seek a full charge before departure, reliance on simulated charger data from limited sites (two residential locations in Korea), and simplification of charger/battery heterogeneity.	Republic of Korea	KDE	The proposed RL method reduced charging costs to 69.7% of the baseline at Site A and 91.1% at Site B, outperforming other RL benchmarks. It also improved the load factor from 0.64 to 0.71, demonstrating effective peak shaving and valley filling.

A unifying reinforcement learning framework for scalable fleet-level EV charging at the regional level integrates decentralized decision-making with centralized coordination to balance grid stability, user satisfaction, and economic efficiency. At its core, the framework models EV charging as an MDP, where each EV or charging station acts as an agent interacting with a dynamic environment shaped by electricity prices, renewable availability, and grid constraints. Deep RL algorithms—such as DQN, DDPG, and constrained RL—are employed to learn adaptive charging and discharging policies under uncertainty. To ensure scalability, MARL allows agents to learn cooperative or competitive strategies, while an aggregator coordinates global objectives such as peak shaving, valley filling, and cost reduction. Regionally, the framework incorporates transfer learning and federated RL so models trained in one city or utility region can generalize across heterogeneous grids. Finally, integration with DSOs enables RL outcomes to align with broader system objectives, including renewable integration, voltage regulation, and carbon reduction. While promising, challenges remain in handling communication delays, ensuring robustness against market feedback loops, and addressing the computational complexity of large-scale training.

Vehicle-To-Grid (V2G)/Vehicle-to-Building (V2B) is a system in which plug-in EVs sell demand response services to the grid. V2G/V2B-Simulator is a tool that models individual EV usage and aggregates large numbers of simulated vehicles to produce grid-scale predictions of the impacts and opportunities from vehicle-grid integration [52]. Most V2G/V2B simulators employ the RL approach to learn the optimal sequential charging and discharging decisions that minimize charging costs for individual users and reduce peak demand. Table 4 summarizes research on scalable V2G/V2B simulators for the load demand response of an EV fleet.

Table 4. Summary of research on V2G/V2B simulators for EV charging behaviors and load demand response of an EV fleet.

Source	Study Description and Limitations	Locations/Region	Applications	Success Metrics/Results
[53]	This paper introduces EV2Gym, an open-source and modular simulation platform. Limitations include the partial modeling of power network impacts compared with full-grid simulators, reliance on predefined datasets for EV behavior, and potential simplifications in charger and transformer interactions under large-scale deployment scenarios.	Netherlands/ Italy	EV Smart charging research and benchmarking Cost Reduction	EV2Gym was validated against real charging data, showing that its two-stage charging/discharging model closely matched laboratory curves for AC and DC charging.
[54]	This paper proposes a novel RL framework for V2B charging, combining DDPG with action masking and MILP-based policy guidance. Limitations include reliance on accurate long-term peak power estimation, computational intensity for training, and assumptions of centralized control across 15 chargers, which may limit scalability in highly decentralized or competitive settings.		Building Energy Cost Reduction Peak Shaving & Valley Filling Scalable Fleet Management Policy & Tariff Testing	The approach consistently outperformed heuristic baselines and closely matched the MILP oracle, reducing monthly building energy bills and peak demand charges.
[55]	This thesis investigates RL approaches for controlling Virtual Power Plants (VPPs) composed of EVs, renewable sources (PV and wind), and household loads. Limitations include reliance on synthetic charging data generated by the Elvis simulator (not fully real-world calibrated) and limited scalability beyond the modeled microgrid (4 households, 4 EVSEs, 40 PV units, and 8 wind turbines).	Germany	Peak Shaving & Valley Filling Self-Consumption Optimization Grid Independence VPPs Scenario Testing	The proposed approach achieved the best performance, delivering effective peak shaving, valley filling, and a stable SoC at departure.
[56]	A cloud-based simulator comprises multiple subsystems, including a real-time power system simulator, an EV charge scheduler system, and a smart energy plaza with various charging stations for large-scale EV charging. However, its validation remains limited to controlled test cases and simulated EV fleets, meaning real-world uncertainties—such as communication delays, heterogeneous user behaviors, and scalability challenges with extreme fast charging—are not yet fully addressed.	USA	Testing of V2G technologies	Results showed effective grid constraint management and frequency regulation in a 460-EV simulation, ensuring SoC targets while avoiding overloads.
[57]	V2G-Sim models the driving and charging patterns of individual plug-in EVs (PEVs) and generates detailed predictions of grid impacts and opportunities over time and space, based on increased PEV adoption. The simulator is scalable and can analyze any number of vehicles, from a single unit to over a million PEVs. Limitations include strong reliance on assumptions about EV adoption rates, charging infrastructure availability, and user charging behavior.	USA	Demand response from PEVs PEVs For Renewables Integration Predicting PEV Charging Demands Spatial Resolution Electricity pricing schemes and their impact on the grid	V2G-controlled charging effectively reduced peak loads and mitigated net load ramps compared to uncontrolled charging.
[58]	The EVs’ user charging behavior simulator consists of two main parts: data generation and trip simulation. Data is generated based on each user’s profile, which includes vehicle details, planned trips (with locations and departure times), and behavioral parameters. While primarily designed for unidirectional charging, the framework can be extended to V2G scenarios, which limits its direct applicability for grid services. Other limitations include simplified assumptions such as user profiles not varying across weekdays, and charging decisions driven mainly by SOC thresholds.	Portugal	Testing the impact of different types of incentives on EV users’ behavior Evaluating the impact of the variation in electricity prices on the behavior of EV users Generating data and scaling it for regional application	For 2500 EVs, variable charging prices yielded user cost reductions of 4% (distance preference), 10% (price preference), and 2% (time preference) compared to fixed prices.
[59]	ACN-Sim is an open-source simulator that offers a modular, extensible architecture, capturing the complexity of real charging systems, including battery charging and user interaction behaviors. Although originally designed for unidirectional charging, ACN-Sim’s modular architecture supports extensions to V2G by adapting charging profiles and optimization models. Other limitations include reliance on data from a limited set of sites, simplifying assumptions in some battery and EVSE models, and a lack of built-in forecasting for loads and generation.	USA	Developing practical solutions for large-scale EV charging	Smart charging with MPC reduced infrastructure needs by delivering 99.8% of demand at 0.234 $/kWh using a 200 kW transformer, compared to uncontrolled charging requiring 685 kW at 0.351 $/kWh.
[60]	The V2G simulator, developed using MATLAB GUI (R2025b) and the MATPOWER engine (Version 8.1), undertakes power flow analysis and generates the charge or discharge patterns. It evaluates the impact of integrating EVs into the power system with or without V2G scenarios. Limitations include simplified assumptions: identical EV battery capacities (60 kWh), fixed travel patterns for weekdays, and a robust communication system between EVs and the control center (idealized).	UK & China	Determining the anticipated patterns of battery cycling associated with driving and V2G operation for frequency support, peak shaving, etc. Investigating the communication and control of temporal and physical information requirements from the battery management system to the grid control system Demonstrating V2G operation	The applied V2G control strategy reduced overall EV charging costs by 13.6% compared with the charging-only scenario, while still meeting minimum state-of-charge requirements for daily travel.

Table 4. Summary of research on V2G/V2B simulators for EV charging behaviors and load demand response of an EV fleet.

Source	Study Description and Limitations	Locations/Region	Applications	Success Metrics/Results
[53]	This paper introduces EV2Gym, an open-source and modular simulation platform. Limitations include the partial modeling of power network impacts compared with full-grid simulators, reliance on predefined datasets for EV behavior, and potential simplifications in charger and transformer interactions under large-scale deployment scenarios.	Netherlands/ Italy	EV Smart charging research and benchmarking Cost Reduction	EV2Gym was validated against real charging data, showing that its two-stage charging/discharging model closely matched laboratory curves for AC and DC charging.
[54]	This paper proposes a novel RL framework for V2B charging, combining DDPG with action masking and MILP-based policy guidance. Limitations include reliance on accurate long-term peak power estimation, computational intensity for training, and assumptions of centralized control across 15 chargers, which may limit scalability in highly decentralized or competitive settings.		Building Energy Cost Reduction Peak Shaving & Valley Filling Scalable Fleet Management Policy & Tariff Testing	The approach consistently outperformed heuristic baselines and closely matched the MILP oracle, reducing monthly building energy bills and peak demand charges.
[55]	This thesis investigates RL approaches for controlling Virtual Power Plants (VPPs) composed of EVs, renewable sources (PV and wind), and household loads. Limitations include reliance on synthetic charging data generated by the Elvis simulator (not fully real-world calibrated) and limited scalability beyond the modeled microgrid (4 households, 4 EVSEs, 40 PV units, and 8 wind turbines).	Germany	Peak Shaving & Valley Filling Self-Consumption Optimization Grid Independence VPPs Scenario Testing	The proposed approach achieved the best performance, delivering effective peak shaving, valley filling, and a stable SoC at departure.
[56]	A cloud-based simulator comprises multiple subsystems, including a real-time power system simulator, an EV charge scheduler system, and a smart energy plaza with various charging stations for large-scale EV charging. However, its validation remains limited to controlled test cases and simulated EV fleets, meaning real-world uncertainties—such as communication delays, heterogeneous user behaviors, and scalability challenges with extreme fast charging—are not yet fully addressed.	USA	Testing of V2G technologies	Results showed effective grid constraint management and frequency regulation in a 460-EV simulation, ensuring SoC targets while avoiding overloads.
[57]	V2G-Sim models the driving and charging patterns of individual plug-in EVs (PEVs) and generates detailed predictions of grid impacts and opportunities over time and space, based on increased PEV adoption. The simulator is scalable and can analyze any number of vehicles, from a single unit to over a million PEVs. Limitations include strong reliance on assumptions about EV adoption rates, charging infrastructure availability, and user charging behavior.	USA	Demand response from PEVs PEVs For Renewables Integration Predicting PEV Charging Demands Spatial Resolution Electricity pricing schemes and their impact on the grid	V2G-controlled charging effectively reduced peak loads and mitigated net load ramps compared to uncontrolled charging.
[58]	The EVs’ user charging behavior simulator consists of two main parts: data generation and trip simulation. Data is generated based on each user’s profile, which includes vehicle details, planned trips (with locations and departure times), and behavioral parameters. While primarily designed for unidirectional charging, the framework can be extended to V2G scenarios, which limits its direct applicability for grid services. Other limitations include simplified assumptions such as user profiles not varying across weekdays, and charging decisions driven mainly by SOC thresholds.	Portugal	Testing the impact of different types of incentives on EV users’ behavior Evaluating the impact of the variation in electricity prices on the behavior of EV users Generating data and scaling it for regional application	For 2500 EVs, variable charging prices yielded user cost reductions of 4% (distance preference), 10% (price preference), and 2% (time preference) compared to fixed prices.
[59]	ACN-Sim is an open-source simulator that offers a modular, extensible architecture, capturing the complexity of real charging systems, including battery charging and user interaction behaviors. Although originally designed for unidirectional charging, ACN-Sim’s modular architecture supports extensions to V2G by adapting charging profiles and optimization models. Other limitations include reliance on data from a limited set of sites, simplifying assumptions in some battery and EVSE models, and a lack of built-in forecasting for loads and generation.	USA	Developing practical solutions for large-scale EV charging	Smart charging with MPC reduced infrastructure needs by delivering 99.8% of demand at 0.234 $/kWh using a 200 kW transformer, compared to uncontrolled charging requiring 685 kW at 0.351 $/kWh.
[60]	The V2G simulator, developed using MATLAB GUI (R2025b) and the MATPOWER engine (Version 8.1), undertakes power flow analysis and generates the charge or discharge patterns. It evaluates the impact of integrating EVs into the power system with or without V2G scenarios. Limitations include simplified assumptions: identical EV battery capacities (60 kWh), fixed travel patterns for weekdays, and a robust communication system between EVs and the control center (idealized).	UK & China	Determining the anticipated patterns of battery cycling associated with driving and V2G operation for frequency support, peak shaving, etc. Investigating the communication and control of temporal and physical information requirements from the battery management system to the grid control system Demonstrating V2G operation	The applied V2G control strategy reduced overall EV charging costs by 13.6% compared with the charging-only scenario, while still meeting minimum state-of-charge requirements for daily travel.

A unifying framework for V2G and V2B simulators integrates EV charging behavior, grid interaction, and demand response strategies into a modular and scalable environment (Figure 6). The framework begins with real-world or synthetic input data (EV sessions, traffic patterns, building loads, renewable generation, and price signals). These inputs feed into simulation engines (e.g., ACN-Sim, EV2Gym) that model EVSE infrastructure, grid constraints, and user heterogeneity. A behavioral modeling layer captures driver preferences, charging profiles, and probabilistic methods (e.g., KDE, GMM), while control and optimization modules apply algorithms such as MPC, heuristic scheduling, or RL (DQN, DDPG, MARL). The V2G/V2B interface manages bidirectional energy flows, accounting for battery degradation, transformer limits, and building demand response needs. Outputs include aggregated load profiles, cost and carbon reduction estimates, peak shaving/valley filling performance, and infrastructure planning insights. Finally, an evaluation layer validates results against benchmarks using metrics such as MAPE, RMSE, peak-to-average ratio, cost savings, and grid stability indices. This integrated framework enables researchers and operators to compare algorithms, optimize demand response, and assess long-term system impacts at both the building and regional levels.

Several traditional methods are used to estimate the regional energy demand of an EV fleet. These include the mixed-integer linear optimization model, the Monte Carlo method, probabilistic forecasts, and the empirical method (Table 5). While these techniques can support utilities in system planning and forecasting EV charging demand, they exhibit notable limitations in dynamic and data-rich environments. Specifically, traditional models often struggle to capture the nonlinear relationships and non-stationary behavior inherent in EV charging patterns, which can lead to significant deviations from actual demand. Moreover, they tend to underutilize detailed fleet-level characteristics and the influence of exogenous factors like weather conditions. In many cases, they also rely on simplified assumptions about user behavior (e.g., fixed travel routines, constant charging rates) and region-specific data that may not generalize across diverse markets. As EV adoption accelerates and usage patterns become increasingly diverse and uncertain, these limitations reduce the predictive accuracy and flexibility of conventional approaches, highlighting the need for more adaptive, data-driven forecasting frameworks.

Table 5. Summary of research on traditional methods for regional energy development.

Source	Study Description	Locations/Region	Models	Success Metrics/Results
[61]	This study develops the robust charging location problem for electric bus networks, expanding the traditional charging location problem to consider charger failures.	Ireland & Spain	Mixed Integer & Large Neighborhood Search (LNS)	The LNS framework consistently outperformed CPLEX, finding solutions with 43% to 97% fewer charging stations in large-scale instances while ensuring robustness.
[62]	This study develops a multiscale model for allocating EV infrastructure to meet residents’ and visitors’ demand for EVs at night and during the day.	Netherlands & Spain	Mixed Integer Linear Optimization Model	The model shows that more than 53% of EVs can charge at their daily destinations under optimized allocation. Under the 40% scenario, the load is distributed evenly across towns, while the 80% scenario concentrates the load in fewer, strategically located areas.
[63]	This paper predicts the charging load of regional EVs based on battery capacity.	China	Improved Monte Carlo method	The improved method accelerates simulations by 7–12 times compared to serial MC.
[64]	This study considers several mobility features of individual users when forecasting the next-day energy demand of individual EVs.	Switzerland & Austria	Probabilistic forecasts	Incorporating mobility features (entropy, radius of gyration, location frequency) improved next-day demand forecasting accuracy over baseline models. Improving the prediction of next-day individual EV energy demand.
[65]	This study suggests regional public EV charging stations based on demand variations among cities and differences in station utilization.	China	Empirical method	Results show that cities Linfen and Lvliang have the highest short- and long-term demand, while residential-area stations have the highest utilization, and workplace stations the lowest.
[66]	This article presents a framework for evaluating load-shifting strategies to reschedule EV charging to lower grid load periods.	Canada	Probabilistic model	The proposed approach achieved significant peak-to-average ratio reduction while avoiding rebound effects, demonstrating grid relief with low implementation cost.
[67]	This work forecasts regional EV charging loads and analyzes their impact on the power grid’s peak–valley difference.	China	Monte Carlo method	The fast-charging expansion increases the peak–valley difference in grid load, but its impact on overall load is relatively limited.
[68]	This study explores hierarchical probabilistic EV load forecasting at regular charging stations.	Slovakia, Italy & Netherlands	Probabilistic method	The proposed method improved forecasting skill by up to 9.5% compared with non-hierarchical approaches, validated on real EV charging data.
[69]	This study proposes a coordinated control of a building cluster with both energy sharing and EV charging.	Sweden	Genetic Algorithm	The coordinated control improved renewable self-consumption by 19% and reduced daily electricity bills by 36% compared to conventional controls.
[70]	This research addresses energy demand at the regional or national levels.	Netherlands, & Australia	Agent-based modeling	Household behavioral changes could cumulatively cut regional energy demand by up to 30% over the modeled period to maximize a region’s emissions reduction potential.
[71]	This work analyzes the regional EV infrastructure to provide guidance on charging infrastructure to regional stakeholders through the U.S. DOE’s Vehicle Technologies Office.	USA	EVI-Pro modeling approach	The simulation estimated that a 57% annual growth in PEV sales will be needed by 2025 to meet projected EV adoption, guiding infrastructure planning.
[72]	This paper presents a time-spatial EV charging power demand forecast model at fast charging stations located in urban areas.	South Korea & Philippines	Markov-chain traffic model and a teleportation approach	The average charging power demand in the “all fully charged” scenario was over 2.5 times higher than in the random charging profile scenario.
[73]	This work proposes a hybrid algorithm for finding the optimal placement of a charging station in a distribution system.	India, South Africa & Denmark	Genetic algorithm and particle swarm optimization	The hybrid algorithm demonstrated improved voltage profiles and reduced power losses compared to standard GA or PSO approaches in charging station siting.
[74]	This paper presents a load prediction method for the conventional electrical load and the charging demand of EV parking lots simultaneously.	China & USA	Autoregressive integrated moving average	The decoupled ARIMA forecaster reduced mean square error compared to integrated approaches and enhanced day-ahead scheduling under stochastic operation scenarios.

Table 5. Summary of research on traditional methods for regional energy development.

Source	Study Description	Locations/Region	Models	Success Metrics/Results
[61]	This study develops the robust charging location problem for electric bus networks, expanding the traditional charging location problem to consider charger failures.	Ireland & Spain	Mixed Integer & Large Neighborhood Search (LNS)	The LNS framework consistently outperformed CPLEX, finding solutions with 43% to 97% fewer charging stations in large-scale instances while ensuring robustness.
[62]	This study develops a multiscale model for allocating EV infrastructure to meet residents’ and visitors’ demand for EVs at night and during the day.	Netherlands & Spain	Mixed Integer Linear Optimization Model	The model shows that more than 53% of EVs can charge at their daily destinations under optimized allocation. Under the 40% scenario, the load is distributed evenly across towns, while the 80% scenario concentrates the load in fewer, strategically located areas.
[63]	This paper predicts the charging load of regional EVs based on battery capacity.	China	Improved Monte Carlo method	The improved method accelerates simulations by 7–12 times compared to serial MC.
[64]	This study considers several mobility features of individual users when forecasting the next-day energy demand of individual EVs.	Switzerland & Austria	Probabilistic forecasts	Incorporating mobility features (entropy, radius of gyration, location frequency) improved next-day demand forecasting accuracy over baseline models. Improving the prediction of next-day individual EV energy demand.
[65]	This study suggests regional public EV charging stations based on demand variations among cities and differences in station utilization.	China	Empirical method	Results show that cities Linfen and Lvliang have the highest short- and long-term demand, while residential-area stations have the highest utilization, and workplace stations the lowest.
[66]	This article presents a framework for evaluating load-shifting strategies to reschedule EV charging to lower grid load periods.	Canada	Probabilistic model	The proposed approach achieved significant peak-to-average ratio reduction while avoiding rebound effects, demonstrating grid relief with low implementation cost.
[67]	This work forecasts regional EV charging loads and analyzes their impact on the power grid’s peak–valley difference.	China	Monte Carlo method	The fast-charging expansion increases the peak–valley difference in grid load, but its impact on overall load is relatively limited.
[68]	This study explores hierarchical probabilistic EV load forecasting at regular charging stations.	Slovakia, Italy & Netherlands	Probabilistic method	The proposed method improved forecasting skill by up to 9.5% compared with non-hierarchical approaches, validated on real EV charging data.
[69]	This study proposes a coordinated control of a building cluster with both energy sharing and EV charging.	Sweden	Genetic Algorithm	The coordinated control improved renewable self-consumption by 19% and reduced daily electricity bills by 36% compared to conventional controls.
[70]	This research addresses energy demand at the regional or national levels.	Netherlands, & Australia	Agent-based modeling	Household behavioral changes could cumulatively cut regional energy demand by up to 30% over the modeled period to maximize a region’s emissions reduction potential.
[71]	This work analyzes the regional EV infrastructure to provide guidance on charging infrastructure to regional stakeholders through the U.S. DOE’s Vehicle Technologies Office.	USA	EVI-Pro modeling approach	The simulation estimated that a 57% annual growth in PEV sales will be needed by 2025 to meet projected EV adoption, guiding infrastructure planning.
[72]	This paper presents a time-spatial EV charging power demand forecast model at fast charging stations located in urban areas.	South Korea & Philippines	Markov-chain traffic model and a teleportation approach	The average charging power demand in the “all fully charged” scenario was over 2.5 times higher than in the random charging profile scenario.
[73]	This work proposes a hybrid algorithm for finding the optimal placement of a charging station in a distribution system.	India, South Africa & Denmark	Genetic algorithm and particle swarm optimization	The hybrid algorithm demonstrated improved voltage profiles and reduced power losses compared to standard GA or PSO approaches in charging station siting.
[74]	This paper presents a load prediction method for the conventional electrical load and the charging demand of EV parking lots simultaneously.	China & USA	Autoregressive integrated moving average	The decoupled ARIMA forecaster reduced mean square error compared to integrated approaches and enhanced day-ahead scheduling under stochastic operation scenarios.

Beyond architectural descriptions, the performance of DL architectures such as LSTM, GRU, and BiLSTM depends heavily on training complexity, hyperparameter sensitivity, and computational scalability—challenges that become pronounced when dealing with large EV–renewable datasets combining high-resolution charging, mobility, and renewable generation data. These recurrent networks require careful tuning of hidden layers, learning rates, and sequence lengths to maintain stable convergence under non-stationary temporal dynamics and heterogeneous spatial inputs. For large-scale datasets, LSTM and BiLSTM models require significant computational resources due to their long-term memory retention and, in the case of BiLSTM, bidirectional processing. In contrast, GRUs achieve faster convergence with lower computational overhead, though sometimes at the expense of reduced temporal modeling depth. In RL applications, scalability is further limited by reliance on synthetic datasets, uniform battery-capacity assumptions, and simplified communication topologies that ignore latency, network congestion, and interoperability across heterogeneous chargers and grid interfaces. These limitations underscore the need for benchmark-scale hybrid frameworks capable of efficiently training on large EV–renewable datasets while integrating real-world variability, system coupling, and multi-agent coordination for more robust and transferable energy management solutions.

4. EV Overnight and Workplace Charging Strategy in the Residential Sector

Over the past few years, scientists have proposed various solutions, including overnight charging, workplace charging, or a combination of these two approaches, to alleviate the increase in peak electricity consumption, reduce the need for network upgrades, and ultimately benefit the consumer [66,75]. Future EV charging strategies, if uncontrolled, may lead to service transformer overloading or voltage violations. According to Cell Reports Physical Science, EV charging can reduce peak electricity demand and store solar energy if managed properly [75]. The study suggests that delayed home and workplace charging are effective strategies to achieve these benefits. Traditional TOU and critical peak pricing schemes were evaluated in residential buildings to explore how price signals could alter consumption patterns [76]. They found evidence of a remarkable time-shifting of charging among 62 households with a PEV based on time-of-use tariffs. The energy consumed by PEV charging during the nighttime increased from 32 percent prior to the pricing experiment to 55 percent during the months with TOU pricing.

Workplace charging can support drivers with long commutes or limited home charging access, while also aligning EV demand with periods of high renewable generation, thus reducing curtailment. Compared to home-dominant strategies, workplace charging can lower evening peaks by offsetting residential demand. Although EV charging strategies are widely studied, load modeling often relies on statistical methods. Recently, ML approaches have emerged to predict load and optimize charging, particularly for reducing peak demand. Table 6 shows a collection of studies focused on overnight and workplace EV charging in residential buildings.

Table 6. Summary of research on model developments for EV overnight and workplace charging strategies.

Source	Study Description and Limitations	Locations /Region	Overnight & Workplace Charging	Models	Success Metrics/Results
[77]	This study simulates charging infrastructure needs with detailed individual characteristics, including dwelling types and activity patterns. Synthetic Swedish data restricts the ability to generalize to other countries.	Sweden	Overnight & Workplace	Agent-based decision support framework	Plan-ahead/event-triggered charging required 2.3–4.5 times more chargers than liquid fuel model.
[78]	This study proposes a multi-objective optimization model to determine the optimal charging infrastructure for transitioning to workplace PEVs. Optimization assumptions might not fully account for real-world uncertainties.	UK, USA, Ireland	Workplace	linear programming approach	Improved trade-offs between cost minimization and coverage in station allocation.
[79]	This study analyzed charging data from a real-world pilot program to evaluate the effectiveness of various smart charging use cases. Limited to a single manufacturer, which reduces its representativeness.	USA	Overnight & Workplace	Smart-charging strategy	Smart charging shifted 15–20% of the load out of peak periods and 20–30% into low-cost periods.
[66]	This article presents a framework to evaluate load shifting strategies to reschedule the EV charging to lower grid load periods. It focuses on a winter-only Quebec case with the assumption of one EV per household.	Canada	Overnight	Probabilistic distribution method	EV overnight charging strategy reduced the peak-to-average ratio by 21% in high-power and battery-size-dominant scenarios and avoided rebound effects.
[80]	This paper quantifies the value of managed charging under a 50% RE grid and PEV adoption scenarios up to California’s 5 million vehicle target. Calibrated to Bay Area 2016 data, excluding Tesla vehicles, may introduce bias in 2025 projections.	USA	Overnight & Workplace	PLEXOS grid simulator	Managed charging reduced costs by $120–690 M and curtailment by up to 40%.
[81]	This paper uses driving patterns from the National Household Travel Survey to simulate workplace charging for parking structures under various charging scenarios. Validated only in simulations with stylized assumptions.	USA	Workplace	Smart-charging strategy	The decentralized protocol balanced the load and reduced peaks, comparable to centralized control.
[82]	This study implements and compares various control schemes for workplace charging to minimize the transformer’s aging. Assumes full knowledge of arrival and departure times.	USA	Workplace	Transformer modeling	Controlled charging reduced costs and supported 67% more vehicles compared to uncontrolled charging.
[83]	This paper investigates the impacts of residential EV charging on a distribution grid. Calibration of a single test grid limits its broader applicability.	Sweden	Overnight	Stochastic charging model	High penetration raised transformer aging and loading; the stochastic model outperformed the deterministic.
[84]	This paper suggests three charging strategies to decrease total charging during peak periods. Validated with Canadian datasets; uncertain for other climates or regions.	Canada	Overnight & Workplace	ML approach	The Overnight and Workplace strategy reduced the peak-to-average ratio by approximately 50% and lowered costs by 54–56% in winter and spring.

Table 6. Summary of research on model developments for EV overnight and workplace charging strategies.

Source	Study Description and Limitations	Locations /Region	Overnight & Workplace Charging	Models	Success Metrics/Results
[77]	This study simulates charging infrastructure needs with detailed individual characteristics, including dwelling types and activity patterns. Synthetic Swedish data restricts the ability to generalize to other countries.	Sweden	Overnight & Workplace	Agent-based decision support framework	Plan-ahead/event-triggered charging required 2.3–4.5 times more chargers than liquid fuel model.
[78]	This study proposes a multi-objective optimization model to determine the optimal charging infrastructure for transitioning to workplace PEVs. Optimization assumptions might not fully account for real-world uncertainties.	UK, USA, Ireland	Workplace	linear programming approach	Improved trade-offs between cost minimization and coverage in station allocation.
[79]	This study analyzed charging data from a real-world pilot program to evaluate the effectiveness of various smart charging use cases. Limited to a single manufacturer, which reduces its representativeness.	USA	Overnight & Workplace	Smart-charging strategy	Smart charging shifted 15–20% of the load out of peak periods and 20–30% into low-cost periods.
[66]	This article presents a framework to evaluate load shifting strategies to reschedule the EV charging to lower grid load periods. It focuses on a winter-only Quebec case with the assumption of one EV per household.	Canada	Overnight	Probabilistic distribution method	EV overnight charging strategy reduced the peak-to-average ratio by 21% in high-power and battery-size-dominant scenarios and avoided rebound effects.
[80]	This paper quantifies the value of managed charging under a 50% RE grid and PEV adoption scenarios up to California’s 5 million vehicle target. Calibrated to Bay Area 2016 data, excluding Tesla vehicles, may introduce bias in 2025 projections.	USA	Overnight & Workplace	PLEXOS grid simulator	Managed charging reduced costs by $120–690 M and curtailment by up to 40%.
[81]	This paper uses driving patterns from the National Household Travel Survey to simulate workplace charging for parking structures under various charging scenarios. Validated only in simulations with stylized assumptions.	USA	Workplace	Smart-charging strategy	The decentralized protocol balanced the load and reduced peaks, comparable to centralized control.
[82]	This study implements and compares various control schemes for workplace charging to minimize the transformer’s aging. Assumes full knowledge of arrival and departure times.	USA	Workplace	Transformer modeling	Controlled charging reduced costs and supported 67% more vehicles compared to uncontrolled charging.
[83]	This paper investigates the impacts of residential EV charging on a distribution grid. Calibration of a single test grid limits its broader applicability.	Sweden	Overnight	Stochastic charging model	High penetration raised transformer aging and loading; the stochastic model outperformed the deterministic.
[84]	This paper suggests three charging strategies to decrease total charging during peak periods. Validated with Canadian datasets; uncertain for other climates or regions.	Canada	Overnight & Workplace	ML approach	The Overnight and Workplace strategy reduced the peak-to-average ratio by approximately 50% and lowered costs by 54–56% in winter and spring.

Across the above studies, model performance depends strongly on context, data richness, and behavioral realism. Simulation and optimization models (e.g., [78,81]) perform well in controlled settings; however, their assumptions limit their applicability when user heterogeneity and uncertainty are high. Probabilistic and stochastic models (e.g., [77,83]) capture variability more effectively, although they require extensive input data and often overestimate infrastructure needs. Machine learning and hybrid forecasting approaches (e.g., [84]) excel in data-rich environments, offering improved predictive accuracy and enabling adaptive charging strategies that reduce peaks and costs. Region- and season-specific strategies (e.g., [66]) demonstrate that contextual factors such as climate, urban environment, and charging culture significantly influence outcomes, highlighting the need for adaptive frameworks tailored to local conditions.

To connect the diverse methodologies reviewed above, this study introduces a unifying conceptual framework (Figure 3, Figure 4, Figure 5 and Figure 6) that organizes the reviewed approaches into a hierarchical, multi-layered architecture linking forecasting, optimization, and policy coupling. The framework captures how supervised, unsupervised, and reinforcement learning models—together with V2G/V2B simulations—interact across stages of data acquisition, feature extraction, model development, optimization-based control, and evaluation.

Forecasting defines predictive baselines for load, renewable generation, and user behavior; optimization transforms these forecasts into real-time scheduling and control decisions; and policy coupling ensures alignment with regulatory, economic, and social objectives. Collectively, these layers illustrate how deep learning–based forecasting integrates with optimization-driven coordination to enable hierarchical decision-making across building, transport, and grid domains. This framework provides the conceptual foundation for the comparative and critical synthesis presented in Section 5.

5. Integrated Frameworks for EVs, Renewable-Buildings Energy Systems: Towards Decarbonization Pathways and Smart Energy Management

Section 5 presents a critical synthesis of integrated frameworks that connect EVs, RE, energy storage, and thermal systems within residential buildings to advance decarbonization and intelligent energy management. Building on the forecasting and control concepts discussed in previous sections, this part examines how data-driven, hybrid, and optimization-based models interact across the domains of prediction, scheduling, and system coupling. The analysis highlights that forecasting accuracy and energy management efficiency are jointly determined by data richness, spatial–temporal granularity, and the depth of cross-domain integration—demonstrating how hybrid learning and coordinated control can accelerate the transition toward low-carbon, grid-interactive neighborhoods.

Integrating EV charging infrastructure with RESs in residential buildings is increasingly recognized as a cornerstone of sustainable energy management. Expanding upon the analytical framework outlined at the end of the previous section, the reviewed studies demonstrate that hybrid and optimization-based approaches—which combine forecasting accuracy with adaptive control—consistently outperform single-model frameworks in addressing the variability of EV charging demand, renewable generation, and electrified building loads. This convergence underscores the importance of linking predictive accuracy with operational scheduling and real-time coordination across data, storage, and control layers. As illustrated in Figure 3, Figure 4, Figure 5 and Figure 6, coordinated and hybrid strategies not only enhance system efficiency and resilience but also deliver measurable benefits in cost reduction, peak shaving, and renewable integration, while highlighting ongoing challenges of scalability, behavioral diversity, and regional generalizability.

Table 7 extends this synthesis by consolidating research on the multi-domain integration of EVs, RESs, heat pumps, and energy storage in residential and community-scale infrastructures. These studies collectively explore all-electric residential systems, community energy solutions, and the coupling of EVs with public and institutional building energy systems, emphasizing coordinated operation and optimization of flexible loads. These investigations span several European countries, with a primary focus on minimizing operational costs, reducing CO₂ emissions, and enhancing the overall efficiency and sustainability of energy systems.

Innovations in this field leverage ML-based optimization and hybrid control frameworks to enhance the coordination of EVs, photovoltaic generation, and thermal and electricity energy storage, thereby enabling greater demand flexibility and renewable penetration. Together, these integrated approaches demonstrate that coupling predictive analytics with coordinated control provides a scalable pathway for residential decarbonization while ensuring grid stability and occupant comfort. A key insight emerging from the reviewed studies is the geographic imbalance in analytical depth and data coverage. While most hybrid forecasting and energy management frameworks are validated in data-rich regions such as Europe and East Asia, research from Africa and Latin America remains underrepresented due to constrained data access and uneven research funding. This disparity limits the transferability of forecasting models and policy recommendations to emerging markets, where grid dynamics, renewable variability, and user behaviors differ markedly. Strengthening regional data infrastructure and funding partnerships is therefore critical to scaling EV-renewable integration in the Global South and ensuring equitable decarbonization pathways.

Table 7. Summary of research on integrating EV charging infrastructure with RESs in residential buildings.

Source	Study Description	Locations/Region	Models	Success Metrics: Appropriate Quantifiable Metrics for Optimization
[85]	This paper investigates the feasibility and design of a building-integrated photovoltaic-powered EV charging system in a typical house using solar energy to meet residential and EV charging demand. The study relies on case-specific simulation scenarios and performance indicators, but does not incorporate variability in user charging patterns, seasonal fluctuations, or large-scale deployment considerations.	Malaysia, India	PVsyst 7.2 (Student version)	The best configuration achieved a performance ratio of 79.78% and a capacity utilization factor of 16.4%, outperforming other tested configurations.
[86]	This study analyzes the charging infrastructure for BEVs in residential buildings. The analysis assumes stable regulatory, tariff, and technical conditions, which may not accurately represent the diverse multi-apartment contexts found elsewhere.	Austria	Load management approach	Dynamic load management reduced peak demand significantly for BEV charging while maintaining high user satisfaction.
[87]	This report introduces the EV charging tool, which was developed under the GEF-7 global program to support countries in shifting to electric mobility. The tool is built on predefined charging profiles and simplified managed charging scenarios.	IEA	EV charging tool	The tool provides weekly demand profiles, managed vs. unmanaged charging comparisons, and CO2 impact estimates, supporting decision-making at system and distribution levels.
[88]	This paper investigates a method involving several strategies to stabilize the grid system and examines the impact of various types of EVs and heat pumps for supplying heat in buildings. However, the limitation is that the approach relies on scenario-driven analysis under uncertain renewable energy variability.	UK	Smart charging algorithm	Findings indicate that the V2H strategy can reduce a household’s carbon footprint by up to 87% and recover ~21.9 kWh/day of surplus renewable energy.
[89]	This paper proposes a two-stage energy management strategy to enhance the flexibility of energy communities with solar PV, EVs, heat pumps, and thermal energy storage systems. The study does not address wider uncertainties in EV user driving patterns and long-term policy effects.	Sweden	Numerical simulations with the Power Factory DigSILENT	The proposed approach reduced operating costs and grid stress, while improving flexibility and preventing load-shedding compared to non-predictive strategies.
[90]	This study presents optimal energy management for controlling energy flow in the smart home that contains photovoltaic generation, integrated with ESS and EV. The study evaluates only two heuristic algorithms and does not compare them against broader optimization or real-world validation scenarios, which limits their generalizability.	China, Sudan	Jaya algorithm and PSO	The Jaya algorithm outperformed PSO in reducing daily electricity costs while reliably meeting both household demand and EV trip requirements.
[91,92]	This paper reviews the impacts of EV charging on 10 distribution electric power systems in residential, commercial, industrial, and mixed-use buildings. The analysis assumes consistent customer responses to TOU price signals and does not capture broader behavioral variability or regional heterogeneity.	USA	Open DSS	The results showed that synchronized off-peak charging can increase peak demand by up to 20%, but randomized charging within off-peak reduced peaks by ~5%.
[93]	This paper proposes an optimal charge schedule for EVs in solar-powered charging stations based on day-ahead forecasting of solar power generation. The scheduling framework is applied to a single PV-powered station in Australia, with limited consideration of scaling or broader network impacts.	Australia	ANN	The ANN-based model reached 99.6% accuracy in solar forecasting, and optimal scheduling with and without PV cut charging costs by 50–100% and 10–20%, respectively.
[94]	This work analyzes highly resolved residential electricity consumption data of Austrian, German, and UK households and proposes an applicable data-driven load model. The proposed models are derived from specific datasets and may not represent load dynamics in regions with different appliance mixes or usage habits.	Germany, UK, & Norway	Stochastic model	The proposed load profiles accurately reproduced high-resolution consumption patterns, outperforming the traditional approach in representing demand fluctuations.
[95]	This work combines a framework of the transportation, building, and electricity sectors to show the operational impacts of demand-side flexibility on both the demand and supply sides of the energy system. The integrated modeling framework is region-specific and does not account for uncertainty in policy adoption or cross-sector scaling to larger jurisdictions.	Canada	Primal-dual interior-point method	The study showed that Regina, Canada, could reach nearly 100% renewable integration by 2050 through rooftop solar, wind, demand response, and storage, staying within 1% of their renewable goal.
[96]	This study proposes an energy-sharing framework that utilizes vehicle-to-grid transfer in coordination with demand response programs for residential buildings. It was tested on a limited residential load profile, assuming ideal IoT-based coordination and predefined travel distances, without undergoing large-scale field validation.	Iran, Vietnam, Denmark, & Qatar	Load management softwareunder the IoT technology	The results showed that the proposed framework effectively reduced peak load and lowered total electricity consumption.
[97]	This work develops an integrated model for the electrical, thermal, and mobility loads of private households. The model is calibrated primarily for German households using survey-based mobility and activity data, which constrains transferability to other regions.	Germany	Modified Markov Chain Process	The integrated model generated consistent load, heating, and mobility profiles, with simulated household electricity use of 2751 kWh/year and vehicle travel distributions.
[98]	This paper proposes two electric energy management systems (EMS) for a grid-connected residential neighborhood, including EVs, battery storage, and solar photovoltaic generation. The EMS strategies were tested in simulation for a single high-density residential case, limiting validation across diverse housing typologies.	Australia	Centralized/Decentralized EMS	Centralized EMS achieved greater cost savings and grid energy reduction compared to decentralized EMS, while accounting for battery degradation costs.
[99]	This study assesses the impact of thermal storage, air heat pumps, and EVs on residential building load. It relies on case-specific assumptions for Swedish clusters and uses a simplified EV load generator for a 24 kWh battery capacity, which limits its generalizability.	Sweden	Genetic algorithm optimization approach	The coupled PV-heat pump-thermal storage-EV system improved PV self-consumption to ~77%, with techno-economic feasibility achieved under shared energy scenarios.
[100]	This paper proposes two system designs, one for home energy storage and the other for community energy storage. The optimization model was validated using a single Swiss community case, where current battery costs render both home and community energy storage economically unfeasible.	Netherlands & Switzerland	Mixed integer linear	Community Energy Storage (CES) outperformed Home Energy Storage (HES) in both cost and CO2 reduction, with sensitivity analysis indicating feasibility under a larger battery storage capacity.
[101]	This article uses highly resolved models of residential power demand and PEV use to assess the impact of uncoordinated in-home PEV charging on residential power demand. The study does not implement smart charging or coordination strategies, relying instead on uncoordinated charging scenarios that limit exploration of demand management solutions.	USA	Stochastic model	Results show that while total electricity consumption increases modestly (~5% for 10% EV share), uncoordinated charging significantly reshapes demand profiles, exacerbates peak loads, and stresses distribution infrastructure at even low EV penetration.
[102]	This work proposes a new prediction model for aggregated loads of buildings and EVs. It remains limited by its reliance on historical data patterns, which may not capture sudden behavioral or policy shifts.	China, US, Iran, Iraq, & Hungary	Enhanced support vector machine	The model reduced forecasting error compared to traditional methods, achieving lower MAPE and RMSE across short-term load prediction scenarios.
[103]	This work provides a novel method for the optimal simultaneous allocation and sizing of RES and EV charging stations and for managing the vehicle charging process. It assumes simplified EV behavior and uniform adoption patterns, which may not reflect diverse user charging habits and stochastic uncertainties of real-world EV fleets.	Iran, USA, & China	Genetic Algorithm-Particle Swarm Optimization	The improved GA-PSO algorithm reduced power losses, voltage fluctuations, and costs more effectively than benchmark methods (DE2), proving its efficiency in IEEE 33-bus simulations.
[104]	This study simulates several potential grid scenarios and resulting demand response priorities. The pilot was limited to only 30 participants and a small regional context, restricting generalization to larger, more heterogeneous EV populations.	Canada	Cloud-based algorithm	The program demonstrated technical feasibility of utility-controlled charging, with curtailments up to 27 kW, while maintaining driver satisfaction and meeting state-of-charge requirements.
[105]	This work develops a decentralized algorithm that minimizes communication and delay to address both grid-level concerns and local-level issues. The proposed algorithm simplifies communication and computation but does not account for broader stochastic variations in EV usage and regional grid diversity.	USA	Decentralized algorithm	Coordinated strategies reduced transformer hot spot temperatures and equivalent aging factors compared to uncoordinated charging, thereby extending transformer life.
[106]	This paper discusses implementing smart charging algorithms to directly control EV charging rates and starting times at residential locations. The framework relies on static TOU and controlled charging strategies, which may not adapt to evolving real-time price signals or heterogeneous user behavior.	USA	Open DSS	Simulation results showed that optimized TOU and smart charging reduced transformer overload and feeder voltage deviations compared to uncontrolled charging.
[107]	This article presents an optimal smart charging algorithm that enables the charging of large numbers of vehicles without adverse effects on the electricity network at residential locations. The optimization relies on linear approximations of voltage drop and assumes the availability of complete network state information, which may limit accuracy under complex or large-scale systems.	Australia	Linear approximation	The method enabled high EV penetration (up to 50%) without major network upgrades, while maintaining voltage and line current within operational limits.
[108]	This paper reviews strategies, algorithms, and methods for implementing a smart charging control system. It synthesizes strategies without implementing quantitative validation, limiting direct performance benchmarking.	Spain	Power system analysis software approaches	The review highlights that centralized and decentralized smart charging can both mitigate voltage deviations and transformer overloads, but practical implementation remains constrained by communication and market barriers.
[109]	This study proposes a novel decentralized valley-filling charging strategy using a day-ahead pricing scheme. The model assumes reliable day-ahead pricing signals and user compliance, overlooking real-world variability in participation and device-level autonomy.	China	Multi-objective charging optimization algorithms	The decentralized strategy achieved a valley-filling effect with 28% lower generation cost compared to uncoordinated charging, and showed robustness to parameter uncertainty with less than 2% deviation.
[110]	This paper presents an EV charging method with a PV system for smart homes/buildings. The proposed EV charging algorithm is designed to determine the optimal schedules of EV charging based on predicted PV output and electricity consumption. The optimization framework assumes error-free PV and load forecasts without reporting uncertainty measures, and it omits grid feedback and V2G interactions, thereby limiting its robustness under real-world operating conditions.	South Korea	Mixed Integer Linear	The proposed scheduling method reduced charging costs by 6% compared to immediate charging (baseline #1) and by 15.2% compared to delayed charging (baseline #2), while ensuring the target SOC was met.

Table 7. Summary of research on integrating EV charging infrastructure with RESs in residential buildings.

Source	Study Description	Locations/Region	Models	Success Metrics: Appropriate Quantifiable Metrics for Optimization
[85]	This paper investigates the feasibility and design of a building-integrated photovoltaic-powered EV charging system in a typical house using solar energy to meet residential and EV charging demand. The study relies on case-specific simulation scenarios and performance indicators, but does not incorporate variability in user charging patterns, seasonal fluctuations, or large-scale deployment considerations.	Malaysia, India	PVsyst 7.2 (Student version)	The best configuration achieved a performance ratio of 79.78% and a capacity utilization factor of 16.4%, outperforming other tested configurations.
[86]	This study analyzes the charging infrastructure for BEVs in residential buildings. The analysis assumes stable regulatory, tariff, and technical conditions, which may not accurately represent the diverse multi-apartment contexts found elsewhere.	Austria	Load management approach	Dynamic load management reduced peak demand significantly for BEV charging while maintaining high user satisfaction.
[87]	This report introduces the EV charging tool, which was developed under the GEF-7 global program to support countries in shifting to electric mobility. The tool is built on predefined charging profiles and simplified managed charging scenarios.	IEA	EV charging tool	The tool provides weekly demand profiles, managed vs. unmanaged charging comparisons, and CO2 impact estimates, supporting decision-making at system and distribution levels.
[88]	This paper investigates a method involving several strategies to stabilize the grid system and examines the impact of various types of EVs and heat pumps for supplying heat in buildings. However, the limitation is that the approach relies on scenario-driven analysis under uncertain renewable energy variability.	UK	Smart charging algorithm	Findings indicate that the V2H strategy can reduce a household’s carbon footprint by up to 87% and recover ~21.9 kWh/day of surplus renewable energy.
[89]	This paper proposes a two-stage energy management strategy to enhance the flexibility of energy communities with solar PV, EVs, heat pumps, and thermal energy storage systems. The study does not address wider uncertainties in EV user driving patterns and long-term policy effects.	Sweden	Numerical simulations with the Power Factory DigSILENT	The proposed approach reduced operating costs and grid stress, while improving flexibility and preventing load-shedding compared to non-predictive strategies.
[90]	This study presents optimal energy management for controlling energy flow in the smart home that contains photovoltaic generation, integrated with ESS and EV. The study evaluates only two heuristic algorithms and does not compare them against broader optimization or real-world validation scenarios, which limits their generalizability.	China, Sudan	Jaya algorithm and PSO	The Jaya algorithm outperformed PSO in reducing daily electricity costs while reliably meeting both household demand and EV trip requirements.
[91,92]	This paper reviews the impacts of EV charging on 10 distribution electric power systems in residential, commercial, industrial, and mixed-use buildings. The analysis assumes consistent customer responses to TOU price signals and does not capture broader behavioral variability or regional heterogeneity.	USA	Open DSS	The results showed that synchronized off-peak charging can increase peak demand by up to 20%, but randomized charging within off-peak reduced peaks by ~5%.
[93]	This paper proposes an optimal charge schedule for EVs in solar-powered charging stations based on day-ahead forecasting of solar power generation. The scheduling framework is applied to a single PV-powered station in Australia, with limited consideration of scaling or broader network impacts.	Australia	ANN	The ANN-based model reached 99.6% accuracy in solar forecasting, and optimal scheduling with and without PV cut charging costs by 50–100% and 10–20%, respectively.
[94]	This work analyzes highly resolved residential electricity consumption data of Austrian, German, and UK households and proposes an applicable data-driven load model. The proposed models are derived from specific datasets and may not represent load dynamics in regions with different appliance mixes or usage habits.	Germany, UK, & Norway	Stochastic model	The proposed load profiles accurately reproduced high-resolution consumption patterns, outperforming the traditional approach in representing demand fluctuations.
[95]	This work combines a framework of the transportation, building, and electricity sectors to show the operational impacts of demand-side flexibility on both the demand and supply sides of the energy system. The integrated modeling framework is region-specific and does not account for uncertainty in policy adoption or cross-sector scaling to larger jurisdictions.	Canada	Primal-dual interior-point method	The study showed that Regina, Canada, could reach nearly 100% renewable integration by 2050 through rooftop solar, wind, demand response, and storage, staying within 1% of their renewable goal.
[96]	This study proposes an energy-sharing framework that utilizes vehicle-to-grid transfer in coordination with demand response programs for residential buildings. It was tested on a limited residential load profile, assuming ideal IoT-based coordination and predefined travel distances, without undergoing large-scale field validation.	Iran, Vietnam, Denmark, & Qatar	Load management softwareunder the IoT technology	The results showed that the proposed framework effectively reduced peak load and lowered total electricity consumption.
[97]	This work develops an integrated model for the electrical, thermal, and mobility loads of private households. The model is calibrated primarily for German households using survey-based mobility and activity data, which constrains transferability to other regions.	Germany	Modified Markov Chain Process	The integrated model generated consistent load, heating, and mobility profiles, with simulated household electricity use of 2751 kWh/year and vehicle travel distributions.
[98]	This paper proposes two electric energy management systems (EMS) for a grid-connected residential neighborhood, including EVs, battery storage, and solar photovoltaic generation. The EMS strategies were tested in simulation for a single high-density residential case, limiting validation across diverse housing typologies.	Australia	Centralized/Decentralized EMS	Centralized EMS achieved greater cost savings and grid energy reduction compared to decentralized EMS, while accounting for battery degradation costs.
[99]	This study assesses the impact of thermal storage, air heat pumps, and EVs on residential building load. It relies on case-specific assumptions for Swedish clusters and uses a simplified EV load generator for a 24 kWh battery capacity, which limits its generalizability.	Sweden	Genetic algorithm optimization approach	The coupled PV-heat pump-thermal storage-EV system improved PV self-consumption to ~77%, with techno-economic feasibility achieved under shared energy scenarios.
[100]	This paper proposes two system designs, one for home energy storage and the other for community energy storage. The optimization model was validated using a single Swiss community case, where current battery costs render both home and community energy storage economically unfeasible.	Netherlands & Switzerland	Mixed integer linear	Community Energy Storage (CES) outperformed Home Energy Storage (HES) in both cost and CO2 reduction, with sensitivity analysis indicating feasibility under a larger battery storage capacity.
[101]	This article uses highly resolved models of residential power demand and PEV use to assess the impact of uncoordinated in-home PEV charging on residential power demand. The study does not implement smart charging or coordination strategies, relying instead on uncoordinated charging scenarios that limit exploration of demand management solutions.	USA	Stochastic model	Results show that while total electricity consumption increases modestly (~5% for 10% EV share), uncoordinated charging significantly reshapes demand profiles, exacerbates peak loads, and stresses distribution infrastructure at even low EV penetration.
[102]	This work proposes a new prediction model for aggregated loads of buildings and EVs. It remains limited by its reliance on historical data patterns, which may not capture sudden behavioral or policy shifts.	China, US, Iran, Iraq, & Hungary	Enhanced support vector machine	The model reduced forecasting error compared to traditional methods, achieving lower MAPE and RMSE across short-term load prediction scenarios.
[103]	This work provides a novel method for the optimal simultaneous allocation and sizing of RES and EV charging stations and for managing the vehicle charging process. It assumes simplified EV behavior and uniform adoption patterns, which may not reflect diverse user charging habits and stochastic uncertainties of real-world EV fleets.	Iran, USA, & China	Genetic Algorithm-Particle Swarm Optimization	The improved GA-PSO algorithm reduced power losses, voltage fluctuations, and costs more effectively than benchmark methods (DE2), proving its efficiency in IEEE 33-bus simulations.
[104]	This study simulates several potential grid scenarios and resulting demand response priorities. The pilot was limited to only 30 participants and a small regional context, restricting generalization to larger, more heterogeneous EV populations.	Canada	Cloud-based algorithm	The program demonstrated technical feasibility of utility-controlled charging, with curtailments up to 27 kW, while maintaining driver satisfaction and meeting state-of-charge requirements.
[105]	This work develops a decentralized algorithm that minimizes communication and delay to address both grid-level concerns and local-level issues. The proposed algorithm simplifies communication and computation but does not account for broader stochastic variations in EV usage and regional grid diversity.	USA	Decentralized algorithm	Coordinated strategies reduced transformer hot spot temperatures and equivalent aging factors compared to uncoordinated charging, thereby extending transformer life.
[106]	This paper discusses implementing smart charging algorithms to directly control EV charging rates and starting times at residential locations. The framework relies on static TOU and controlled charging strategies, which may not adapt to evolving real-time price signals or heterogeneous user behavior.	USA	Open DSS	Simulation results showed that optimized TOU and smart charging reduced transformer overload and feeder voltage deviations compared to uncontrolled charging.
[107]	This article presents an optimal smart charging algorithm that enables the charging of large numbers of vehicles without adverse effects on the electricity network at residential locations. The optimization relies on linear approximations of voltage drop and assumes the availability of complete network state information, which may limit accuracy under complex or large-scale systems.	Australia	Linear approximation	The method enabled high EV penetration (up to 50%) without major network upgrades, while maintaining voltage and line current within operational limits.
[108]	This paper reviews strategies, algorithms, and methods for implementing a smart charging control system. It synthesizes strategies without implementing quantitative validation, limiting direct performance benchmarking.	Spain	Power system analysis software approaches	The review highlights that centralized and decentralized smart charging can both mitigate voltage deviations and transformer overloads, but practical implementation remains constrained by communication and market barriers.
[109]	This study proposes a novel decentralized valley-filling charging strategy using a day-ahead pricing scheme. The model assumes reliable day-ahead pricing signals and user compliance, overlooking real-world variability in participation and device-level autonomy.	China	Multi-objective charging optimization algorithms	The decentralized strategy achieved a valley-filling effect with 28% lower generation cost compared to uncoordinated charging, and showed robustness to parameter uncertainty with less than 2% deviation.
[110]	This paper presents an EV charging method with a PV system for smart homes/buildings. The proposed EV charging algorithm is designed to determine the optimal schedules of EV charging based on predicted PV output and electricity consumption. The optimization framework assumes error-free PV and load forecasts without reporting uncertainty measures, and it omits grid feedback and V2G interactions, thereby limiting its robustness under real-world operating conditions.	South Korea	Mixed Integer Linear	The proposed scheduling method reduced charging costs by 6% compared to immediate charging (baseline #1) and by 15.2% compared to delayed charging (baseline #2), while ensuring the target SOC was met.

The reviewed studies (published between 2014–2023) on integrating EV charging infrastructure with RESs in residential buildings consistently show that coordinated and optimization-based models consistently outperform uncoordinated or static approaches, improving grid stability, cost reduction, and renewable integration. Early works (e.g., [101,104]) highlight the risks of unmanaged charging, with increased peak demand and localized grid stress, though pilot projects confirm that smart charging can be implemented while maintaining user satisfaction. More advanced hybrid and optimization approaches (e.g., [102,103,109]) demonstrate strong performance in forecasting and minimizing technical losses, but simplified assumptions about user behavior and uniform adoption patterns limit their effectiveness. More recent scenario-based and field-tested studies (e.g., [85,86,88]) emphasize local context, renewable integration, and emissions reductions, though they often rely on geographic- and policy-specific assumptions. Tool-based and review works (e.g., [87,108]) broaden the discussion but lack empirical validation.

Across the reviewed studies, methodological success depends on data richness, temporal granularity, and system coupling depth. Supervised approaches (e.g., LSTM, BiLSTM, GRU) achieve high forecasting accuracy under well-structured datasets but degrade under sparse or noisy conditions. Their performance further weakens when spatial planning of charging infrastructure and operational scheduling are not co-optimized with renewable generation or grid constraints, limiting real-world scalability. In contrast, hybrid frameworks combining unsupervised clustering (e.g., K-means, GMM) and RL control achieve greater adaptability by linking predictive accuracy with dynamic scheduling and multi-domain coordination. As illustrated in Figure 3, Figure 4, Figure 5 and Figure 6, the unifying conceptual framework organizes these methodologies into a hierarchical architecture connecting forecasting, optimization, and policy coupling layers. This structure bridges spatial siting, temporal scheduling, and regulatory integration across building, transport, and grid systems, enabling scalable and context-aware energy management.

Despite algorithmic progress, real-world deployment remains constrained by practical barriers: data privacy regulations restrict access to high-resolution mobility and consumption records; interoperability issues across charging standards and data formats hinder integration; and computational demands limit large-scale hybrid and RL applications. Furthermore, model interpretability and regulatory acceptance remain persistent challenges. Embedding explainable AI, privacy-preserving learning (e.g., federated learning), and open benchmarking datasets can accelerate the translation from simulation to real-world deployment.

Overall, the synthesis of reviewed studies demonstrates that integrating data-driven forecasting with coordinated spatial planning and operational scheduling yields the most resilient and cost-effective pathways for EV–renewable–building energy systems. Hybrid and context-aware approaches enhance predictive accuracy, adaptability, and interoperability across technical, behavioral, and regulatory dimensions. To operationalize these advances, policymakers and utilities should promote standardized data-sharing protocols, interoperable communication frameworks, and transparent benchmarking tools to ensure replicable, secure, and equitable deployment.

6. Conclusions

DL has become an essential tool for predicting EV charging demand, building loads, and RE generation at scale, as well as for managing energy in residential communities. This review critically examined DL and hybrid optimization approaches for forecasting, control, and the integration of EVs with RESs in residential and community contexts. Supervised models such as LSTM, BiLSTM, and GRU achieve high forecasting accuracy under structured datasets but are sensitive to data quality, temporal resolution, and exogenous variability. Hybrid architectures that integrate supervised, unsupervised, and reinforcement learning consistently outperform single-model approaches by coupling predictive precision with adaptive control. When combined with spatial planning and real-time operational strategies that account for renewable variability and grid constraints, these systems enhance forecasting reliability, operational efficiency, and overall energy system resilience.

As detailed in Section 5, the proposed unifying conceptual framework links forecasting, optimization, and policy coupling within a hierarchical architecture, enabling coordinated decision-making across building, transport, and grid layers. By integrating predictive modeling with optimization-based control and regulatory alignment, the framework provides a scalable pathway for real-world EV–renewable–building energy systems.

Although DL and hybrid optimization frameworks have demonstrated high predictive accuracy and adaptability, their large-scale deployment remains constrained by data privacy regulations, interoperability among heterogeneous systems, and computational resource requirements. Embedding explainable AI modules, federated learning, and open benchmarking datasets will be essential to ensure transparent, secure, and reproducible implementation in real-world energy systems. Moreover, coordinated policy action—through standardized data-sharing protocols, interoperable communication infrastructures, and incentive-based regulatory mechanisms—will accelerate technology adoption and ensure equitable access to smart charging and renewable integration solutions.

To make these approaches actionable, application-specific strategies should combine rich data acquisition, targeted model optimization, and deliberate cross-domain integration. Detailed EV session logs, traffic flows, household load data, and renewable output forecasts should be systematically collected to capture temporal, spatial, and behavioral heterogeneity. These datasets can then inform optimized forecasting and control pipelines, where supervised DL captures short-term variability, unsupervised clustering reveals user archetypes, and RL or MPC balances grid and user objectives under operational constraints. Cross-domain integration of transport, building, and grid models—augmented by robust optimization heuristics and demand-response mechanisms—ensures that EV charging, local renewables, and household flexibility are co-optimized. Collectively, these strategies operationalize the hierarchical framework linking forecasting, optimization, and policy coupling, enabling scalable, context-aware, and resilient energy management across residential and community systems.

7. Future Work and Policy Pathways

The findings of this review indicate that continued exploration is needed to enhance the sustainability and resilience of integrated energy systems through the combined use of DL, EV integration, and community-based RE solutions. While this study advances understanding of how EVs and RESs can be integrated into community grids, several methodological and implementation challenges remain. A key limitation is the availability and quality of high-resolution data, particularly for EVs equipped with large-capacity batteries and 20 kW chargers, since accurate DL training requires extensive datasets often restricted by privacy concerns in residential communities. DL models can also be computationally demanding, making real-time control and deployment in community-scale systems challenging without adequate computational resources. Additionally, interoperability with diverse EV models, charging hardware, and legacy grid systems must be addressed to ensure operational stability amid renewable generation and demand uncertainty.

Future research should prioritize comprehensive cluster analyses to better characterize residential charging behaviors and load profiles, particularly for next-generation EVs with higher power capacity. Applying advanced RL to EV scheduling can further enhance smart charging coordination under dynamic grid conditions while maintaining user convenience. Expanding model training to include field data, regional climate influences, and stochastic renewable fluctuations will improve robustness and adaptability for practical deployment. Post-processing tools such as global sensitivity analysis should also be explored to improve interpretability and provide greater transparency in model decision-making, thereby strengthening trust in DL-based forecasting and control frameworks.

Moreover, future studies should address regional and climatic diversity—especially the underrepresentation of cold weather contexts such as Canada—by developing climate-specific models and integrating complementary technologies such as thermal energy storage, building envelope retrofits, and cold-climate heat pumps within unified optimization frameworks. Beyond the residential context, future research should extend the proposed analytical framework to large-battery fleets such as electric buses and heavy-duty vehicles, which operate across multiple public charging depots and introduce additional challenges in spatial planning, power allocation, and grid coordination. Developing integrated deep learning and optimization models for these large-scale systems will be essential to support regional infrastructure planning, operational scheduling, and renewable energy alignment at the city and corridor levels. Policymakers and utilities can accelerate adoption by supporting standardized data-sharing protocols, interoperable communication infrastructures, and transparent benchmarking tools. Coordinated research across technical, behavioral, and regulatory domains will be vital to achieving scalable, equitable, and carbon-neutral energy communities where predictive accuracy, operational flexibility, and societal trust co-evolve.