Overview of the Application of Artificial Intelligence in China’s Park-Level Integrated Energy System: Current Status, Challenges, and Future Paths

Abstract

The global low-carbon energy transition relies on the orderly integration of a high proportion renewable energy. As an important carrier of demand-side energy systems, parks are responsible for local balancing and the accommodation of distributed renewable energy. However, the energy systems of parks exhibit the integrated characteristics of heterogeneous energy sources, including electricity, heat, and gas. It also encompasses the entire source–network–load–storage process, which renders it huge and complex. For this reason, as a systematic review article, this paper aims to summarize the overall application of artificial intelligence technology in China’s park-level comprehensive energy system. First, the current status of technology applications in the corresponding scenarios is analyzed based on three dimensions: prediction, scheduling, and security. Subsequently, key challenges in applying AI technologies to these scenarios are identified, including multi-temporal and spatial synergy issues in source–load forecasting, multi-agent equilibrium problems in dispatch optimization, and cross-modal matching challenges in security operation and maintenance (O&M). Thereafter, the feasible directions to solve these bottlenecks will be discussed comprehensively in light of the latest research advancements. Finally, we propose a phased roadmap for technological development and to identify the key gaps in this research field, such as the lack of publicly available benchmark datasets, data exchange standards, and cross-campus validation frameworks. This article aims to provide a systematic theoretical reference and development framework for the in-depth empowerment of AI technology in the integrated energy system of industrial parks.

1. Overview of the Development of Demand-Side Park-Level Integrated Energy Systems

In response to global climate change, China has proposed the ‘dual-carbon’ strategy (i.e., carbon peak and carbon neutrality) as part of its development approach. As policies such as carbon tariffs, carbon accounting, and other related measures continue to advance, the decarbonization of the energy system is becoming increasingly critical. Among these efforts, the large-scale development of demand-side distributed new energy has become a key pathway for constructing China’s new power system. According to statistics released by the National Energy Administration (21 July 2025), as of June 2025, the cumulative installed capacity of China’s distributed new energy (mainly distributed photovoltaics) has reached 493 million. As a key carrier for integrating distributed new energy, park-level integrated energy systems have gained growing importance in national top-level planning. On 12 December 2024, China’s Central Economic Work Conference, for the first time, proposed the ‘establishment of a number of zero-carbon parks’ and designated it as a core task for 2025. This initiated intensive efforts in the development of zero-carbon parks; subsequently, a series of policies were introduced in this field, laying a solid foundation for demand-side decarbonization transformation at the park level.

A park-level integrated energy system typically refers to an energy supply system that integrates multiple energy carriers, such as electricity, heat, cold, and gas, within a specific geographic area (such as industrial parks, commercial parks, or technology parks) to achieve integrated and coordinated energy production, conversion, storage, transmission, and consumption. Compared with the commonly used microgrids (which focus more on power systems) or campus energy systems (which have relatively single load types) internationally, park-level integrated energy systems typically cover an area of 5–20 square kilometers (which can be expanded to tens of kilometers with the development of long-distance heating technology), serving various user types, such as industry, commerce, and residents. Their multi-energy coupling characteristics are more significant, and they are a key node connecting macro-energy networks with micro-user needs, as well as an important lever for China to achieve decarbonization goals. The system comprises distributed renewable energy; combined cooling, heating, and power (CCHP); cold storage; heat storage; electricity storage; and other energy equipment, forming a complete demand-side ‘source–network–load–storage’ system [1]. The rapid development of new energy technologies over the past two decades has driven the evolution of RIES in functional positioning and bottleneck demands, and its stage division and characteristics are illustrated in Figure 1.

In this context, it is crucial to maximize carbon reduction benefits through local accommodation and the balancing of distributed new energy resources in parks. The essence of this demand lies in fully leveraging the role of multi-energy interactions and autonomous coordination in terminal, distributed integrated energy systems. With the large-scale integration of distributed new energy, new-type energy storage, electric vehicles, and other decentralized resources within parks, the source–load imbalance over time and space will continue to intensify [2]. Moreover, the multi-agent characteristics and operational uncertainty will make distribution network planning [3], scheduling, and control more complex [4]. Although the development of an intelligent distribution network with “micro-distribution synergy” has become an industry consensus, the intelligent regulation technology for park-level electricity–heat–gas heterogeneous energy networks remains largely theoretical [5]. The construction of digital twin IoT infrastructure is progressing rapidly, but the lack of reliable and fast decision-making and control mechanisms means that AI empowerment in the park-level integrated energy sector is still lagging behind and urgently needs rapid development.

The park-level integrated energy system features a complete “source–network–load–storage” structure, with construction and operational regulation needs similar to those of both the main grid and distribution networks. In this paper, the three dimensions of prediction, scheduling, and security are distinguished according to their corresponding application scenarios and relevant matching algorithms, with a focus on source–load prediction, scheduling optimization, and security operation and maintenance (O&M). First, an overview of the current status of technology in park-level integrated energy systems is presented based on the common application scenarios of the three technologies, elaborating on the specific manifestations of the primary stage of the parks’ smart transformation and the main challenges under current development trends. Second, considering the differentiated AI technology development needs in each scenario, the three latest application directions are examined in detail. Finally, the future evolution of AI-based technologies in park-level integrated energy systems is discussed. The current status and development trends of AI empowerment in the park-level integrated energy sector are illustrated in Figure 2.

2. Technology Status and Bottlenecks in the Park-Level Integrated Energy Sector

Driven by the ‘dual-carbon’ goal (i.e., carbon peak and carbon neutrality), park-level integrated energy systems have become a key scenario for energy transition due to their close coupling of multiple energy types, complex multi-agent interactions, and strong heterogeneity of multi-level equipment. Their significant individual differentiation has further highlighted the separation between data-driven paradigms in existing AI applications and physical mechanisms, thus demanding higher model generalization. As park-level systems already feature a complete source–network–load–storage energy system and function as a coupled system of heterogeneous energy types (electricity, heat, and gas), the application of AI technologies in this field faces three core challenges: multi-dimensional synergy, multi-timescale coupling, and multi-modal data heterogeneity.

Currently, the intelligent transformation of the park-level integrated energy sector remains in its primary stage, with AI applications mainly limited to basic data analysis and simple decision support. In practical scenarios, the initial integration of traditional mechanistic models and data-driven methods is the mainstream approach, and most parks have not yet deployed complex AI systems on a large scale. Based on actual application scenarios in the park-level integrated energy sector, the current status and bottlenecks of AI technology applications are analyzed from three common technical scenarios: source–load forecasting, scheduling optimization, and security operation and maintenance (O&M).

2.1. Distributed Renewable Energy Output and User Load Forecasting

2.1.1. Application Status: Initial Integration of Traditional Models and Data-Driven Techniques

Currently, source–load forecasting in park energy systems is still dominated by traditional statistical methods, but the penetration of data-driven techniques is gradually increasing. Taking Long Short-Term Memory Network (LSTM) as an example, this model can effectively capture the long-term dependencies of time series, making it perform well in short-term load forecasting [6,7,8]. A typical application is to integrate meteorological data (temperature, irradiance, etc.) with historical load curves and control the 24 h prediction error of photovoltaic outputs within the range of 12–15% (for example, the experimental results in reference [6] were from a specific park dataset). However, LSTM lacks adaptability to complex scenarios, such as device aging and local shadow occlusion, and hyperparameter adjustment has a significant impact on prediction accuracy, requiring manual experience correction. Some parks have attempted to enhance robustness by integrating mechanistic models [9], e.g., introducing building heat transfer equations in cold load forecasting to limit the output range of neural networks through physical constraints. It is worth noting that current park-level source–load forecasting often only involves hyperparameter adjustment, using similar deep learning algorithms without in-depth exploration of their respective characteristics [10,11].

In medium- and long-term forecasting, Random Forest and Gradient Boosting Decision Trees (GBDT) [12], as well as Monte Carlo simulation [13,14], are dominant. The advantage of GBDT lies in its ability to handle nonlinear relationships and provide feature importance ranking, such as being used to identify the impact weight of industry types and workday/holiday patterns on load, thereby reducing load forecasting errors [12]. The former can identify the impact weights of industry types, how weekday/holiday patterns can be used to construct an annual wind speed distribution model through 100,000 random samples, and how the configuration error of energy storage capacity can be controlled by correcting Weibull distribution parameters. In addition, to address the multi-energy flow coupling characteristics of parks, some applications have explored hybrid modeling methods [15,16] that combine Random Forest with thermodynamic system modeling. However, such methods rely on expert experience for feature engineering, limiting their generalizability [17,18].

Table 1. Comprehensive review of distributed renewable energy and load forecasting methods.

Prediction Target	Time Scale	Data Modality	Model	Evaluation Metrics	Ref.
PV Forecasting	Short- Term	Meteorological Data Sky Images/Satellite Images Historical Data SCADA Data NWP Data Regional Wind Power Aggregation Data PVGIS Data	LSTM	RMSE/ Skill Scores (SS)	[20,21,22,23,24]
			Stacking Ensemble Models (LSTM + GRU, SVR)	R2/NMSE/NRMSE	[25,26]
			Temporal Fusion Transformers (TFT)/ LSTM with Dual-Attention Mechanisms	MAE/R2/PICP WMAPE	[26,27]
	Medium-Term		RF-LSTM Hybrid Model	MAE/RMSE	[18,28]
	Long- Term		Transformer + M4 Fusion Strategy	MAE/WMAPE/ R2	[29]
Wind Power Forecasting	Short- Term	SCADA Data NWP Data Historical Data Geographic Data Meteorological Data	Bi-LSTM/ S-R CNN	MAE/CRPS/PICP NMAE/NMSE/Bias	[23,30]
	Medium- term		xLSTM	MAE/RMSE/R2	[31]
	Long- term		Bi-LSTM/GNN	MAE/PICP	[23,30]
Load Forecasting	Short- Term	IoT Data Historical Load Data Meteorological Data Economic Data PV Data SCADA Data Electrolyzer Data PV Panel Data	LSTM families (Bi-LSTM, GRU) Shuffle Transformer Multi- Head Attention Net	MAE/NRMSE/ R2	[32,33,34,35]
	Medium-Term		LSTM families Parallel LSTM-MLP XGBoost	MAE/RMSE/MAPE	[36,37]
	Long- Term		LSTM + XGBoost Transformer families RF, XGBoost, SVR, MLP	MAE/R2/NRMSE	[37,38]

systematically summarizes the key technical elements of photovoltaic (PV), wind power, and load forecasting across different time scales, covering data mo-dalities, model families, and evaluation metrics. It aims to provide a structured framework for method selection and performance evaluation in various forecasting tasks. In PV forecasting, short-term prediction often relies on the integration of mete-orological and IoT data, while medium-term forecasting emphasizes the model’s abil-ity to capture periodic features. Studies indicate that—through effective Numerical Weather Prediction (NWP) integration strategies, such as feature concatenation or de-coder input—LSTM and Transformer models improve short-term forecasting accuracy by 55.1% and long-term accuracy by 62.4% compared to scenarios without NWP input [19], demonstrating strong synergy between model architecture and data strategy.

In wind power forecasting, SCADA and meteorological data serve as the primary inputs. Attention-based Transformer variants outperform conventional LSTM models in short-term forecasting, reducing the Mean Absolute Error (MAE) by 18.2% and enhancing prediction stability under sudden wind speed changes by 23.5% [30], confirming the capability of advanced model families to capture the intermittency of wind power. For load forecasting, models must balance multi-scenario adaptability and computational efficiency. For example, a Transformer architecture incorporating meteorological and historical electricity consumption data significantly reduces both Root Mean Square Error (RMSE) and MAE across datasets with different sampling rates while maintaining a low parameter count [24]. Notably, the application scope of load forecasting is gradually expanding to integrated scenarios, such as green hydrogen production. Through load–green hydrogen collaborative modeling, the production forecast error (Mean Absolute Percentage Error, MAPE) can be controlled within 8.7%, and the accuracy can be improved by more than 14% [35].

2.1.2. Bottleneck Problems: Deep Contradictions Between Data–Mechanism Synergy and Spatiotemporal Correlation Modeling

(1)

Limitations of the Data-Driven Paradigm and Challenges of Dual-Driven Fusion

Current prediction models are overly reliant on data-driven approaches, ignoring physical mechanism constraints, which leads to a sharp drop in prediction accuracy in complex microclimate and multi-shading scenarios within parks. Purely data-driven models (e.g., LSTM and Transformer) can capture statistical patterns in historical data but cannot embed physical constraints, such as PV panel temperature decay equations or wind turbine aerodynamic equations. Consequently, prediction errors increase significantly under unexpected abnormal operating conditions, such as equipment failures and extreme weather. While physical models ensure mechanistic correctness, they struggle to handle nonlinear disturbances, like sudden local weather changes and shadow occlusion in parks. The key challenges in data-model dual-driven fusion include the following: (1) the choice of coupling methods for physical equations and neural network weights (hard constraints/soft regularization); (2) balancing real-time computational efficiency and mechanistic modeling accuracy; and (3) the collaborative mechanism between online identification of equipment-level physical parameters and model updates.

(2)

Spatiotemporal Correlation Modeling Dilemma for Collaborative Source–Load Forecasting

Distributed PV, energy storage, and flexible loads in parks form a strong spatiotemporal coupling system. Traditional independent forecasting models result in excessive standby capacity redundancy. Thus, the joint extraction of source–load features must be achieved through a multi-task learning framework [39,40]. However, two major challenges exist: (1) dynamic construction of spatiotemporal correlation matrices [41] (for example, the impact of PV output fluctuations on charging pile loads exhibits a time-varying lag (5–30 min), requiring an adaptive time-delay compensation module); (2) development of a cross-modal feature sharing mechanism to achieve semantic alignment [42] of heterogeneous data (meteorological data [continuous values], equipment status [discrete alarms], and load demand [probability distribution]) in the shared layer.

(3)

Differences in Multi-Time Scale Uncertainty Modeling

The probability distribution of wind and solar output exhibits non-Gaussian and multi-peak characteristics. Traditional assumptions of normal distribution lead to distorted risk assessments. For an AI-based uncertainty characterization of wind and solar output, significant differences exist across time scales: minute-level (cloud cover), hourly level (weather systems), and seasonal-level (climate patterns). Conventional probabilistic prediction models with a single temporal resolution [43] (e.g., GAN [Generative Adversarial Networks] and VAE [Variational Autoencoders]) struggle to simultaneously capture: (1) the spiky and thick-tailed characteristics of minute-level fluctuations; (2) the time dependence of hourly level prediction intervals; and (3) the non-stationarity of seasonal-level pattern drifts [44].

(4)

Bottleneck of Association Analysis for Knowledge Transfer Under Data Scarcity

Newly built parks face insufficient cold-start data [45], while existing parks experience fundamental load changes due to user occupancy, expansion, or renovation [46]. Traditional transfer learning ignores inter-park association topologies [47], which include the following: (1) climate zone similarity (temperature/irradiance distribution matching); (2) industrial load characteristics (differences in energy usage patterns between manufacturing parks and commercial parks); and (3) equipment configuration correlations (compatibility between PV inverter models and energy storage systems). In such cases, knowledge transfer and sharing must be conducted under the premise of ensuring user data privacy [48].

2.2. Integrated Energy System Scheduling Decision-Making

2.2.1. Application Status: Conventional Algorithms Dominated by Shallow Reinforcement Learning Pilots

Currently, conventional algorithms, such as mixed-integer linear programming (MILP), are still core features for scheduling optimization [49]. By combining them with two-stage optimization, Model Predictive Control (MPC), bilevel optimization, and robust optimization [50], they can solve different single-objective or multi-objective problems in the electricity–heat system. Such methods are computationally time-consuming but remain preferred by most parks due to flexible constraints and verifiable results. When evaluating these methods, comparable indicators, such as total operating costs, carbon dioxide emissions, constraint violation rates (such as power imbalance), and computation time, are typically used as benchmarks [51,52].

Shallow reinforcement learning techniques have begun pilot applications [53]. For example, Deep Q-Network (DQN) has attracted attention in energy storage optimization scheduling due to its ability to handle high-dimensional state spaces. A typical case is the construction of an energy storage control strategy based on DQN, which maps the state space of electricity price signals and load fluctuations to increase frequency regulation revenue by about 5–10% [54,55]. However, the DQN equivalent function method has limitations when dealing with continuous action spaces. However, these applications are generally limited by state space dimensions [56], handling only 5–8 key variables, with complex scenarios still requiring manual intervention. One park attempted to expand the state space to 15 dimensions (including wind and solar output, storage SOC (state of charge), grid tariff, etc.), but training time surged from 3 to 72 h, limiting practicality [57,58] (Figure 3).

2.2.2. Bottleneck Problems: Synergistic Dilemma of Multi-Energy Coupling and Multi-Agent Game

(1)

Multi-Timescale Conflict in Fine Scheduling

Dynamic responses of the electricity–heat–gas multi-energy flows at the park level differ significantly [59]: power system responses are millisecond-level, heat networks have hourly delays, and gas network pressure wave propagation is minute-level. Existing studies use differential algebraic equations (DAEs) to describe multi-energy coupling, but neural networks have poor numerical stability with DAEs—especially in heat network delay links, which easily trigger gradient explosions. Thus, it is critical to achieve breakthroughs in the differentiable transformation of multi-energy flow equations, develop DAE-based neural network structures to accurately characterize the impact of heat network delay characteristics on scheduling decisions, coordinate the conflict between offline (hourly) and online control (second-level), and build a complementary cooperative feature extraction system for historical and real-time data.

(2)

Equilibrium Diversity Requirements of a Multi-Agent Game

New energy operators, power sales companies, users, and other agents in parks form a complex game network. Traditional Nash equilibrium assumes complete rationality and information symmetry, but practical scenarios need to accommodate the following: (1) evolutionary games [60] (user response strategies dynamically evolve with tariff signals); (2) cooperative games (revenue distribution mechanisms for multi-agent sharing alliances); and (3) master–slave games (Stackelberg games between operators and aggregators). These high-dimensional interest conflicts need to be disassembled and combined to construct a multi-agent game model with real-time equilibrium capabilities [61].

2.3. Fault Diagnosis and Response Management

2.3.1. Application Status: Primary Application of Threshold Alarms and Image Recognition

Currently, threshold alarm systems remain the main means of fault detection. For example, deploying an anomaly detection system based on the sliding window mean method [62] triggers an alarm when the string current of a PV module deviates from the reference value by 10% for three consecutive sampling points, with an accuracy of up to 85%. However, its false alarm rate is as high as 20%; it was especially prone to frequent false triggers in cloudy weather, requiring O&M personnel for manual review. Evaluating such methods should give particular consideration to indicators specific to rare fault categories, such as accuracy, recall, area under the curve (AUC-PR), false positive rate (FPR), and mean time to detect (MTTD) faults [62,63,64].

AI-based image recognition technology has advanced in small-scale pilots, e.g., using the YOLOv3 model to analyze UAV inspection images [63]. However, model generalization is poor: accuracy varies by over 30% across different wind turbine models, requiring separate training for each device [64]. Primary applications of knowledge graphs show value: park energy management systems can associate equipment ledgers, maintenance records, and SCADA data [65] to build an O&M knowledge base with tens of thousands of nodes, reducing average fault localization time by 40%. The current data/model fusion strategies, such as expert mixing and cross attention mechanisms, are still in the exploratory stage and face typical challenges such as category imbalance, data drift, and poor cross-campus transferability [64]. The key path from laboratory results to deployable solutions lies in building high-quality, well-labeled fault datasets and establishing cross-campus model validation benchmarks.

In summary, although some parks have deployed smart meters and environmental sensors across the board, data quality issues still restrict AI applications. High-performance algorithms rely on on-site GPU clusters [66], and AI infrastructure faces dual challenges of coverage and quality. Thus, AI applications in park-level integrated energy systems remain in the pilot stage of conventional machine learning algorithms.

2.3.2. Bottlenecks: Technical Bottlenecks in Cross-Modal Diagnosis and Fast Recovery

(1)

Semantic Gap in Multimodal Data Fusion

Equipment condition monitoring involves multimodal data, such as infrared images (spatial features), vibration signals (time-frequency features), and SCADA data (temporal features) [67,68,69]. Existing fusion methods face two challenges: (1) feature scale mismatch (difficulty aligning image local features (e.g., hotspots) with global vibration spectrum features); (2) poor tolerance for missing modalities (anomalies in a single modality invalidate overall diagnosis). It is necessary to extract common features through multimodal synergy and establish cross-modal cooperative weight assignment methods [70,71].

(2)

Dilemma of Causal Correlation Decoupling for Composite Faults

Electrical faults (e.g., insulation breakdown) and information faults (e.g., communication interruption) [72] may occur concurrently with implicit causality: (1) fault propagation path masking (information link interruptions hide real electrical fault features); (2) cross-domain causal confusion (SCADA data anomalies may stem from sensor faults or actual equipment degradation). Existing graph neural networks struggle to distinguish spurious correlations, requiring analysis of internal causal deduction logic for composite faults and construction of a dual composite fault identification system with correlation structure and data recognition [73].

(3)

Dual Timeliness Constraints for Fast Recovery Decision-Making

Fault recovery must satisfy the following: (1) fault-tolerant control timeliness (isolating the fault area within 10 s); (2) power supply recovery real-time performance (reconstructing the optimal supply path within 5 min). Existing lightweight diagnostic models suffer excessive accuracy loss [74], while digital twin simulation and deduction are time-consuming [75]. Thus, it is necessary to combine high-precision knowledge/data; dual-driven fault diagnosis methods [76]; refined scheduling optimization technologies [77]; and a cloud-edge collaborative architecture to achieve fast fault identification and response recovery [78].

2.4. Analysis of Key Indicators for AI Methods in Park-Level Integrated Energy Systems

To further clarify the applicable boundaries of existing AI methods in park-level integrated energy systems, this section conducts a critical comparison of three mainstream methods—traditional machine learning, deep learning, and deep reinforcement learning—from four core dimensions (with specific indicators as shown in

Table 2. Critical comparison table of key indicators for AI methods in park-level integrated energy systems.

Evaluation Dimension	Traditional Machine Learning (RF, GBDT, Monte Carlo Simulation)	Deep Learning (LSTM/GRU, PINNs Transformer Multi-Modal Model)	Deep Reinforcement Learning (Q-L, MADDPG, Generative Reinforcement Learning)
Data Requirements
Data Type	Structured data	Multi-source heterogeneous data	Environmental interaction data
Sample Size	Medium	High	Extremely high
Sensitivity to Data Quality	Low	High	Medium
Computational Complexity
Training Phase	Low	High	Extremely high
Inference Phase	Low	Medium–high	Medium
Hardware Dependence	None	Strong	Medium–high
Real-Time Adaptability
Inference Latency	Low	Medium–high	Medium
Dynamic Scenario Response	Medium	High	High
Scalability to Park-Scale
Small-Scale	High	Medium	Low
Medium-Scale	Medium	High	Medium
Large-Scale	Low	Medium	High
Core Advantages	Low data/deployment cost, strong real-time performance, suitable for basic prediction	High accuracy in multi-modal fusion/ nonlinear fitting, suitable for medium-scale multi-energy flow scenarios	Excellent multi-agent game/dynamic decision-making capabilities, suitable for large-scale cross-domain collaboration
Disadvantages	Complete failure in multi-modal/multi-agent scenarios	High data/computational cost, insufficient multi-agent collaboration for large-scale parks	Long training cycle, poor economy for small-scale parks

): data requirements, computational complexity, real-time adaptability, and scalability to park-scale.

This comparison provides a quantitative basis for parks to select suitable AI technologies based on their own scale (e.g., small-scale parks/large-scale zero-carbon parks) and scenario requirements (e.g., basic load forecasting/multi-agent scheduling). Meanwhile, it highlights the core bottlenecks of current technology applications (e.g., excessive data dependence of deep learning and high computational costs of deep reinforcement learning) and offers targeted guidance for the directions of subsequent core technology breakthroughs.

In summary, the complexity of park-level systems requires AI technologies to address the three-fold problem of multi-energy flow, multi-agent, and multi-modality, with key challenges focusing on physical knowledge embedding, game optimization integration, and multi-modal causal reasoning. The goal is to construct a new paradigm of AI applications with explainable mechanisms, coordinable games, and traceable faults, supporting efficient and reliable operation of park-level new energy systems.

3. Core Development Directions of AI Technology in Park-Level Integrated Energy Sector

3.1. Cross-Modal, Transferable Source–Load Collaborative Forecasting

The multimodal Transformer model requires the fusion of high-dimensional data, such as satellite cloud images and drone inspection videos, necessitating high GPU computing power from cloud computing centers. Federated learning relies on the local computing power of edge nodes for model training, only uploading encrypted model parameters. Its core requirement is to balance communication overhead and model performance. Among them, there are semantic gaps in the multimodal data alignment process, privacy leakage risks under federated learning frameworks (although parameters are encrypted, they may still be inferred backwards), and unstable model convergence. Currently, multimodal fusion prediction is at TRL 3-4 (laboratory validation), and federated learning frameworks are at TRL 4-5 (small-scale pilot). The goal for the next 5 years is to achieve TRL 6-7 (large-scale demonstration application) [79,80,81].

In recent years, the demand for fusing multi-source heterogeneous information (e.g., meteorological data, equipment status, and user behavior) in parks has driven the development of Transformer-based multi-modal joint modeling technology. For example, in a commercial complex project, a “meteorology–image–load” three-modal pre-trained large model was constructed [79], which dynamically aligns features of satellite cloud maps, building thermal imaging and POS transaction data via the self-attention mechanism. Experiments show that the model can capture the correlation between air conditioning load surges and PV output drops under heavy rainfall, reducing the 72 h prediction error by 31% compared to single-modal models. To address equipment heterogeneity in parks, multi-source transfer learning [80] (Figure 4) and federated multi-task learning frameworks [81] enable cross-park knowledge transfer as follows: three industrial parks with similar climates share encrypted parameters to jointly train wind and solar output forecasting models, with each park retaining a local equipment characteristic modeling branch. This ultimately reduces the cold-start prediction error of new parks by 42%.

3.2. Collaborative Scheduling for Multi-Agents and Heterogeneous Energy Sources

(1)

Multi-Agent Deep Reinforcement Learning (MADRL) for Multi-Agents

The training of MADRL agents requires a substantial volume of interaction data collection from the environment, usually conducted in simulated environments, which consumes a huge amount of computing resources and is suitable for offline training in the cloud. When applied online, the trained policy network can be deployed to the edge controller for execution. However, the non-stationarity between intelligent agents leads to training difficulties, convergence to local optima rather than global optima, and numerical stability issues when applied in rigid systems, such as power flow equations [82,83]. Currently, MADRL is in TRL 3-4 in park scheduling. The goal is to achieve TRL 6 in specific scenarios, such as virtual power plants, by 2030 [84,85,86].

Dynamic games among new energy operators, power sales companies, and users in parks [82] have promoted the expansion of MADRL into complex strategy spaces [83,84]. For example, a virtual power plant project adopted an improved MADDPG algorithm [85], defining a 53-dimensional state space (including electricity prices, carbon emission factors, storage SOC, etc.) and a 17-dimensional action space. It trains agents in phases via a curriculum learning mechanism: Phase 1 focuses on basic energy trading strategies, Phase 2 introduces demand response subsidy rules, and Phase 3 simulates emergency collaboration under extreme weather. Ultimately, the Nash equilibrium solution time across six agent types is reduced from 3.2 h to 11 min, with intraday trading revenue increased by 28%. A more complex large language model (LLM)-assisted decision architecture [86] (Figure 5) enables inputting unstructured texts (e.g., park scheduling rules, equipment manuals, and market policies) into a fine-tuned Llama2 model to generate interpretable scheduling strategy recommendations, improving manual review efficiency by 65%.

(2)

Multi-Timescale Optimization of Heterogeneous Energy via Neural Differential Equations (NDEs)

NDEs require high-frequency sensor data to characterize dynamic processes, with high demands for data quality and continuity. The solving process involves numerical integration, which has a higher computational complexity than feedforward neural networks and requires specialized hardware acceleration. Among them, the numerical solution of rigid differential equations (such as describing fast power transient processes) is unstable, which may lead to gradient explosion or disappearance, affecting the training effect [87,88]. At present, the application of neural differential equations in energy systems is in TRL 2-3 (proof of concept). The goal is to achieve TRL 5-6 in integrated demonstration projects within the next 5–10 years [87,88].

To address dynamic response differences in electricity–heat–gas multi-energy flows (millisecond to hourly), NDEs achieve cross-scale collaboration through differential equation modeling. For example, a regional energy internet project constructed a heat network delay dynamics model using Neural Ordinary Differential Equations (Neural ODE) to describe temperature field conduction in heat storage tanks, combined with LSTM to capture electricity price fluctuations. This forms a three-layer optimization architecture: “second-level response–minute-level adjustment–hour-level planning”. Operational data show the model reduces heat network supply–demand imbalance from 15% to 2.3% and peak shaving costs by 19% [87] (Figure 6). In an extreme scenario response, Generative Reinforcement Learning (GRL) offers unique value: in a wind–solar–storage–hydrogen park, a diffusion model generates a million-level virtual scenario library covering 10 extreme events (e.g., typhoons and extreme cold), training DRL agents to form robust strategies and increase power supply reliability from 82% to 96% [88].

3.3. Cross-Modal, Large-Model Security O&M Systems

(1)

Cross-Modal Comparative Learning (CMCL)

Cross-modal contrastive learning (CMCL) requires alignment learning of multimodal data (such as vibration, temperature, and images) from the same device, resulting in high data acquisition and annotation costs. Model training requires GPU support, but after training, lightweight encoders can be deployed on the edge side for real-time diagnosis. However, the performance of the model decreases when there is a lack of modality, as well as the failure of contrastive learning caused by inconsistent distribution of data from different modalities [89,90]. At present, CMCL is at TRL 3-4 in equipment diagnosis. The goal is to achieve demonstration applications of TRL 6 in specific industries, such as wind power and photovoltaics, within the next 3–5 years [91,92].

The need for the multi-dimensional perception of park equipment status (vibration, temperature, images, etc.) has driven breakthroughs in multi-modal fusion diagnosis. For example, a wind farm deployed a CMCL framework [89], achieving cross-modal correlation detection of early faults by aligning SCADA data spectral features, acoustic signal time-frequency maps, and infrared thermal imaging embedding spaces. When micron-level wear occurs in wind turbine gearbox bearings, the model synergistically analyzes high-frequency anomalies in acoustic signals and vibration spectrum shifts [90], advancing early warning time from 24 h (traditional threshold method) to 72 h, with false alarm rates reduced to <5%. Progress has also been made in the visual–physical joint modeling for PV module hidden crack detection [91], fusing electroluminescence (EL) image intensity distribution with IV curve characteristics [92], as well as in constructing a Graph Attention Network (GAT) to improve hidden crack detection rate from 78% to 94% (false alarm rate < 3%) [93] (Figure 7).

(2)

Large Language Models (LLMs) for Security O&M Scenarios

The large language model (LLM) requires a large amount of high-quality domain text data (manuals, regulations, and cases) for fine-tuning, with extremely high computational requirements, and it can usually only be deployed and serviced in the cloud. Among them, there may be a problem of “illusion”, that is, generating incorrect but seemingly reasonable operation and maintenance suggestions, which may lead to security accidents. In addition, there is a risk of a lack of domain knowledge and insufficient interpretability of decisions [94,95,96]. At present, the LLM is in TRL 2-3 in operation and maintenance. The long-term goal is to combine knowledge graphs and causal reasoning to achieve TRL 5-6 and become an advanced decision-making tool [97,98].

The explosive growth of unstructured data in parks (equipment ledgers, maintenance records, and expert experience) has spurred the development of LLM-based intelligent O&M assistants [94]. For example, in a zero-carbon park, the GPT-4 model is fine-tuned with 300,000 equipment parameter documents and 50,000 historical fault cases to build an interactive O&M knowledge engine. When transformer oil chromatography data is abnormal, the system automatically correlates similar cases from the past three years, manufacturer technical bulletins, and IEEE standards to generate a decision report (including fault probability ranking, disposal steps, and spare part lists), reducing average fault diagnosis time from 2.5 h to 18 min [95]. Additionally, augmented reality (AR)-based causal diagnosis systems show potential in substations [96]: O&M personnel use Hololens glasses to view 3D equipment heat maps superimposed with fault propagation paths [97,98,99], combining voice commands to invoke LLMs for root cause reasoning and improving complex fault localization efficiency by 70% [100].

The latest technological advancements indicate that park-level integrated energy AI applications are shifting from single-point breakthroughs to system reconstruction. Transformer-based multi-modal large models enable cross-vendor equipment unification via dynamic feature alignment, solving data silos caused by differences in park equipment brands, models, and communication protocols. The accurate portrayal of coupling effects between hourly heat network delays and millisecond-level grid fluctuations via NDEs addresses time-scale differences in electricity–heat–gas multi-energy flow dynamic responses. Furthermore, dynamic weighted federated learning frameworks combined with multi-agent technology resolve conflicts between data privacy and economic games among operators, users, and third-party service providers.

In summary, these breakthroughs focus on core features, such as park equipment heterogeneity, multi-agent gaming, and multi-energy flow coupling, marking energy system intelligence entering a new stage of “cross-domain collaboration–autonomous evolution”. Demand-oriented integration with underlying technologies (quantum computing chips and neuromorphic hardware) is expected to significantly improve AI real-time performance and energy efficiency, providing core impetus for new energy system construction.

4. Suggested Development Path for the Park-Level Integrated Energy Sector

In the future, AI technologies will gradually reshape the underlying architecture of park energy systems, enhancing real-time performance and energy efficiency of regional energy systems. This will promote the transition from local automation to full-domain intelligence. By 2030–2035, parks will become intelligent energy entities characterized by “autonomous decision-making, dynamic collaboration, and ecological friendliness.” With AI as the core driver, parks will integrate the entire energy production, transmission, and consumption chain, forming a sustainable paradigm featuring multi-energy complementarity, game equilibrium, and self-healing operation. Through deep integration of algorithms and physical systems, AI will overcome energy spatiotemporal constraints: Physics-Informed Neural Networks (PINNs) will build cross-domain prediction capabilities for multi-energy flows, achieving accurate matching of wind–solar–storage output and flexible loads. Multi-agent reinforcement learning will drive the transformation of scheduling systems into autonomous decision-making networks, supporting dynamic balance among various stakeholders and global resource optimization. Cross-modal causal reasoning and edge intelligence will endow the system with a “sensing–diagnosis–repair” closed-loop capability, upgrading equipment health management from threshold alarms to molecular-level self-healing. At the same time, parks will evolve into replicable zero-carbon technology benchmarks. AI will maximize clean energy penetration on the production side, dynamically optimizing energy consumption patterns through behavioral profiling on the consumption side, building multi-level collaboration mechanisms in scheduling networks, and enabling minute-level fault prediction and autonomous repair in O&M systems.

Based on the 5–10-year development vision (as shown in Figure 8), AI empowerment in park-level integrated energy systems can be phased into short-term, medium-term, and long-term plans. The goal is to deeply integrate AI into energy systems, making each park an AI-active node in a sustainable ecological network. In the process of promotion, it is necessary to simultaneously pay attention to and address the cost, regulatory, and cybersecurity challenges associated with technological evolution, to ensure the stability and feasibility of the development path.

(1)

Source–Load Multi-Temporal–Spatial Collaborative Forecasting

The future of source–load forecasting requires building a physics–data dual-driven hybrid intelligent architecture, promoting technology integration and scenario deepening in phases. The short-term focus is to achieve initial integration of data and mechanisms: embed physical constraints (e.g., PV panel temperature decay equations wind turbine aerodynamic equations) into deep learning models via Physics-Informed Neural Networks (PINNs), pilot solutions to wind–solar output prediction errors caused by local meteorological disturbances, and simultaneously use federated learning frameworks to realize encrypted transfer of cross-park equipment parameters. Medium-to-long-term goals include the deep synergy of multi-modal data: build a park-level multi-modal large model based on the Transformer architecture; integrate meteorological satellite cloud maps, equipment infrared thermal imaging, and user load time-series data; and dynamically align cross-modal features via self-attention mechanisms. Additionally, combine generative AI to generate a million-level virtual scenario library covering extreme weather (e.g., typhoons and extreme cold) to improve extreme event coverage. Ultimately, the prediction model will gain adaptive energy network evolution capabilities, providing high-fidelity inputs for personalized park-level integrated energy configuration and scheduling strategies. The large-scale deployment of large models should focus on considering computing power costs and energy consumption, and exploring model lightweighting and green computing technologies. Meanwhile, the security and credibility of data generated by generative AI require the establishment of auditing and authentication mechanisms. At the same time, the introduction of cutting-edge technologies such as quantum computing has extremely high costs and requires the promotion of research and development through major national science and technology plans or industry alliances, as well as advanced planning of their integration and transition paths with existing classical computing infrastructure.

(2)

Multi-Energy and Multi-Agent Collaborative Scheduling

Scheduling optimization technology needs to evolve from single-agent control to multi-agent collaboration, gradually solving multi-energy flow coupling and game equilibrium challenges. The short term focuses on the large-scale application of single-agent deep reinforcement learning (DRL): use DRL algorithms to realize real-time optimization of multi-scenario energy storage control strategies, match differentiated demand response scenarios of grid-connected virtual power plants, pursue accurate prediction and rapid response, and build a minute-level response mechanism for park-level “source–network–load–storage” systems. Coordinate games among new energy operators, power sales companies, users, and aggregators via Multi-Agent Deep Reinforcement Learning (MADRL); additionally, simultaneously use Neural Differential Equations (NDEs) to model hourly delay characteristics of heat networks, initially coordinating differences in electricity–heat dynamic responses. Medium-to-long-term goals are focusing on building a brain-like decision system: deploy neuromorphic computing chips to support Spiking Neural Networks (SNNs) for microgrid second-level frequency regulation and generate interpretable scheduling strategy reports based on large language models (LLMs) to reduce manual review complexity. This will enable second-level cross-domain allocation of wind–solar–storage resources and dynamic carbon flow tracking, and it will help build a blockchain-based distributed consensus mechanism. Multi-agent games involve complex distribution of interests and market rules, requiring regulatory agencies to take the lead in establishing new market mechanisms and game rules that are adapted to AI decision-making, while ensuring their fairness and transparency. The ‘energy metaverse’ also faces significant initial investment and ongoing cybersecurity challenges, requiring the establishment of a comprehensive network security protection system that can resist advanced persistent threats (APTs) across the entire system, and the development of strict policies for data sovereignty and cross-border flow regulation.

(3)

Cross-Modal Diagnosis and Rapid Recovery

Security O&M systems need to upgrade from passive alarming to active autonomy, realizing the closed-loop intelligence of fault diagnosis and recovery in phases. The short term promotes the transition from threshold alarms to AI diagnosis: piloting Cross-Modal Comparative Learning (CMCL) in key equipment; reducing false alarm rates by aligning embedding spaces of vibration spectra, infrared images, and SCADA time-series data; and simultaneously deploying lightweight digital twin systems to achieve the rapid monitoring of battery module-level temperature fields, thermal runaway propagation path deduction, and fault isolation responses. High-quality and standardized sensor deployment requires significant investment, which is the foundation of AI diagnosis, but also brings initial cost pressure. However, in addition to fault response, high-precision data perception can also be combined with AI regulation to improve overall system energy efficiency, and this investment still has economic value. Medium-to-long-term goals deepen causal reasoning and cluster autonomy: use Causal Graph Neural Networks (CausalGNNs) to analyze the implicit correlations between electrical faults and information interruptions; build a multi-agent autonomous recovery architecture; and realize the parallel localization of faulty segments and power supply reconstruction via edge-side consensus algorithms.

5. Conclusions

This review focuses on parks as a key entity to demand-side energy, starting with the current state of applications in China. It breaks down the AI-enabled aspects of park-level integrated energy systems into three application scenarios (corresponding to the mainstream directions of deep learning, reinforcement learning, and their combination, respectively): prediction, scheduling, and security. Through analysis of the current status in the field, key issues in these three scenarios—such as multi-temporal–spatial, multi-agent, and cross-modal characteristics—are identified, and cutting-edge AI technologies (including large models) are summarized. Finally, recommendations for future development in this field are provided, which could offer valuable theoretical insights for AI-enabled technologies in regional integrated energy systems, including demand-side parks and building clusters. To promote further development in this field, this article believes that the following research gaps need to be urgently filled:

In the field of prediction, there is currently a lack of publicly available, high-quality, and multimodal benchmark testing datasets for park-level energy, which seriously hinders the fair comparison and progress of algorithms. In the future, it is necessary to promote the establishment of data collection and exchange standards, as well as to conduct cross-validation research on models across multiple sites and climate zones.

In the field of scheduling optimization, there is a gap between the simulation environment of multi-agent games and the real system, and a benchmark testing platform that is closer to reality needs to be developed. At the same time, the interpretability and security evaluation system of AI decision-making has not been established, which restricts its application in high-risk scheduling scenarios.

In the field of security operations, the scarcity of fault samples makes model training difficult, and the application of small sample and zero sample learning techniques is key. In addition, it is necessary to build an autonomous operation and maintenance benchmark test scenario covering the entire chain of “perception–diagnosis–decision–recovery” to evaluate the overall system performance.