A Review of Hybrid LSTM Models in Smart Cities

Abstract

Rapid global urbanization poses complex challenges that demand advanced data-driven forecasting solutions for smart cities. Traditional statistical and standalone Long Short-Term Memory (LSTM) models often struggle to capture non-linear dynamics and long-term dependencies in urban time-series data. This review critically examines hybrid LSTM models that integrate LSTM with complementary algorithms, including CNN, GRU, ARIMA, and SVM. These hybrid architectures aim to enhance prediction accuracy, integrate diverse data sources, and improve computational efficiency. This study systematically reviews principles, trends, and real-world applications, quantitatively evaluating hybrid LSTM models using performance metrics such as mean absolute error (MAE), root mean square error (RMSE), and the coefficient of determination (R²), while identifying key study limitations. The case studies considered include traffic management, environmental monitoring, energy forecasting, public health, infrastructure assessment, and urban waste management. For example, hybrid models have achieved substantial accuracy improvements in traffic congestion forecasting, reducing their mean absolute error by up to 29%. Despite the inherent challenges related to structural complexity, interpretability, and data requirements, ongoing research on attention mechanisms, model compression, and explainable AI has significantly mitigated these limitations. Thus, hybrid LSTM models have emerged as vital analytical tools capable of robust spatiotemporal prediction, effectively supporting sustainable urban development and data-driven decision-making in evolving smart city environments.

1. Introduction

As global urbanization accelerates, cities are increasingly facing complex challenges, such as traffic congestion, rising energy consumption, environmental pollution, and aging infrastructure. However, conventional administrative approaches have proven inadequate to address these multifaceted issues. Consequently, a paradigm shift toward data-driven decision-making and real-time forecasting has become essential. In this context, the concept of the “smart city” has emerged—defined as an intelligent urban system that integrates information and communication technology (ICT), the Internet of Things (IoT), and artificial intelligence (AI) to manage urban resources efficiently and provide improved public services [1]. To operate effectively, smart cities require the ability to collect, process, and analyze vast volumes of urban data and make optimal real-time decisions based on that analysis.

A key feature of smart city operations is their reliance on time-series data generated from sensors embedded in infrastructure, vehicles, and the environment. Examples include energy usage, traffic speed, waste levels, and air quality indices. Accurate forecasting of such data is critical for responsive and sustainable urban management. Traditional statistical models such as autoregressive integrated moving average (ARIMA) and machine learning algorithms such as Support Vector Machines (SVMs) and Random Forests have historically been used for time-series forecasting. However, these models often exhibit limited performance in capturing non-linear patterns or long-term dependencies. For example, while ARIMA models may yield high short-term prediction accuracy, they frequently fail to capture sudden fluctuations or non-linear patterns in traffic forecasting [2]. This results in significantly increased errors over extended time horizons.

To overcome these limitations, deep learning-based approaches have gained prominence, particularly Recurrent Neural Networks (RNNs) and, more specifically, the Long Short-Term Memory (LSTM) model. Originally proposed by Hochreiter and Schmidhuber in 1997, the LSTM architecture introduces cell states and gating mechanisms—input, forget, and output gates—that enable the model to effectively learn long-term dependencies [3]. Since then, LSTM has been successfully applied to various smart city domains, such as traffic forecasting, energy demand prediction, and air quality monitoring. Despite its strengths, the standard LSTM model remains constrained by its fixed structure and limited adaptability, which may hinder its prediction performance in highly dynamic environments.

To address these shortcomings, recent studies have focused on hybrid LSTM models [4]. Hybrid LSTM integrates LSTM with other machine learning or deep learning algorithms, such as CNNs, GRUs, ARIMA, and SVMs, to leverage the complementary strengths of each component. The concept of hybrid time-series modeling can be traced back to Zhang’s [5] study in 2003 titled “Time series forecasting using a hybrid ARIMA and neural network model.” The study introduced a new paradigm by combining ARIMA for linear trends and artificial neural networks (ANNs) for non-linear components. Zhang’s approach has spurred extensive research in hybrid modeling, with its principles adapted to LSTM architecture, leading to various innovative hybrid designs. For instance, Bhanja and Das [6] proposed a hybrid deep learning model that effectively enhanced air quality time-series prediction performance by integrating convolutional layers with LSTM units.

Smart cities generate massive volumes of heterogeneous and dynamic data from various sources, including traffic sensors, energy meters, air quality monitors, and IoT devices. These data streams are often non-linear, highly volatile, and exhibit complex temporal and spatial dependencies. Managing such complexity requires predictive models capable of learning from multivariate time-series data while adapting to rapidly changing conditions.

Traditional models such as CNNs, SVMs, GRUs and others have made significant contributions to time-series forecasting but show inherent limitations when applied in smart city contexts. CNNs are effective in spatial pattern extraction but struggle with long-term dependencies. SVMs perform well on smaller datasets but lack the scalability and temporal learning capacity needed for real-time urban data. GRUs offer computational efficiency but are often less accurate in modeling deep temporal patterns. Hybrid LSTM models address these challenges by integrating the long-term dependency learning strength of LSTM with the complementary features of other algorithms. This integration enables more accurate, robust, and adaptive forecasting, making hybrid LSTM models essential tools for supporting complex decision-making in smart city systems.

The primary objectives of hybrid LSTM models are threefold. First, they enhance prediction accuracy and robustness across varying time horizons, as demonstrated in vehicle speed prediction tasks that integrate ARIMA’s linear modeling with LSTM’s non-linear pattern recognition [2]. Similarly, Yadav [7] applied a CNN-LSTM hybrid model for short-term load forecasting in smart grids, demonstrating notable improvements in accuracy and computational efficiency. Second, they enable the integration of heterogeneous data types, such as combining time-series data with spatial inputs or unstructured formats such as images and text. This capability is particularly valuable for air quality forecasting, in which sensor data and satellite imagery can be jointly analyzed. Third, to optimize computational efficiency, lightweight architectures such as CNN-GRU-LSTM allow for high-performance training even in resource-constrained environments. In smart cities, hybrid LSTM models serve as key analytical tools by synthesizing diverse data sources to support policymaking and real-time operations, thereby contributing to sustainable urban development [8].

The adoption of hybrid LSTM models is driven not only by technological advances, but also by societal demands. Accelerated urbanization, climate change, energy crises, and post-pandemic public safety concerns have highlighted the need for sophisticated real-time data analysis and forecasting systems. Therefore, hybrid LSTM models have been widely studied and deployed in both academic and industrial settings. The proliferation of IoT sensor networks across cities has enabled the continuous collection of vast time-series datasets. Hybrid LSTM models are particularly well suited to processing these data in real-time, using LSTM to capture long-term patterns and auxiliary algorithms to account for short-term variations and environmental fluctuations [9]. These capabilities position hybrid LSTM as a core technology in a smart city ecosystem.

Prediction model requirements vary significantly across different smart city applications. To meet these diverse and domain-specific demands, the architecture of hybrid LSTM models can be flexibly adapted and designed accordingly. For instance, traffic management systems often require real-time, low-latency predictions to support immediate operational decisions. In contrast, air quality monitoring and public health/safety applications demand high spatial resolution and the capacity to capture spatiotemporal variability. Furthermore, most application domains require models capable of handling seasonal trends, non-linear dynamics, and data irregularities.

To the best of the authors’ knowledge, no prior review has specifically examined hybrid LSTM models within the comprehensive context of smart cities. While existing reviews have addressed AI applications or hybrid models in isolated domains—such as transportation or energy—these studies do not offer a unified framework centered on hybrid LSTM approaches across multiple smart city sectors. This paper addresses this gap by providing a comprehensive analysis of hybrid LSTM model applications across diverse smart city domains.

This review aims to analyze the application of hybrid LSTM models in smart city domains, such as transportation, environmental monitoring, energy management, public health, infrastructure maintenance, and waste disposal. It also provides a comprehensive perspective on the evolution, limitations, future expansion, and market potential of this emerging technology, underscoring the strategic significance of hybrid LSTM models in advancing smart city innovation.

2. Principles and Characteristics

This section analyzes the principles, development process, strengths, and weaknesses of hybrid LSTM models.

2.1. Mechanism

The operating principle of hybrid LSTM models leverages the strengths of each combined algorithm to analyze and predict time-series data from multiple perspectives. Fundamentally, LSTM is a type of recurrent neural network with an exceptional capability to learn data patterns over time. The core of the LSTM architecture consists of cell states and gate mechanisms, as shown in Figure 1. This structure, also represented in Figure 2a, which comprises input, forget, and output gates, regulates the importance of previous information, selectively incorporates new data, and determines appropriate outputs [10]. This architecture enables effective learning even with time-series data with long-term dependencies.

In hybrid LSTM models, the LSTM structure is combined with other algorithms. For example, gated recurrent units (GRUs) feature simplified gate mechanisms that enable rapid training. When integrated with LSTM, both computational efficiency and performance increase, allowing for the superior handling of problems requiring the simultaneous management of both long- and short-term dependencies.

Furthermore, the integration of ARIMA, a statistical model, into LSTM creates a structure in which ARIMA captures linear trends and seasonality while LSTM learns non-linear patterns. In this approach, data first passes through ARIMA to remove or extract linear components, and then the LSTM models the non-linearity of the residuals.

The data flows in the hybrid models vary according to integration method. In parallel integration, identical input data are fed simultaneously into multiple algorithms, generating outputs that are combined for final predictions. Conversely, in serial integration, data flow sequentially through algorithms. Examples include CNN outputs serving as LSTM inputs and the LSTM processing of ARIMA residuals.

Through these mechanisms, hybrid LSTM models can analyze and predict complex time-series patterns from multiple perspectives, delivering high performance across various smart city applications.

2.2. Technical Characteristics

Hybrid LSTM models enhance time-series prediction performance by combining LSTM’s fundamental architecture with other algorithms. Their core attributes include complexity, flexibility, and scalability. Complexity refers to the ability to learn complex data patterns by exploiting the advantages of multiple algorithms. Flexibility denotes the capability of adjusting the model structure for various domains and data types. Scalability represents the potential for model expansion using additional modules or layers. These characteristics are crucial in the multidimensional and complex data environments of smart cities. Complexity captures non-linear urban data characteristics, flexibility enables adaptation to diverse urban challenges, and scalability addresses expanding data volumes as cities grow.

Hybrid integration methods can be categorized as either parallel or serial. LSTM-Prophet models represent a notable example of predicting energy consumption by eliminating seasonality and irregularities. CNN-LSTM models offer another important implementation. In these models, while the CNN processes spatial data, the LSTM handles temporal data. This architecture supports applications in network security and solar power generation forecasting. Particularly noteworthy are the GRU-based RNNs that efficiently handle missing values and noise using Soft GRU mechanisms. This capability makes them effective for traffic congestion prediction.

2.3. Development

Hybrid LSTM models have evolved continuously in terms of structure and function alongside technological advancements. Early hybrid models simply combined LSTM with other algorithms to simultaneously model linearity and non-linearity. With the increase in data complexity, diversification of data types, and emergence of various applications, hybrid LSTM models have developed into more sophisticated structures.

Recent models incorporate attention mechanisms, allowing them to highlight critical components of their input data. This approach complements LSTM’s long-term dependency learning capabilities, improving prediction accuracy. Hybrid models combined with transformer architectures are also being developed. These strengthen parallel-processing capabilities and improve long-term dependency learning.

In smart city applications, models that consider spatial characteristics, such as CNN-LSTM and ConvLSTM, have received significant attention. ConvLSTM integrates convolutional operations into the LSTM structure and effectively learns spatiotemporal patterns. This integration exhibits excellent performance in traffic flow prediction and meteorological data analyses.

Hybrid models also incorporate ensemble techniques and improve prediction accuracy by combining predictions from various models. Recent studies have shown the active progression of model compression and optimization to increase computational efficiency. Examples include reducing the model parameters and training lightweight models through knowledge distillation.

In addition to the approaches already discussed, further extensions of LSTM have been exploited. Bidirectional LSTM (Bi-LSTM) processes input sequences in both forward and backward directions, improving context understanding in time-series prediction. LSTM with residual or skip connections enhances gradient flow and training stability, particularly in deeper architectures. Moreover, LSTM has been effectively combined with optimization algorithms (e.g., Particle Swarm Optimization, genetic algorithms) and signal processing techniques (e.g., wavelet transforms, empirical mode decomposition) to improve hyperparameter tuning, feature extraction, and noise reduction.

These developments have expanded the application scope of hybrid LSTM models and improved their performance across various domains.

2.4. Advantages and Disadvantages

Hybrid LSTM models have the following advantages:

• High prediction accuracy and generalization ability: Compared to single models, hybrid LSTM models achieve higher prediction accuracy by combining the strengths of multiple algorithms. Considering both linearity and non-linearity simultaneously enables the effective learning of complex data patterns and ensures consistent performance across various scenarios.
• Processing of complex non-linear data: These models handle complex non-linear data patterns by leveraging LSTM’s long-term dependency learning capabilities alongside the specialized functions of other algorithms. This is particularly useful for analyzing complex time-series data in smart cities, including traffic flow, energy consumption, and environmental changes.
• Application potential across diverse fields: Hybrid LSTM models can be applied to various smart city domains, including transportation, energy, environment, health, and public safety. Their flexible structural adjustments based on data characteristics enable a broad application scope.
• Real-time data-processing capability: These models effectively process and predict large volumes of real-time data, contributing to smart city real-time monitoring and rapid decision-making.

They also have the following disadvantages:

• Complexity of model structure: Their complex structures combining multiple algorithms increase the training time and complicates model implementation and tuning. This results in increased computational resource requirements.
• Requirement of large data and risk of overfitting: More complex models require larger training datasets. Insufficient data creates an overfitting risk, which degrades the generalization ability of the model.
• Limitations in interpretability: The complex structures of hybrid models create difficulties in interpreting internal mechanisms and prediction results. This can affect the transparency and reliability of the models.
• Importance of data quality and preprocessing: When combining various data sources, data quality and preprocessing significantly influence model performance. Handling data noise and missing values are crucial.

Several studies have addressed these limitations. Computational efficiency has been improved through model compression and optimization. Regularization techniques and data augmentation can prevent overfitting. Explainable AI (XAI) technology has been actively incorporated to enhance model interpretability. These efforts have aimed to increase the practicality and reliability of hybrid LSTM models.

3. Case Studies

This section examines the implementation of hybrid LSTM models across key smart city domains by analyzing real-world applications and their performances. This review analyzes use cases across six major areas: traffic management, air quality and environmental monitoring, energy management, health and public safety, urban infrastructure monitoring, and waste management.

3.1. Traffic Management

Traffic prediction for management purposes is challenged by the heterogeneous nature of data collected from IoT devices, sensors, and GPSs. These data are often noisy, incomplete, and exhibit complex temporal and spatial dependencies. Traditional models like ARIMA, SVMs, and CNNs struggle to capture both the linear and non-linear dynamics of urban traffic flow. Therefore, hybrid deep learning approaches are increasingly essential for achieving accurate and robust traffic forecasting.

IoT-based hybrid architectures have been widely explored for traffic management applications. Chahal et al. [11] proposed a hybrid model that combined seasonal autoregressive integrated moving average (SARIMA) with bidirectional long short-term memory (Bi-LSTM) for traffic congestion forecasting. SARIMA captures the linear components of IoT-based time-series data, whereas Bi-LSTM addresses non-linear patterns. The model further incorporates a Backpropagation Neural Network (BPNN) to optimize its combined output. By modeling the non-linear relationship between predictions, this multi-model approach improves forecasting accuracy and offers a more advanced alternative to basic model stacking.

These models exhibit statistically significant improvements over conventional methods by effectively capturing complex non-linear patterns in urban traffic data.

Figure 2. (a) LSTM architecture; (b) Bi-LSTM architecture [12].

Figure 2. (a) LSTM architecture; (b) Bi-LSTM architecture [12].

The dataset was sourced from the CityPulse EU FP7 project data, which included time-series data on vehicle speed and traffic volume. The proposed model was evaluated using mean absolute error (MAE), mean squared error (MSE), root mean square error (RMSE), and mean absolute percentage error (MAPE). The hybrid model consistently outperformed both the standalone LSTM and ARIMA models across all evaluation metrics.

The quantitative results demonstrated significant performance gains: the hybrid model achieved an MAE of 0.499 (29% improvement over standalone LSTM’s 0.705), MSE of 0.337 (34% reduction compared to LSTM’s 0.510), RMSE of 0.58 (compared to LSTM’s 0.715), and MAPE of just 3% [11]. These metrics confirm the substantially enhanced predictive capabilities.

The results validate the effectiveness of the hybrid approach in addressing single-model limitations through the comprehensive modeling of both linear and non-linear data components. The proposed SARIMA–Bi-LSTM architecture represents a viable solution for smart city traffic prediction and management, with direct implications for urban mobility optimization and service quality enhancement.

In a follow-up study, Chahal et al. [12] introduced ARIMA-Bi-LSTM using an Aquila Optimizer (AB-AO) to improve traffic congestion forecasting in smart cities. As depicted in Figure 2b, the model combines ARIMA for capturing linear trends with Bi-LSTM for learning non-linear patterns. The performance was further enhanced through the hyperparameter optimization of the Bi-LSTM component using an Aquila Optimizer (AO). The model was tested using traffic data from three cities provided by the CityPulse EU FP7 project.

The AB-AO model outperformed standalone ARIMA, Bi-LSTM, and the conventional ARIMA-LSTM combination across all evaluation metrics. It achieved the lowest MAE of 3.18, compared to 4.63 (ARIMA), 4.20 (Bi-LSTM), and 3.98 (ARIMA-LSTM) [12]. Similarly, it recorded the lowest MSE (18.74) and the lowest MAPE (0.21), significantly improving the forecast accuracy over competing models [12].

These results confirm the ability of the AB-AO model to effectively capture complex traffic patterns and highlight the value of AO-based optimization for enhancing predictive performance. This study positions the AB-AO approach as a high-performance and scalable solution for intelligent traffic management in smart cities.

Focusing on the practical application of IoT-based traffic data, Culita et al. [13] proposed the incorporation of Autoregressive Moving Average (ARMA) into LSTM to address congestion issues in Bucharest, Romania. By comparing the performance of the ARMA–LSTM model with that of a standard LSTM model, their study evaluated the benefits of the hybrid approach in traffic forecasting. Both models were assessed under short-term (25 min) and long-term (50 min) forecasting scenarios using RMSE and MAE as the primary evaluation metrics, as shown in Figure 3.

In short-term forecasting, the ARMA–LSTM model delivered RMSE values 8–12% lower than those of standalone LSTM [13]. For example, in one segment, LSTM recorded an RMSE of 13.22, while the hybrid model achieved an RMSE of 12.14 [13]. The MAE results followed a similar pattern, indicating greater precision in short-term predictions.

For long-term forecasting, the performance gap increases further. The ARMA–LSTM model recorded an RMSE of 16.42, compared to 18.77 from the LSTM, representing a 13% reduction in error [13]. The MAE was also 10–15% lower on average, demonstrating the effectiveness of the hybrid model in enhancing forecast accuracy across both time horizons [13].

The results demonstrate that the ARMA–LSTM model effectively addresses the structural limitations of standard LSTM models and delivers improved prediction accuracy. This enhancement is driven by the ARMA component, which enables the model to better capture the complex temporal patterns inherent in time-series data. Particularly for dynamic, non-linear datasets such as traffic flow, the ARMA–LSTM model offers more refined forecasts, underscoring its practical value in real-time traffic management applications.

Zafar et al. [14] proposed a hybrid LSTM-GRU model to predict traffic speed in smart cities by integrating heterogeneous data sources, including IoT sensor feeds, Google Maps traffic data, weather conditions, and holiday schedules, as shown in Figure 4. The model combines the temporal learning strength of LSTM with the computational efficiency of GRUs to capture complex spatiotemporal traffic patterns. To increase prediction accuracy, the model was trained and validated using large-scale traffic data collected at 15 min intervals in Islamabad, Pakistan.

According to Figure 5, the model demonstrated superior predictive accuracy over the standalone models. The hybrid LSTM-GRU model achieved an RMSE of 4.5, outperforming LSTM (4.86) and the GRU (5.05) [14]. It also recorded a lower MAE of 2.03, compared with 2.13 for LSTM and 2.29 for the GRU [14]. In terms of the MAPE, the hybrid model achieved the lowest error of 6.67%, whereas the LSTM and GRU recorded errors of 6.95% and 7.7%, respectively [14].

These results confirm the effectiveness of the hybrid LSTM-GRU model in processing complex urban traffic data. This approach offers practical benefits for real-time traffic management and system optimization in smart city applications.

Sengupta et al. [15] proposed a hybrid traffic flow prediction model that combined a Hidden Markov Model (HMM) with LSTM to improve short-term forecasting, ranging from 5 min to 1 h. Their objective was to enhance traffic management and alleviate congestion by accurately capturing short-term fluctuations.

The proposed model combines the ability of the HMM to identify shifts in the traffic state with LSTM’s ability to effectively learn temporal dependencies. Among the two model architectures presented, the concatenated hybrid (C-Hybrid) structure, which combines the outputs from two LSTM networks, achieved the highest prediction accuracy.

Model performance was evaluated using real-world data from the California Department of Transportation’s Performance Measurement System (PeMS). The C-Hybrid model recorded an RMSE of 0.4203 [15]. This is significantly lower than those of the stacked LSTM (0.8235) and AR-HMM (0.8766) models, reflecting an improvement of up to 52% [15]. These results demonstrate the superior ability of the hybrid model to predict sudden changes in traffic flow.

The study highlights the potential of combining an HMM and LSTM to improve prediction accuracy and reliability, offering practical value for advanced traffic management systems.

Chaoura et al. [16] suggested a hybrid LSTM model incorporating Particle Swarm Optimization (PSO) to enhance intersection traffic flow prediction. Their study aimed to mitigate intersection congestion and improve road safety, emergency response time, and overall traffic efficiency.

The hybrid architecture integrates LSTM’s robust temporal learning capabilities with PSO’s exceptional optimization functionality. PSO, which mimics collective intelligence in natural systems to identify optimal solutions, effectively optimizes the LSTM hyperparameters (learning rate and network architecture) to enhance predictive accuracy. This optimization process identifies the optimal model configurations, thereby enabling precise traffic predictions.

The study utilized 48,120 actual vehicle passage records collected from four distinct intersections in Nanjing, China, between November 2015 and June 2017 [16]. The data were preprocessed to optimize the characteristics of each intersection. The performance evaluation demonstrated that the proposed model outperformed the conventional LSTM, Random Forest (RF), k-neighbor (KN), and Decision Tree (DT) models across all intersections. For instance, at intersection 1, the model achieved an RMSE of 0.15258, indicating highly accurate predictions.

The performance assessment of the LSTM-PSO model showed its superior performance at all intersections compared with standalone LSTM and the widely used RF, KN, and DT models, as illustrated in Figure 6 and Figure 7. At intersection 1, the model recorded an RMSE of 0.15258 and an MAE of 0.0898, demonstrating exceptional accuracy [16]. These results indicate the robust predictive capabilities of the LSTM-PSO model across complex and diverse intersecting environments.

In addition, the results validate the practical applicability of the LSTM-PSO hybrid approach for real-time traffic signal control and traffic management system implementation. Moreover, the results have significant potential contributions to the development and deployment of smart cities and intelligent transportation systems (ITSs).

In bicycle traffic flow forecasting, hybrid models have been proposed to address the non-linearity and complex distribution characteristics of time-series data. Myhrmann and Mabit [17] developed a Long Short-Term Memory Mixture Density Network (LSTMMDN) model for precise hourly bicycle flow prediction across urban areas. This architecture provides accurate exposure data for policy analyses such as bicycle traffic safety assessments.

The LSTMMDN integrates Mixture Density Network (MDN) components to predict various possible bicycle flow distributions probabilistically under identical conditions. This incorporation enables more realistic predictions than conventional models, which only forecast mean values, effectively capturing the traffic volume variability across different times of the day, days of the week, and weather conditions.

This study analyzed 64,664 hourly observations from bicycle counters in Copenhagen, Denmark, from 2017 to 2020, incorporating 17 weather and time variables [17]. Performance comparisons against the Danish Road Directorate’s Seasonal Variation Factor (SVF) method and standard LSTM showed that the LSTMMDN achieved an up to 77% accuracy improvement compared with SVF methods, with an MSE of 0.102 [17]. Meanwhile, SVF recorded an MSE of 0.377 [17]. The model also consistently outperformed the standard LSTM implementations.

Visual analysis in Figure 8 confirmed the LSTMMDN’s superior representation of actual traffic patterns. Using only 5% of the available data, the model maintained more than 50% higher accuracy than SVF, demonstrating strong performance with limited datasets [17]. In safety applications, LSTMMDN predictions improved the explanatory power of accident risk analysis by 5.5%, highlighting the importance of precise exposure data in safety assessments [17]. This research establishes the LSTMMDN as an effective tool for traffic management and policy development.

Miao and Liao [18] proposed a hybrid CNN–LSTM-PSO model to improve traffic congestion forecasting in smart cities by leveraging real-time IoT-enabled sensor data. The goal was to enhance prediction accuracy and enable dynamic traffic signal control based on urban mobility patterns.

The model architecture integrates CNNs for spatial feature extraction, LSTM for capturing temporal dependencies in traffic patterns, and PSO for hyperparameter tuning. Sequential trends—such as vehicle flow, speed, and incidents—were modeled by embedding LSTM layers within the CNN framework. PSO effectively tuned key parameters, including learning rate and filter configurations, to accelerate convergence and enhance model stability.

Real-time traffic data were collected over one year from 50 intersections, generating approximately 1.2 million sensor records, with features such as vehicle count, speed, traffic density, incident indicators, and weather. Preprocessing included interpolation, normalization, and augmentation. The dataset was split into training (70%), validation (15%), and testing (15%).

The CNN–PSO model achieved 92% accuracy in congestion classification and reduced the MSE by 15% compared to baseline models [18]. It recorded an RMSE of 0.030 and an MAE of 0.025, outperforming SVM and RF approaches [18]. Its precision, recall, and F1-scores were 90%, 89%, and 89.5% [18]. Real-time deployment simulations showed a 20% reduction in average traffic delay and a 25% improvement in traffic flow efficiency [18].

A notable feature of the study was its real-time integration with traffic infrastructure, enabling dynamic signal control based on live LSTM-informed predictions. The model demonstrated low latency, high adaptability across urban conditions, and reduced computational overhead through PSO-based tuning compared to exhaustive search methods.

This work highlights the effectiveness of hybrid LSTM models in smart cities—fusing spatial and temporal learning with intelligent optimization to address complex, dynamic traffic systems. By embedding LSTM within a multimodal architecture, the model accurately captured non-linear and evolving patterns in urban mobility, offering a scalable, low-latency solution for intelligent transportation systems (ITSs) and real-time traffic management

3.2. Air Quality and Environmental Monitoring

Air quality and environmental factor prediction in smart cities is difficult due to noisy, incomplete, and varied sensor data, and rapidly changing conditions influenced by traffic, industry, and weather. Traditional models struggle with long-term and spatiotemporal patterns, while hybrid deep learning models improve accuracy and enable real-time decisions.

In a smart city, managing the environment through accurate predictions directly affects citizens’ quality of life. Environmental monitoring primarily focuses on forecasting air quality and energy production.

Kataria and Puri [19] proposed a hybrid CNN-LSTM model for air quality prediction in smart cities. Their proposed architecture efficiently processes atmospheric pollution data to accurately forecast air quality index (AQI) values. They enhanced the prediction accuracy by combining the CNN architecture, which effectively extracts key features from data, with traditional LSTM, which learns temporal dependencies. Model performance was evaluated using two extensive public datasets from Beijing and Chennai.

The CNN-LSTM architecture processes data sequentially, beginning with the CNN learning spatial characteristics and transferring the extracted features to LSTM, followed by LSTM analysis of temporal dependency. This approach enables more sophisticated data interpretation and higher prediction accuracy compared with the operations performed by conventional LSTM models.

The results in Figure 9 demonstrated that the CNN-LSTM model significantly outperformed standalone LSTM models. Using a dataset from Beijing, China, the hybrid model recorded an RMSE of 23.14, which was approximately 3.5% lower than that of the LSTM model (23.97) [19]. For MAE, CNN-LSTM achieved 14.89, which was approximately 6.3% lower than LSTM’s 15.89 [19]. The Chennai dataset showed similar improvements, with CNN-LSTM recording an RMSE of 30.85 and MAE of 26.69, both lower than LSTM’s 31.52 and 29.89, respectively [19].

These findings confirm that the CNN-LSTM model processes the spatiotemporal complexity of air quality data more effectively than standalone LSTM architectures. This approach is a valuable tool for predicting air pollution in smart cities, implementing early warning systems, supporting environmental policy development, and mitigating health risks.

To forecast the photovoltaic output in smart urban environments, Kim et al. [20] incorporated a seasonal autoregressive integrated moving average with exogenous factors (SARIMAX) model into LASTM in a stacked ensemble framework. Figure 10 shows the overall flow of Kim et al.’s study. Kim et al. attempted to incorporate SARIMAX’s temporal processing capabilities into LSTM’s advanced pattern recognition to improve prediction accuracy across diverse geographical locations.

The research team used predictions from the SARIMAX model, which excels in capturing seasonal characteristics and exogenous factors, as training data for the LSTM. It could handle both long- and short-term dependencies in time-series data to generate more refined forecasting results.

SARIMAX excels in analyzing data by considering seasonal characteristics and exogenous variables and effectively capturing temporal dependencies and seasonality. LSTM can simultaneously process both long- and short-term dependencies in time-series data using predictions from SARIMAX as training data to generate more refined results. The study evaluated the performance of the proposed SARIMAX-LSTM model using solar power plant data from three regions in South Korea: Incheon, Busan, and Yeongam.

As depicted in Figure 11, the empirical analysis revealed substantial improvements with the integrated approach: an MAE of 64.730, representing a 15.8% reduction compared to traditional LSTM’s 76.913; an RMSE of 95.800, outperforming standard LSTM’s 106.123 by 9.7%; and a symmetric mean absolute percentage error (SMAPE) of 19.891%, further confirming the superior performance compared to LSTM’s 23.369% by 15% [20].

These findings confirm the superior capability of the SARIMAX–LSTM model in reflecting seasonality and long-term dependencies in time-series data, providing enhanced precision for smart city solar power generation forecasting. This model is an important tool that can contribute to establishing efficient power supply plans and promoting the effective utilization of energy resources.

Mukhtar et al. [21] introduced a CNN-LSTM-ANN hybrid architecture for solar radiation prediction in smart cities. This framework focuses on learning spatiotemporal data patterns to more accurately forecast hourly solar radiation variations. As shown in Figure 12, the model employs a CNN to extract non-linear features, LSTM to learn temporal dependencies, and an ANN to perform the final prediction calculations, thereby enhancing the overall performance.

The distinctive feature of the CNN-LSTM-ANN structure is its hierarchical organization: the CNN extracts spatial patterns from input data, LSTM processes long-term sequential dependencies, and the ANN computes the comprehensive prediction values. This architecture was specifically engineered to effectively manage the spatiotemporal complexity of environmental data. The study evaluated the model’s performance across varying climatic conditions using datasets from ten African countries.

The assessment results demonstrated that the CNN-LSTM-ANN model had superior predictive capabilities compared with the traditional ANN and CNN-ANN implementations. In Ethiopia, the CNN-LSTM-ANN recorded an MAE of 25.89 W/m², RMSE of 72.27 W/m², and MAPE of 9.31%, showing improvements over the ANN model’s metrics (MAE: 28.13 W/m²; RMSE: 73.81 W/m²; MAPE: 10.28%) [21]. Similarly, in Somalia, the hybrid approach achieved an MAE of 16.60 W/m², RMSE of 51.54 W/m², and MAPE of 5.725%, again outperforming the ANN model (MAE: 22.21 W/m²; RMSE: 52.80 W/m²; MAPE: 7.78%) [21].

The study demonstrates how the CNN-LSTM-ANN hybrid architecture enhances prediction accuracy by capturing complex spatiotemporal patterns in environmental data. This model contributes to the efficient management and utilization of renewable energy resources in smart urban environments. Such advanced modeling approaches are critical in improving the reliability of energy management systems against climate variability and in strengthening sustainability initiatives in smart city development.

Zaini et al. [22] proposed a hybrid deep learning framework to improve the accuracy of urban PM2.5, aiming to maintain environmental sustainability and reduce health risks from air pollution. Their research addresses the deteriorating air quality in Kuala Lumpur, Malaysia, resulting from rapid industrialization, urbanization, traffic congestion, and increased energy consumption. The accurate prediction of PM2.5, which has a significant impact on public health and the environment, is essential.

The research team developed a hybrid architecture based on LSTM, integrating the Ensemble Empirical Mode Decomposition (EEMD) data decomposition technique with the metaheuristic optimization algorithms PSO and Sparrow Search Algorithm (SSA). Initially, they constructed PSO-LSTM and SSA-LSTM models to optimize the hyperparameters. Then, they developed EEMD-PSO-LSTM and EEMD-SSA-LSTM models based on the best-performing input configurations. In this architecture, LSTM learns long-term dependencies in time-series data, EEMD decomposes complex non-linear data into simpler components, and PSO/SSA handles LSTM hyperparameter optimization. The SSA demonstrated a more stable convergence than PSO and utilized the Levy Flight Trajectory to address local optimization issues.

This study utilized hourly time-series data collected from nine air quality monitoring stations and three meteorological stations in Malaysia from 1 January to 31 December 2019. The dataset included six air pollution variables (PM2.5, PM10, SO₂, NO₂, O₃, and CO) and five meteorological variables (temperature, relative humidity, wind speed, wind direction, and precipitation). The primary target areas were Batu Muda and Cheras, Kuala Lumpur. Missing values were addressed using linear interpolation, and the data were split into 70% training and 30% testing segments.

The performance evaluation used RMSE, MAE, MAPE, and the coefficient of determination (R²). The initial models showed that the input variable combination M4 (air pollutants excluding PM2.5, SO₂, temperature, and relative humidity) delivered the best prediction performance. The SSA-LSTM model demonstrated a more stable performance than the PSO-LSTM model across multiple iterations. Models incorporating EEMD (EEMD-SSA-LSTM and EEMD-PSO-LSTM) generally outperformed the models without decomposition techniques. Notably, the EEMD-SSA-LSTM model achieved the best performance when PM2.5 was included from adjacent monitoring points (M5) as an additional input. In the Batu Muda area, this approach improved the RMSE by 2.65% and MAE by 9.31%, whereas in Cheras the improvements were 20.00% and 25.30%, respectively, as shown in Figure 13 [22].

The study demonstrates that a hybrid approach combining LSTM, data decomposition (EEMD), metaheuristic optimization (SSA), feature selection, and spatial data (neighboring PM2.5 information) effectively enhances the accuracy of air quality prediction. Such forecasting models can support practical response strategies, including environmental policy development, early warning systems, and emission control during pollution peaks, thereby contributing to environmental sustainability and public health protection. Future research could include incorporating meteorological variables from adjacent areas, improving computational efficiency, enhancing the day/night prediction performance, and applying the model to datasets from diverse regions.

Kaveh et al. [23] introduced a novel hybrid model called OA-LSTM for air pollution management, specifically for PM2.5 prediction. This architecture combines the Orchard Algorithm (OA) with Long Short-Term Memory networks to address the fundamental limitations of traditional LSTM implementations, including local minimum entrapment, high computational costs, and dependency on continuous objective functions, as illustrated in Figure 14. The OA employs a metaheuristic global search strategy inspired by orchard dynamics to efficiently perform optimization in high-dimensional parameter spaces. Additionally, the researchers implemented the Binary Orchard Algorithm (BOA) for sensitivity analysis and feature selection, optimizing the model by maintaining key predictive variables while minimizing prediction errors.

The study utilized meteorological data, topographical features, PM2.5 pollution levels, and satellite imagery from Tehran, Iran, from 2014 to 2016. To prepare the model input dataset, data preprocessing involved noise reduction using Savitzky–Golay filtering, spatial interpolation through inverse distance weighting (IDW), and missing-value reconstruction via spline interpolation. The performance evaluation of the OA-LSTM framework employed multiple metrics, including RMSE, R², Standard Deviation Error (SDE), convergence trends, and computational complexity.

When compared against five machine learning models—traditional LSTM, an RNN, a Deep Neural Network (DNN), Random Forest (RF), and a Support Vector Machine (SVM)—the OA-LSTM architecture demonstrated superior performance, with an RMSE of 3.01 µg/m³ and R² of 0.88 [23]. In contrast, conventional LSTM recorded an RMSE of 9.53 and R² of 0.65, indicating substantially lower performance [23]. OA-LSTM also exhibited greater stability with an SD Error of 0.05, outperforming all comparison models [23]. Computational efficiency testing revealed that OA-LSTM reached an RMSE below 10 µg/m³ within 173 s, as shown by the convergence trend in Figure 15 [23]. Meanwhile, the RNN required 634 s, and the other models either failed to achieve this threshold or demanded significantly more processing time [23]. Through BOA-based feature selection, OA-LSTM achieved an RMSE of 5.12 µg/m³ using only nine critical features, surpassing the performance of the RF and SVM models that utilized more extensive feature sets [23].

This study demonstrates that integrating LSTM with an innovative OA optimization technique effectively addresses air pollution prediction challenges. The OA optimizes the LSTM weight and bias configurations, enhancing the convergence speed and model efficiency, whereas the BOA maximizes the prediction accuracy while minimizing variable redundancy. The proposed framework offers a scalable and efficient solution for urban air quality management and early warning systems for extreme pollution events.

Ratković et al. [24] conducted a comparative analysis of SVM and hybrid LSTM regression models for air pollution prediction in Niksic, Montenegro. Their study aims to forecast the Common Air Quality Index (CAQI) for upcoming periods and proposes a hybrid model that integrates a Genetic Algorithm (GA) with LSTM for hyperparameter optimization. The GA systematically searches for optimal combinations of time-step quantities and LSTM units per layer, thereby enhancing prediction accuracy and model reliability while preventing overfitting. Additionally, researchers have implemented bagging techniques to create an ensemble LSTM configuration, train multiple models on different data subsets, and combine their predictions to further improve performance.

This study used air pollution concentration data collected from monitoring stations in the Niksic region of Japan. After data preprocessing and feature engineering, the dataset was partitioned into training and testing segments and normalized. The performance evaluation employed multiple metrics, including the MSE, MAE, MAPE, and R², with an additional statistical analysis of the MAE values, summarized in

Table 1. Evaluation results of SVM, optimized LSTM, and ensemble models for hourly CAQI prediction [24].

Performance Metric	SVM Model	Best LSTM Model	Ensemble Model
MSE	428.2885	58.8046	56.5810
MAE	13.8739	4.6657	4.5560
MAPE	0.2685	0.0923	0.0912
$R^2$	0.3820	0.9135	0.9168

.

The ensemble LSTM architecture demonstrated superior performance compared with the SVM model, with an MSE of 56.5810, MAE of 4.5560, MAPE of 0.0912, and R² of 0.9168 for CAQI prediction [24]. In contrast, the SVM model achieved a considerably lower performance, with an MSE of 428.2885, MAE of 13.8739, MAPE of 0.2685, and R² of 0.3820 [24]. Furthermore, the ensemble LSTM model showed improved metrics compared with the single LSTM implementation (MSE: 58.8046; MAE: 4.6657) and outperformed conventional LSTM architectures, as illustrated in Figure 16 [24].

The study was the first demonstration of the effectiveness of a hybrid model combining LSTM, GA, and bagging techniques for air pollution prediction in Montenegro. The findings confirm that this integrated approach can significantly enhance the accuracy and reliability of hourly air pollution forecasts. The proposed modeling framework can be provided to relevant authorities to support pollution-level predictions and facilitate appropriate response measures.

Zeng et al. [25] proposed an innovative forecasting model combining the Extended Stationary Wavelet Transform (ESWT) with Nested Long Short-Term Memory (NLSTM). This was carried out in order to overcome the prediction limitations faced by conventional deep learning models, owing to time-series data volatility and the need for extensive historical records. ESWT extends the traditional Stationary Wavelet Transform (SWT) by decomposing the original data into high- and low-frequency components and then applying SWT again to generate six more-stable sub-signals, as illustrated in Figure 17. These components are individually processed through separate NLSTM networks for prediction, and the results are integrated using an Inverse Wavelet Transform (IWT). The NLSTM architecture enhances the long-term dependency learning capability by embedding additional LSTM cells within existing LSTM structures, thereby effectively expanding the memory capacity.

This study utilized univariate PM2.5 time-series data collected from 12 monitoring stations around Beijing, China, from 2013 to 2017. The data underwent decomposition through ESWT into six sub-signals before being fed into the individual NLSTM models. The performance evaluation employed multiple metrics, including MAE, RMSE, MAPE, and R². The comparison of final prediction results and values for a subset of 72 data samples is shown in Figure 18.

The proposed ESWT-NLSTM framework demonstrated superior performance across all evaluation metrics compared to various machine learning- and deep learning-based time-series prediction models. Specifically, ESWT-NLSTM recorded an MAE of 3.456, RMSE of 5.579, MAPE of 11.61%, and R² of 0.990, significantly outperforming both standard NLSTM (MAE = 9.290; RMSE = 17.968; MAPE = 31.60%; R² = 0.895) and EMD-NLSTM (MAE: 7.248; RMSE: 11.292; MAPE: 22.75%; R²: 0.959) implementations [25]. The absolute error comparison between the proposed and existing models is illustrated in Figure 19. Notably, the model substantially reduced prediction lag effects, providing crucial advantages for peak value forecasting.

This study represents the first application of NLSTM for air pollution prediction. It demonstrates that the integration of ESWT with NLSTM creates an optimal hybrid architecture for highly volatile univariate time-series data. Improvements in prediction accuracy and a reduction in lag effects make this approach particularly suitable for operational air quality forecasting systems, especially pollution alert systems, which potentially play a vital role in public health protection.

Dalal et al. [26] developed a hybrid PSO–LSTM–RNN deep learning model for real-time air quality prediction in smart cities, targeting pollutants—such as SO₂, CO, O₃, and NO₂—to support urban environmental management and health risk mitigation.

The hybrid architecture integrated LSTM’s ability to model long-term dependencies with RNNs’ capability to capture sequential relationships in time-series data. PSO was employed specifically to optimize the LSTM neural network’s weights, enhancing predictive accuracy by efficiently exploring high-dimensional parameter spaces. Additionally, a dropout regularization technique was integrated into the LSTM layer to address overfitting, improving the model’s generalization capability in dynamic environments.

Compared to Gradient-Boosted Tree Regression, standalone LSTM, and SVM regression, the hybrid PSO–LSTM–RNN demonstrated superior performance—recording a significantly lower RMSE of 0.0184, MAE of 0.0082, and higher R² of 0.1227 [26]. This highlights its effectiveness in managing complex, non-linear air quality patterns common in urban environments.

The LSTM component was instrumental in leveraging gating mechanisms—input, forget, and output gates—to model complex time-series patterns, particularly in fluctuating pollutant concentrations from sensor networks. This contributed to enhanced long-term learning essential for real-time environmental monitoring.

This study demonstrates the practical applicability of hybrid LSTM-based models in smart cities for environmental monitoring and decision support. By incorporating PSO for precise weight optimization and combining multiple robust deep learning techniques, the proposed hybrid architecture offers an accurate and computationally efficient solution for real-time air quality forecasting in urban settings.

While the reviewed hybrid LSTM models demonstrate high accuracy in air quality and environmental monitoring, it is worth noting that most of the reviewed models were trained and validated using historical datasets. Consequently, their robustness against extreme data variability, long-term climate change, or rare events remains largely untested. This highlights a critical research gap and underscores the need for future studies to evaluate model performance under non-stationary or atypical conditions.

3.3. Energy Management

This section focuses on electrical energy consumption forecasting, covering various energy sources including natural gas, hydroelectric, wind, solar, and waste-to-energy conversion. The energy management approaches discussed primarily target smart grid infrastructure, smart homes, and smart buildings within urban environments.

Energy consumption forecasting in smart cities is challenging due to high-frequency, non-linear, and seasonal patterns driven by factors like population growth and weather. Traditional models often miss complex dependencies, while deep learning models may overfit or generalize poorly. Hybrid deep learning with decomposition techniques helps manage time-series variability, enabling more accurate predictions for planning and sustainability.

Consumption forecasting is a core research topic in the energy management domain. Arslan [27] proposed a hybrid model combining Prophet with Stacked Bidirectional LSTM to enhance energy consumption prediction in smart cities. Prophet offers strengths in decomposing time-series data into seasonality, trend, and irregularity components for separate analysis, whereas Stacked Bidirectional LSTM can simultaneously leverage both temporal sequences and future information, as illustrated in Figure 20. This integrated approach was designed to improve prediction accuracy by preserving seasonal patterns while reducing the impact of irregular fluctuations.

The key feature of the hybrid architecture lies in its complementary methodology: using Prophet to maintain seasonal characteristics while employing Stacked Bidirectional LSTM to learn non-linear and complex patterns in decomposed data. The framework implements Seasonal-Trend Decomposition using loess (STL) to separate data into seasonality, trend, and residual components. Prophet processes the original data to predict basic trends, including seasonality, whereas Stacked Bidirectional LSTM learns from the residual data to generate more sophisticated predictions. The final forecast combined the outputs of both models.

The performance evaluation demonstrated the hybrid model’s superior results compared to standalone Prophet, Bi-LSTM, and deseasonalized LSTM (deBi-LSTM) implementations. For the Canadian dataset, the hybrid model achieved an RMSE of 2348.23, approximately 43% lower than Prophet’s 4119.49 and 12% lower than Bi-LSTM’s 2670.47 [27]. Similarly, for the French dataset, the hybrid model recorded an RMSE of 2147.37, showing improvements of 45% compared with Prophet (3931.96) and 1.3% compared with Bi-LSTM (2176.40) [27].

In conclusion, the hybrid model integrating Prophet with Stacked Bidirectional LSTM demonstrated enhanced accuracy in smart city energy consumption forecasting by simultaneously addressing seasonality and irregularity, compared with conventional approaches. This advancement has the potential to improve energy management efficiency in smart cities and enhance the precision of resource allocation and planning processes.

Ali et al. [28] introduced two hybrid models––ConvLSTM and CNN-LSTM—to predict energy consumption in smart homes. The primary objective was to improve the energy management efficiency in smart homes through accurate forecasting based on time-series data. ConvLSTM features an architecture capable of simultaneously processing two-dimensional spatial features and temporal dependencies. In contrast, CNN-LSTM is designed such that the CNN learns spatial features first, followed by LSTM learning temporal dependencies.

The key characteristic of the ConvLSTM model is its ability to learn spatiotemporal patterns through 2D ConvLSTM layers and then optimize temporal dependencies in subsequent LSTM layers to deliver high prediction accuracy. The CNN-LSTM operates by extracting key features through 1D Conv layers and learning time-series data flows via LSTM layers. Both architectures effectively combine spatial and temporal patterns to provide superior results compared with conventional approaches.

Performance evaluation demonstrated that ConvLSTM had the highest accuracy across all test scenarios. In the 1-day, 3-day, and 6-day prediction scenarios, ConvLSTM achieved MAE values of 3.69, 4.07, and 4.65, respectively, outperforming CNN-LSTM’s MAE (3.80, 4.15, 4.69) and standalone LSTM (3.90, 4.24, 4.72) [28]. For MAPE, ConvLSTM recorded 18.48% for 1-day predictions, 21.27% for 3-day predictions, and 25.35% for 6-day predictions, showing improvements over CNN-LSTM (18.59%, 22.65%, and 25.20%, respectively) and LSTM (22.75%, 26.09%, and 28.03%, respectively), as illustrated in Figure 21 [28]. The RMSE values for ConvLSTM were also the lowest at 5.39, 5.88, and 6.54 for the 1-day, 3-day, and 6-day forecasts, respectively [28].

In conclusion, both ConvLSTM and CNN-LSTM demonstrated superior performance in smart home energy consumption prediction compared to the conventional standalone LSTM and CNN models. The ConvLSTM model maintained high accuracy even in long-term predictions and demonstrated excellent spatiotemporal pattern-learning capabilities. This research validates the idea that such hybrid architectures can serve as powerful tools for enhancing energy management efficiency in smart cities.

Alhussein et al. [29] proposed a hybrid CNN-LSTM model to enhance short-term electricity consumption forecasting for individual households in smart cities. Their study aimed to efficiently manage power supply and demand in smart grids while improving the accuracy of energy usage predictions. The model was evaluated using data from Australia’s Smart Grid Smart City (SGSC) project, leveraging the complementary characteristics of CNNs and LSTM to significantly enhance the predictive performance, as shown in Figure 22.

The key feature of the CNN-LSTM architecture lies in its dual processing approach: the CNN extracts spatial patterns from data, whereas LSTM learns temporal dependencies, effectively capturing both short- and long-term patterns in time-series data. The CNN filters noise from the input data and extracts critical features, which are then passed to the LSTM layers to learn sequential dependencies. This hybrid approach combines the strengths of both techniques to deliver a higher prediction accuracy than conventional models.

The results demonstrated that the CNN-LSTM model outperformed traditional standalone LSTM models across all evaluation metrics. In single-step forecasting, the CNN-LSTM model achieved a MAPE of 40.38%, which was approximately 8.3% lower than that of the conventional LSTM model (44.06%) [29]. For multi-step forecasting, the CNN-LSTM model surpassed the LSTM model, with MAPE reductions of 4.01%, 4.76%, and 5.98% for its 1 h, 2 h, and 6 h predictions, respectively, as shown in Figure 23 [29].

In conclusion, the CNN-LSTM hybrid model demonstrated excellent performance in predicting individual household energy consumption in smart cities. This architecture is a powerful tool that can enhance the efficiency of power management systems and contribute to energy conservation and cost reduction strategies.

Karri et al. [30] proposed an IoT-based hybrid LSTM-MLP model aimed at improving waste-to-energy conversion efficiency in smart cities. Their study focused on effectively processing extensive IoT data to achieve accurate energy predictions. Within this architecture, LSTM handles the learning of temporal characteristics in IoT data, whereas a Multilayer Perceptron (MLP) manages complex feature extraction. This structural design was developed specifically to enhance the accuracy and reliability of massive IoT datasets.

The study utilized time-series data obtained from the Kaggle platform, incorporating waste generation volumes, energy conversion rates, environmental factors such as temperature and humidity, and IoT sensor data. Data preprocessing involved handling missing values, detecting and processing outliers, and applying normalization techniques (Min-Max Scaling, Z-score Normalization). Feature engineering implements temporal characteristics (day of the week and time), interaction features, correlation analysis, Gini importance, and Principal Component Analysis (PCA) for feature selection. During model training, the parameters were optimized through backpropagation using loss functions (MSE, Cross-Entropy Loss), with hyperparameter tuning conducted via cross-validation and grid search methodologies.

The performance evaluation employed metrics such as accuracy, precision, recall, and F1-score. The results demonstrated that the hybrid IoT-LSTM-MLP model achieved an excellent performance, with 97.6% accuracy, 98.2% precision, 96.8% recall, and 97.5% F1-score [30]. These metrics represent more than a 6% improvement in accuracy compared with the conventional ARIMA and SVM models [30]. These findings validate the effectiveness of temporal data processing using LSTM combined with feature extraction via the MLP. From a practical application perspective, this model can be integrated into dynamic energy management systems in smart cities, significantly contributing to waste-to-energy generation optimization and operational efficiency improvements.

Kaya et al. [31] conducted research utilizing a hybrid CNN-LSTM model to enhance consumption and production forecasting accuracy for multiple energy sources across Turkey. Their study aimed to overcome the limitations of existing prediction models for handling data complexity and to implement a more precise predictive performance. Within this architecture, the CNN extracts localized patterns from data, whereas LSTM learns long-term temporal dependencies.

The study employed hourly energy consumption and production data from 2018 to 2023 provided by the Energy Exchange Istanbul (EXIST). The dataset encompasses total consumption and production figures, as well as production data for various energy sources, including natural gas, hydroelectric dams, lignite, run-of-river hydroelectric, wind, and fuel oil. After transforming the data using a sliding window approach, they applied MinMaxScaler for normalization and partitioned the dataset into training (80%), validation (10% of training data), and testing (20%) segments.

The performance evaluation demonstrated that the hybrid CNN-LSTM model achieved superior results compared to conventional Linear Regression (LR), Random Forest (RF), Support Vector Regression (SVR), and standalone CNN and LSTM models [31]. Notably, the model recorded an R² value of 0.990 for total energy consumption prediction and exhibited the lowest RMSE and MAE values, with the highest R² scores across most energy source predictions [31]. These results confirm that the combination of a CNN and LSTM effectively captures both the spatial and temporal complexities in this data, significantly enhancing prediction accuracy. This research contributes to improving energy resource utilization efficiency and power system management precision while suggesting potential future extensions incorporating additional variables such as economic indicators and meteorological conditions.

Du [32] proposed a hybrid SVR-LSTM model with Random Forest optimization aimed at improving building energy consumption prediction accuracy to enable effective energy resource allocation and waste reduction. Through an SVM-LSTM-stacking architecture combining the strengths of Support Vector Regression and Long Short-Term Memory networks, the study sought to effectively analyze the complex and diverse energy consumption data characteristics that conventional models struggle to process. Within this framework, SVR handles complex, high-dimensional data patterns, whereas LSTM effectively learns the temporal characteristics of the energy consumption data. Random Forest was implemented using a stacking approach for feature selection and model optimization in order to enhance the overall prediction performance.

The study used data from 539 buildings collected between 2017 and 2020. The dataset comprised 12 variables: building size, construction year, total floor area, cooling area ratio, average monthly occupancy, chiller age, central cooling plant efficiency, LED usage status, and annual energy consumption. The data preprocessing primarily involved normalization and correlation analysis for appropriate variable selection because the dataset contained minimal missing values and outliers.

The performance evaluation employed metrics such as accuracy, mean squared error (MSE), and coefficient of determination (R²). The results demonstrated that the hybrid model achieved over 90% accuracy on both the training and testing data, representing substantial performance improvements compared with standalone LSTM (R² of 0.542, approximately 54%) and SVR models (R² of 0.151, approximately 15%) [32]. Furthermore, the RF Stacking model exhibited a very low MSE (0.00047) with a high explanatory power (R² = 0.898) [32].

This hybrid architecture can be integrated into real-time monitoring and control systems for smart buildings to maximize energy efficiency while providing valuable data-driven guidelines for policy development and building design.

Sivarajan and Jebaseelan [33] proposed an innovative approach that combines a hybrid CNN-LSTM model with a Modified Sea Lion Algorithm (MSLA) for accurate energy demand prediction in IoT-based smart grids. To overcome the limitations of existing models in processing data complexity, they integrated a CNN’s feature extraction capabilities with the LSTM’s ability to learn time-series dependencies. Additionally, the MSLA was designed to further enhance model performance through hyperparameter tuning by implementing Levy Flight Trajectory techniques to address local optimization issues in the conventional Sea Lion Algorithm.

The study utilized the actual energy consumption data from various household appliances in smart homes, including televisions, lighting, air conditioners, refrigerators, fans, and water heaters. The data underwent authentication, encryption, and normalization processes, as well as preprocessing, to improve prediction accuracy and processing efficiency.

The model evaluation employed metrics including MSE, RMSE, MAE, MAPE, and prediction and training times. The results demonstrated that the CNN+LSTM+MSLA hybrid model outperformed the conventional CNN-LSTM, standalone LSTM, and Linear Regression models across per minute, hourly, daily, and weekly data intervals. For instance, with per minute data, the hybrid model recorded an MSE of 0.034, an RMSE of 0.200, an MAE of 0.084, and a MAPE of 12.67%, representing the best performance across all metrics [33]. Similarly, the lowest MSE (0.198) and RMSE (0.438) were achieved for the hourly data [33]. The model also demonstrated improvements in reducing the training and prediction times.

This study validates that the combination of a hybrid CNN-LSTM architecture with MSLA optimization is highly effective for energy demand prediction in smart grids. The proposed model enables the efficient allocation of energy resources and minimization of waste, contributing to the enhanced stability and sustainability of smart grid systems.

Jin et al. [34] proposed a hybrid LSTM–BPNN-to-BPNN model to enhance long-term electricity peak load forecasting, a critical task for smart city energy planning and infrastructure resilience. The model addresses the non-linear and multivariate nature of peak demand, which is influenced by historical load patterns, economic indicators, and weather conditions.

The architecture consists of three components: an LSTM module to extract temporal features from historical peak load data, a BPNN module to capture the non-linear structure of projected future variables (e.g., temperature forecasts, economic development targets), and a final BPNN that fuses both outputs to predict monthly peak loads. The highest predicted monthly value is then selected as the forecasted annual peak load. This structure allows the model to leverage both temporal dependencies and static future information, offering a robust framework for forward-looking energy demand estimation.

The model was trained on a real-world dataset from a Chinese city (2012–2021), where it demonstrated strong performance. Compared to conventional models—ARIMA (MAPE: 4.63%), SVR (2.04%), LSTM (3.55%)—the hybrid model achieved the lowest MAPE of 1.47%, outperforming even standalone LSTM and BPNN approaches [34]. Its ability to incorporate exogenous and seasonal variables across multiple time resolutions proved essential to its accuracy.

In the context of smart cities, where energy systems must anticipate long-term load dynamics while adapting to policy and climate shifts, this hybrid model exemplifies the value of integrating LSTM’s temporal learning with complementary machine learning layers. Its modularity and performance make it a viable candidate for energy infrastructure planning, peak load mitigation, and grid optimization strategies.

Although hybrid models for energy consumption prediction exhibit promising accuracy, the majority of evaluations are based on historical datasets and conducted under specific regional or contextual conditions. Their performance under extreme data variability, long-term climate trends, or atypical events remains largely unexplored. Furthermore, the reviewed studies utilize diverse datasets and evaluation metrics, hindering the direct comparability across models. This variability restricts the generalization and scalability of these current findings. Recognizing these limitations, future research should prioritize the development of standardized benchmarking frameworks to facilitate more interpretable and robust cross-context model evaluation.

3.4. Health and Public Safety

Health and public safety prediction in smart cities depends on diverse, often incomplete data from sources like IoT sensors and surveillance. Challenges include complex, non-linear, and spatiotemporal patterns in events like disease outbreaks or cyber threats. Traditional models fall short in handling such dynamics, making hybrid deep learning models vital for accurate, real-time crisis management.

Research has primarily focused on epidemic prediction and cybersecurity in the health and public safety domains. Khan et al. [35] proposed a hybrid CNN-LSTM model to predict COVID-19 hotspots in smart cities. The objective of this study was to develop a data-driven prediction model to facilitate the planning and implementation of efficient lockdown strategies. In this model, a CNN was employed to extract multiscale temporal features from the input time-series data. In contrast, LSTM was used to learn temporal dependencies based on those features, thereby enhancing the overall prediction performance.

The CNN-LSTM model integrates the strengths of both CNNs and LSTM, with a CNN being responsible for extracting spatial features from the data and LSTM handling temporal dependencies. Notably, the model was designed to utilize multiscale temporal features extracted from various CNN layers, enabling it to learn short-, medium-, and long-term dependencies simultaneously. This multiscale approach provides more accurate predictions than models that rely solely on single-scale features.

The performance evaluation demonstrated that the CNN-LSTM model significantly outperformed existing models, including ARIMA, standalone CNN, standalone LSTM, and FFNN. The CNN-LSTM model achieved a mean absolute error (MAE) of 74.27, compared to 120.76 for LSTM and 201.75 for ARIMA [35]. Regarding the root mean square error (RMSE), the model scored 106.90, which was considerably lower than the 179.41 for LSTM and 283.46 for ARIMA [35]. These results indicate that the hybrid model achieved an over 40% improvement in prediction accuracy compared with traditional models, as presented in Figure 24 [35]. This study confirmed that CNN-LSTM can be a powerful tool for managing public health emergencies by supporting proactive and precise containment strategies in smart cities.

Kompunt et al. [36] proposed a hybrid MLP-LSTM model to predict the spread of COVID-19 across 77 provinces in Thailand. Figure 25 represents the overall architecture of the proposed system. The primary goal of this study was to forecast newly confirmed cases, total infections, and death counts that would support efficient disease control and resource allocation planning. In this model, LSTM processed time-series data by learning temporal dependencies. Meanwhile, an MLP handled static data, such as regional attributes and vaccination rates, to improve the model’s predictive capability.

A key advantage of this MLP-LSTM model is its ability to learn dynamic temporal patterns and static features jointly. LSTM effectively handles long-term dependencies in time-series data, and the MLP increases prediction accuracy by incorporating diverse static variables. The model processes dynamic and static data streams separately and then integrate the results to generate the final prediction. The prediction outputs were visualized using a geographic information system (GIS) and made accessible via a web application.

In performance evaluations, the MLP-LSTM model consistently outperformed the standalone LSTM and unoptimized models. For example, when predicting the total infection cases, the LSTM model alone achieved an R² value of 0.9959 and an MAE of 300.97 [36]. In contrast, the MLP-LSTM model improved to 0.9980 and 100.60, respectively, indicating a 66.6% reduction in the error [36]. In death prediction, the LSTM model recorded an R² of 0.9944 and an MAE of 1.42, whereas MLP-LSTM achieved values of 0.9964 and 1.04, representing a 26.8% improvement in MAE [36]. These results validate the superior performance of the hybrid model in integrating heterogeneous data sources for pandemic management.

Sah et al. [37] developed a stacked hybrid LSTM-GRU model combining deep learning with statistical forecasting methods such as Prophet and ARIMA. Their study aimed to provide accurate predictions of COVID-19 confirmed and active cases across India to support healthcare resource planning. The Prophet and ARIMA models are effective at modeling linear trends and seasonality but are limited in handling non-linear, long-term dependencies. To overcome these shortcomings, a stacked LSTM-GRU architecture was implemented using an Adam optimizer and ReLU activation functions, enhancing its ability to learn complex time-series patterns.

The hybrid model significantly outperformed the individual statistical models. Specifically, in terms of prediction accuracy (R²), the LSTM-GRU model achieved 0.74, whereas ARIMA and Prophet recorded 0.56 and 0.46, respectively [37]. For RMSE, the hybrid model scored 69.92, whereas ARIMA and Prophet scored 1260 and 568.58, respectively [37]. These results demonstrate that deep learning-based hybrid models are not only applicable to epidemic forecasting but can also be extended to other smart city domains such as traffic, energy, and environmental monitoring.

In the field of cybersecurity, Razib et al. [38] proposed a hybrid DNN-LSTM model to detect IoT-based security threats in smart environments. This model, as shown in Figure 26, was integrated with a Software-Defined Networking (SDN) framework to detect and mitigate various cyberattacks, such as DDoS, brute force, and infiltration threats. The DNN component learns non-linear patterns from the network traffic, whereas the LSTM component captures temporal dependencies. The fusion of these components allows the system to detect suspicious activities with greater precision in real time.

The DNN-LSTM model can analyze heterogeneous IoT data and monitor network traffic in real time. DNNs are well suited to learning complex patterns in high-dimensional data, whereas LSTM models have long- and short-term temporal dependencies. When integrated with SDN, the model enhances the network adaptability and scalability by dynamically controlling the data flow between IoT devices.

The performance evaluation showed that DNN-LSTM achieved 99.55% accuracy, significantly outperforming DNN-GRU (98.67%) and BLSTM (98.9%) [38]. Furthermore, the model recorded a precision of 99.36%, recall of 99.44%, and F1-score of 99.42%, all of which were 1–2% higher than those of the alternative models [38]. These findings demonstrate that DNN-LSTM is highly effective for the real-time detection and mitigation of cyber threats in IoT-driven smart city infrastructures.

Mohbey et al. [39] proposed a CNN-LSTM hybrid model to analyze sentiments related to monkeypox on social media. This study utilized 61,379 tweets collected between May and June 2022 to classify public sentiment into positive, negative, and neutral categories. The CNN component captured local text features, such as word position and contextual relationships, whereas the LSTM component modeled the sequential emotional flow within the tweets. The preprocessing steps included stop-word removal, punctuation, URL cleaning, text normalization, and stemming.

The performance metrics demonstrated that the CNN-LSTM model outperformed other classification algorithms, including standalone CNN, LSTM, SVM, Random Forest, and Decision Tree approaches. It achieved an accuracy of 94%, which was 6% higher than that of the LSTM model (88.1%) [39]. The ability of the hybrid architecture to capture long-term dependencies in lengthy texts proved advantageous for sentiment classification. This study validated CNN-LSTM as an effective tool for analyzing public sentiment during infectious disease outbreaks and guiding public health communication strategies in smart cities. Future studies should consider expanding the model to perform topic classification and sentiment analysis simultaneously.

Hassan et al. [40] proposed a hybrid deep learning model, CNN/weight-dropped LSTM (WDLSTM), for efficient intrusion detection in big data environments by leveraging the strengths of both convolutional neural networks and regularized recurrent units. The architecture includes two 1D convolutional layers for extracting spatial features, a max-pooling layer for dimensionality reduction, and a WDLSTM layer that maintains long-term temporal dependencies while preventing overfitting through drop-connect regularization.

The model was evaluated on the UNSW-NB15 dataset, comprising over 2.2 million instances across 10 attack categories. It achieved 97.17% accuracy in binary classification and 98.43% in multi-class classification, significantly outperforming baseline models such as the TSDL (two-stage deep learning) model, which showed only 89.13% accuracy on the same dataset [40]. Moreover, the CNN–WDLSTM model demonstrated a lower average execution time per instance (0.002383 ms) compared to TSDL (0.003372 ms), making it more suitable for real-time intrusion detection [40]. A Z-test confirmed the statistical significance of this performance improvement at a 95% confidence level.

While these studies demonstrate strong performance in epidemic prediction and cybersecurity, the generalizability of hybrid models to other smart city domains—characterized by different data structures, infrastructure configurations, and resource limitations—remains an open challenge. The cross-domain applicability of hybrid LSTM architectures has not been extensively examined, limiting insights into their transferability across diverse urban contexts. Future research that adapts these modeling strategies to structurally analogous domains may offer meaningful contributions to expanding the utility of hybrid approaches in smart city analytics.

3.5. Urban Infrastructure Monitoring

This section examines predictive models for key urban infrastructure types including seaports and maritime facilities, residential housing markets, coastal and tidal infrastructure, urban water supply systems, and structural health monitoring of bridges and reinforced concrete buildings.

Urban infrastructure monitoring in smart cities faces challenges due to heterogeneous, high-dimensional, and often incomplete sensor data. Prediction models must account for spatial–temporal dependencies and evolving infrastructure conditions. Traditional models struggle with adaptability to dynamic urban environments and real-time data processing. Advanced hybrid deep learning models are needed to enhance accuracy, resilience, and scalability in forecasting structural health and resilience.

Research on urban infrastructure monitoring has addressed logistics and real-estate forecasting. Cuong et al. [41] proposed a hybrid model combining Long Short-Term Memory (LSTM) and Random Forest (RF) to enhance the prediction of container throughput at the Busan Port, as shown in Figure 27. This model focuses on increasing forecast accuracy by addressing non-linear patterns and external disruptions in port operations. In the hybrid architecture, LSTM manages time-dependent patterns in the data, whereas RF processes the residuals produced by LSTM to correct prediction errors.

The LSTM-RF model was designed to integrate the complementary strengths of the two algorithms: LSTM captured both long- and short-term dependencies to predict trends and patterns, whereas RF refined the remaining errors to improve the precision of the final predictions. Compared to the individual models, this hybrid approach demonstrated a stronger and more stable performance under variable conditions.

Performance evaluations demonstrated that the LSTM-RF model outperformed both the standalone LSTM and RF models. In the long-term prediction scenarios, LSTM-RF achieved a root mean square error (RMSE) of 0.019, compared with 0.1664 for LSTM and 0.0966 for RF [41]. Similarly, in terms of the mean absolute error (MAE), LSTM-RF achieved a value of 0.0004, which was significantly lower than 0.0277 (LSTM) and 0.0093 (RF) [41]. These findings highlight the ability of the hybrid model to handle non-linearity and external variability more effectively than traditional models.

Almaliki and Khattak [42] introduced a hybrid Temporal Convolutional Network (TCN) and LSTM model for tidal-level forecasting in coastal and port infrastructure. Their study aimed to construct a reliable forecasting framework capable of predicting both short-term (T+5 and T+10 days) and long-term (T+30 and T+60 days) tidal fluctuations. The proposed model was designed to combine TCN’s ability to detect time-series patterns with LSTM’s capacity to capture long-term dependencies, enabling the accurate modeling of both localized and periodic changes.

The study used high-precision hourly tidal data collected from Ras Tanura, Saudi Arabia, between 2016 and 2022 (±0.005 m accuracy). A sliding window of 30 d was applied to generate input sequences, and hyperparameters were optimized using the Covariance Matrix Adaptation Evolution Strategy (CMA-ES), thereby reducing human bias and improving learning efficiency.

In the performance assessments, the TCN-LSTM hybrid outperformed the standalone LSTM, GRU, and CNN models across all forecasting horizons. For example, in the T+5 d prediction, TCN-LSTM achieved an MAE of 0.073 and RMSE of 0.08 m, with a 25.5% lower MAE and 32.2% lower RMSE than LSTM [42]. For the T+30 and T+60 forecasts, the hybrid achieved MAEs of 0.050 and 0.054 and MAPEs of 5.39% and 4.93%, respectively [42]. The CNN exhibited the poorest performance, with MAE and MAPE values up to 52% higher than those of the hybrid model [42].

Uncertainty analysis further demonstrated the reliability of the hybrid model. For the T+60 predictions, TCN-LSTM recorded the lowest average uncertainty (0.301) and uncertainty ratio (0.376), substantially outperforming the GRU (0.545), LSTM (0.533), and the CNN (0.604) [42]. The researchers suggested expanding the model to include meteorological factors (e.g., storm surges) for integrated coastal forecasting and highlighted its applicability to urban planning, tidal energy management, and port design.

Temür et al. [43] used monthly housing sales data from Turkey (January 2008–April 2018) to propose a hybrid ARIMA-LSTM model. In the model, ARIMA models the linear components, whereas LSTM processes the residuals to capture non-linear relationships. This two-step process enables the independent modeling of both linear and non-linear patterns, with the final outputs generated through their combination, as illustrated in Figure 28.

The hybrid model outperformed the standalone ARIMA and LSTM models. It achieved a mean absolute percentage error (MAPE) of 0.072, compared with 0.121 for ARIMA and 0.15 LSTM, corresponding to error reductions of approximately 40% and 52%, respectively [43]. For the mean squared error (MSE), the hybrid model scored 175.6, compared to 280.4 (ARIMA) and 473.3 (LSTM), reflecting approximately 37% and 63% lower errors, respectively [43]. These error values are summarized in

Table 2. Comparison of error values by predictive learning model [43].

Comparison of All Models
Model	MAE	MSE	RMSE	MAPE
ARIMA (0,1,2)	12.801	280.400	16.745	0.121
LSTM (1000 epochs)	17.853	473.346	21.757	0.150
Hybrid (LSTM 1500 epochs–ARIMA (1,1,1))	8.847	175.611	13.252	0.072

.

Pu et al. [44] proposed a CNN-LSTM hybrid model for short-term urban water demand forecasting. The goal was to capture temporal and spatial patterns from complex water demand data influenced by variables such as temperature, rainfall, and hourly fluctuations. Accurate forecasting supports efficient water supply management and cost savings.

Performance comparisons showed that CNN-LSTM outperformed the ANN, Conv1D, GRUN, and LSTM models. For the 15 min forecasts, CNN-LSTM achieved a MAPE of 1.25% [44]. This was significantly lower than that of ANN (2.59%), Conv1D (3.05%), GRUN (2.39%), and LSTM (2.32%) [44]. In one-hour forecasts, the model achieved a MAPE of 2.97%, again outperforming ANN (4.08%), Conv1D (4.67%), GRUN (4.31%), and LSTM (4.49%) [44]. The RMSE values for CNN-LSTM were 79.23 m³/h (15 min) and 201.13 m³/h (1 h), reflecting substantial improvements in prediction accuracy [44].

Das and Guchhait [45] proposed a parallel hybrid GRU–LSTM(P) deep learning model for structural damage detection using time-series vibration data. Their objective was to identify and quantify the location and severity of structural damage based on experimentally obtained acceleration response data. The model was validated using real-world data from the Z24 bridge and achieved nearly 100% accuracy across 17 damage scenarios.

In the initial simulation phase, a two-span continuous beam was modeled as a simplified steel girder bridge, and 60 damage cases were created by simulating 10%, 30%, and 50% thickness reductions. Gaussian noise was added to simulate the real sensor environments. The data were input in raw form without preprocessing, thereby enhancing the practicality of the model.

Model comparisons showed that under high-noise conditions (10%), the standalone GRU achieved 91.77% accuracy and a 91.76 F1-score, and LSTM achieved a 93.04% accuracy and a 93.03 F1-score [45]. In contrast, the GRU–LSTM(P) model achieved superior results of 93.32% accuracy and a 93.30 F1-score [45]. The proposed model converged quickly (within 20–30 epochs) and was computationally efficient for real-time inference on platforms such as Google Colab.

The validation of the Z24 bridge confirmed the classification accuracy of the model across all damage scenarios. Compared to the CNN-LSTM model proposed by Yessoufou and Zhu [46], which achieved 86% accuracy, the GRU–LSTM(P) model demonstrated superior performance. The T-SNE visualizations revealed clear damage clusters, and the model accurately localized and quantified the damage using only raw acceleration data.

The study demonstrated the feasibility of real-time structural health monitoring using a high-performance deep learning model capable of processing raw vibration signals without preprocessing [46]. The model was resilient to noise and variations in the damage type, making it suitable for practical deployment in real-world infrastructure systems.

In 2022, Parida et al. [47] developed a hybrid CNN–LSTM deep learning model to monitor bond strength degradation in reinforced concrete structures using time-series data from both vibration and ultrasonic wave signals. Their objective was to evaluate bond quality degradation in rebar–concrete interfaces, a critical aspect of structural durability, through continuous, non-destructive monitoring.

The study employed experimental specimens with varying levels of artificially induced bond damage using accelerated corrosion, and data were collected using embedded sensors during cyclic loading.

Compared to standalone CNN and LSTM models, the hybrid CNN–LSTM achieved a higher predictive performance. Specifically, the CNN–LSTM model recorded an RMSD of 5.3% at a 10 mm displacement level, outperforming the standalone CNN (RMSD = 10.1%) and LSTM (RMSD = 7.0%) architectures [47]. These results underline the hybrid model’s effectiveness in both regression and classification tasks.

A key strength of the study was its use of both ultrasonic and vibration data, enabling the model to fuse complementary signal characteristics. This multimodal input significantly improved the model’s ability to detect early-stage degradation. The CNN–LSTM model also required fewer training epochs to converge and showed robustness under sensor noise and real-time fluctuations.

This work demonstrated the practical applicability of hybrid deep learning in structural health monitoring for concrete infrastructure. By combining spatial and temporal learning capabilities and leveraging multiple sensor modalities, the proposed approach offered a high-accuracy, low-latency solution for early warning systems in civil structures.

Although these studies demonstrate high predictive performance, most have not been tested under the practical constraints of real-world infrastructure, such as limited sensor coverage, data inconsistencies, and requirements for model interpretability. These implementation challenges, including the explainability of predictions and integration into operational decision systems, remain underexplored in the current literature. Future research should address these factors to support real-time adoption in smart city infrastructure monitoring.

3.6. Urban Waste Management

Predicting waste management in smart cities is challenging due to the dynamic influence of socioeconomic factors and seasonal trends. The data is often non-linear, multi-source, and exhibits temporal dependencies. Traditional models fail to capture complex relationships between input variables and waste output. Hybrid deep learning models have proven effective in improving prediction accuracy for sustainable waste management.

Hybrid LSTM models have also been applied in urban waste management, particularly to forecast waste generation, predict waste bin fill-levels, and develop smart waste classification systems.

Lin et al. [48] proposed a hybrid attention–LSTM-CNN model to accurately predict municipal solid waste (MSW) generation in smart cities. The primary goal of their study was to develop a forecasting system that integrated time-series data with socioeconomic variables (e.g., average household size, budget revenues, and public investment) to support waste policy and management decisions. The model leveraged the complementary strengths of the attention mechanism, 1D-CNN, and Bidirectional LSTM (Bi-LSTM) to enhance prediction accuracy, as shown in Figure 29.

The key feature of the Attention–LSTM-CNN model is its ability to process spatial and temporal patterns simultaneously. The attention mechanism selects important features from the input, which are then passed on to LSTM to learn the temporal dependencies. LSTM handles both long- and short-term data dependencies, whereas the CNN extracts spatial features, ultimately improving the prediction precision. This structure was specifically designed to effectively determine the non-linear relationships between complex socioeconomic variables and waste generation.

The experimental results demonstrated that the Attention–LSTM-CNN model outperformed the other hybrid and individual models. Based on the R² (coefficient of determination), the CNN-LSTM–Attention configuration achieved a value of 95.31%, compared with 89.45% for the Attention–CNN-LSTM (A-C-L) structure and 90.77% for the CNN–Attention–LSTM (C-A-L) structure [48]. The model also recorded the lowest mean squared error (MSE) of 0.032, confirming its superior predictive performance.

In conclusion, the Attention–LSTM-CNN model provided a promising solution for supporting data-driven decision-making in smart city waste management systems. Improving the accuracy of urban waste forecasting will contribute to more effective resource allocation, policy development, and the promotion of sustainable urban environments.

Lilhore et al. [49] introduced a smart waste classification framework that combined a CNN-LSTM hybrid model with transfer learning (TL) to address sustainable waste management in urban environments. This study aimed to automatically classify waste into organic and recyclable categories, thereby reducing the limitations of manual sorting, data imbalance, and model overfitting.

The proposed model was designed such that the CNN extracted spatial features from waste images, whereas the LSTM analyzed the temporal sequence relationships within those features. Transfer learning was applied using pre-trained weights from ImageNet to reduce training time and enhance generalization. The model was trained on the TrashNet image dataset from Kaggle, which included 27,025 images—17,005 of organic waste and 10,020 of recyclable waste.

Performance comparisons against models such as VGG-16, ResNet-34, ResNet-50, and AlexNet were conducted under two conditions (80:10:10 and 70:15:15 train/validation/test splits) and across 50 and 100 epochs. In all cases, the CNN-LSTM+TL hybrid model achieved the best performance, with 95.7% accuracy, 95.4% precision, 94.3% recall, and a 93.1% F1-score—improvements of 4–10% over conventional models [49]. The model also achieved the lowest error rate (0.0166%) and shortest training time (7586 s) in the 50-epoch setting [49].

To further enhance model performance, the study incorporated advanced data augmentation techniques, including background enhancement (BackMod), image rotation and flipping, and noise injection. Optimization was conducted using the Adaptive Moment Estimator (AME) algorithm, which contributed to faster convergence and effective loss minimization.

Lilhore et al. [49] demonstrates the potential for deploying low-cost, high-accuracy, automated waste classification systems in smart cities. Their study makes important contributions to public sanitation, resource recovery, and environmental protection. Future work will explore the multi-class classification of waste images and the development of predictive models for waste generation using socioeconomic and service expenditure variables.

Fatovatikhah et al. [50] proposed a hybrid LSTM-SVM model to predict waste surges during flooding events. Their study focused on providing data for proactive responses and resource allocation by accurately forecasting various types of waste generated by floods.

The dataset was sourced from the U.S. Environmental Protection Agency’s report “Advancing Sustainable Materials Management: Facts and Figures 2015,” covering MSW generation and recycling data across nine time points between 1960 and 2015. To enhance the data volume, linear interpolation was applied between years, resulting in 56 data points that were converted to daily frequencies for model training.

In the hybrid architecture, the LSTM component learned sequential patterns, whereas the SVM was responsible for capturing non-linear relationships through classification or regression. The experiments compared the performances of standalone LSTM and hybrid LSTM-SVM models across different waste categories. The model was trained using the Adam optimizer (learning rate = 0.005) for up to 250 epochs.

Performance evaluation using RMSE and MAPE metrics showed that the LSTM-SVM hybrid model outperformed LSTM alone across all categories. On average, LSTM-SVM achieved an RMSE of 4.48 and a MAPE of 0.26, compared to 5.79 and 0.74 for LSTM, respectively [50]. For instance, in the Paper and Paperboard category, the RMSE decreased from 32.06 (LSTM) to 16.76 (LSTM-SVM), and in the Textile category, the MAPE dropped from 4.91 to 0.59 [50]. These results illustrate the model’s ability to capture complex multivariate time-series patterns more effectively.

In addition to predictive modeling, the study proposes a waste trend analysis model based on the recycling potential across the U.S. regions (Northeast, South, Midwest, and West). An LSTM-based regression model was developed to rank regions according to their recycling capacities, offering practical insights into waste separation policies and regional resource recovery forecasting.

In 2023, Xie et al. [51] examined a hybrid modeling approach that integrated Temporal Convolutional Networks (TCNs) with Long Short-Term Memory (LSTM) networks to enhance the prediction accuracy of total nitrogen (TN) levels in wastewater treatment plant (WWTP) effluent, particularly in response to increasingly stringent discharge regulations. The model was developed using real-time operational data from a WWTP in Jiangsu Province, China.

By combining the strengths of both architectures, the TCN–LSTM model effectively captures temporal dynamics. The TCN component allows for modeling long-range dependencies with flexible receptive fields, while the LSTM layer further refines long-term temporal patterns, helping to preserve important features and mitigate overfitting, which often limits the performance of standalone deep learning models. This architecture proved resilient to large fluctuations in influent characteristics and the complexity of WWTP operations.

Evaluation using test data showed that the hybrid model outperformed standalone TCN and LSTM models, delivering a 33.1% improvement in real-time hourly prediction accuracy [51]. Compared to a traditional feedforward neural network (FFNN), the TCN–LSTM model demonstrated a 63.6% increase in predictive performance [51]. Additionally, it provided highly reliable early warnings of TN effluent concentrations up to eight hours in advance, with over 10% better reliability than baseline models.

Given its adaptability to varying influent and operational conditions, the proposed hybrid model holds promise for broader application across diverse WWTPs.

While these promising studies are based on controlled datasets, they do not address geographic variation, sensor infrastructure disparities, or real-time data constraints, especially in developing or disaster-prone areas. These limitations may affect the scalability and reliability of these models in real-world waste management systems. Further research is needed to assess their performance under varying conditions and capacities.

In conclusion, the hybrid LSTM-SVM model demonstrated robust and accurate forecasting capabilities for flood-induced waste management. The study offers a viable foundation for future applications in disaster preparedness and sustainable waste resource planning.

These developments have expanded the application scope of hybrid LSTM models and improved their performance across various domains.

3.7. Comparative Study

Comparisons and summaries regarding the studies in the six different domains are presented in

Table 3. Comparative summary of hybrid LSTM models across six application domains.

Domain	Study/ Authors	All Models Used	Best Performing Model	Evaluation Metrics	Best Reported Performance	Data Source/Notes
Traffic management	Chahal et al. [11]	Bi-LSTM, SARIMA, BPNN	SARIMA–Bi-LSTM combination	MAE, MSE, RMSE, MAPE	MAE: 0.499 (↓29%), RMSE: 0.58	CityPulse EU FP7, vehicle speed/volume
	Chahal et al. [12]	Bi-LSTM, ARIMA, Aquila Optimizer	ARIMA-Bi-LSTM with Aquila Optimizer	MAE, MSE, MAPE	MAE: 3.18 (↓32%), MAPE: 0.21	CityPulse EU FP7, 3 cities’ traffic data
	Culita et al. [13]	LSTM, ARMA	ARMA–LSTM hybrid	RMSE, MAE	RMSE ↓8.2% (e.g., 13.22 → 12.14)	IoT traffic data, Bucharest (Romania)
	Zafar et al. [14]	LSTM, GRU	LSTM-GRU hybrid	RMSE, MAE, MAPE	MAE ↓10–15%	Islamabad traffic + weather + maps
	Sengupta et al. [15]	LSTM, HMM	C-Hybrid (HMM + LSTM)	RMSE	RMSE: 0.4203	California PeMS traffic data
	Chaoura et al. [16]	LSTM, PSO	LSTM optimized by PSO	RMSE, MAE	MAE: 0.0898 RMSE: 0.15258	Nanjing intersection data (48,120 records)
	Myhrmann & Mabit [17]	LSTM, MDN	LSTM–Mixture Density Network	MSE	MSE: 0.102	Copenhagen bicycle data (64,664 h)
	Miao & Liao [18]	LSTM, CNN, PSO	LSTM-CNN-PSO	MSE, RMSE, MAE,	RMSE: 0.03 MSE↓ 15%	City traffic data
Air quality and environmental monitoring	Kataria & Puri [19]	LSTM, CNN	CNN-LSTM hybrid	RMSE, MAE	Beijing RMSE ↓3.5%, MAE ↓6.3%	Air quality data (Beijing, Chennai)
	Kim et al. [20]	LSTM, SARIMAX	SARIMAX-LSTM stacked ensemble	MAE, RMSE, SMAPE	MAE ↓15.8%, RMSE↓ 9.7%	Solar power output (Incheon, Busan, Yeongam)
	Mukhtar et al. [21]	LSTM, CNN, ANN	CNN-LSTM-ANN hybrid	MAE, RMSE, MAPE	MAPE↓ 9.43%	Solar radiation, 10 African countries
	Zaini et al. [22]	LSTM, EEMD, PSO, SSA	EEMD-SSA-LSTM with neighbor data	RMSE, MAE, MAPE, R2	MAE↓ 9.31% RMSE↓ 2.65%	PM2.5 and meteorology (Malaysia, 2019)
	Kaveh et al. [23]	LSTM, Orchard Algorithm (OA), BOA	OA-LSTM with BOA-based feature selection	RMSE, R2, SD Error	R2: 0.65 → 0.88	Tehran pollution, satellite + weather
	Ratković et al. [24]	LSTM, GA, Bagging, SVM	Ensemble LSTM optimized with GA	MSE, MAE, MAPE, R2	MAPE↓ 66.0%	Niksic air quality, hourly CAQI data
	Zeng et al. [25]	NLSTM, ESWT	ESWT-NLSTM combination	MAE, RMSE, MAPE, R2	RMSE ↓ up to 69%, MAE ↓62.8%	Beijing PM2.5, ESWT signal decomposition
	Dalal et al. [26]	LSTM, PSO, RNN	PSO–LSTM–RNN	MAE, RMSE, R2	RMSE: 0.0184	City pollutants data
Energy management	Arslan [27]	Bi-LSTM, Prophet	Prophet + Bi-LSTM (stacked)	RMSE	RMSE ↓43% (Canada), ↓45% (France)	Smart city energy (Canada and France)
	Ali et al. [28]	LSTM, CNN, ConvLSTM	ConvLSTM (spatiotemporal forecasting)	MAE, MAPE, RMSE	MAE: 3.69–4.65, MAPE ↓18.5~18.8%	Smart home energy consumption
	Alhussein et al. [29]	LSTM, CNN	CNN-LSTM hybrid	MAPE (1-step and multi-step)	MAPE ↓6%	Australia SGSC smart grid data
	Karri et al. [30]	LSTM, MLP	LSTM-MLP hybrid for IoT data	Accuracy, precision, recall, F1-score	Accuracy↑ 6%	Kaggle IoT data on waste-to-energy
	Kaya et al. [31]	LSTM, CNN	CNN-LSTM for energy source forecasting	RMSE, MAE, R2	R2: 0.99	Turkey energy mix (2018–2023, EXIST)
	Du [32]	LSTM, SVR, RF	RF-stacked SVR-LSTM	Accuracy, MSE, R2	Accuracy 90%	Building energy data (539 cases)
	Sivarajan & Jebaseelan [33]	LSTM, CNN, Modified Sea Lion Algorithm (MSLA)	CNN-LSTM + MSLA	MSE, RMSE, MAE, MAPE	RMSE: 0.438 → 0.2	IoT smart grid household appliances
	Jin et al. [34]	LSTM, BPNN	LSTM–BPNN-to-BPNN	MAPE	MAPE: 4.63%→1.47%	Copenhagen bicycle data (64,664 h)
Health and public safety	Khan et al. [35]	LSTM, CNN	CNN-LSTM (COVID-19 hotspot prediction)	MAE, RMSE	MAE: 74.27 (↓40%), RMSE: 106.90	COVID-19 time series (multiscale features)
	Kompunt et al. [36]	LSTM, MLP	MLP-LSTM integrated model	MAE, R2	MAE ↓66.6%,	COVID-19 cases across 77 Thai provinces
	Sah et al. [37]	LSTM, GRU, Prophet, ARIMA	Stacked LSTM-GRU	R2, RMSE	R2: 0.74 RMSE: 69.92	India COVID data (confirmed, active cases)
	Razib et al. [38]	LSTM, DNN	DNN-LSTM for cyberattack detection	Accuracy, precision, recall, F1-score	Accuracy: 99.55%	IoT cyberattack detection in smart city SDN
	Mohbey et al. [39]	LSTM, CNN	CNN-LSTM for social sentiment analysis	Accuracy	Accuracy: 88.1%→ 94%	Monkeypox-related tweets (61,379 samples)
Urban infrastructure monitoring	Cuong et al. [41]	LSTM, Random Forest	LSTM-RF hybrid for port throughput	RMSE, MAE	MAE: 0.004	Busan port container throughput
	Almaliki & Khattak [42]	LSTM, TCN	TCN-LSTM for tidal forecasting	MAE, RMSE, MAPE, uncertainty	MAE ↓ 25.5% RMSE ↓ 32.3%	Ras Tanura (Saudi) tidal levels
	Temür et al. [43]	LSTM, ARIMA	ARIMA-LSTM for real estate	MAE, MSE, RMSE, MAPE	MAPE ↓ 52% MSE ↓ 63%	Turkey housing sales (2008–2018)
	Pu et al. [44]	LSTM, CNN	CNN-LSTM for water demand forecasting	MAPE, RMSE	MAPE ↓40–52%, RMSE ↓60.6%	Urban water usage (15 min and 1 h)
	Das & Guchhait [45]	LSTM, GRU	Parallel GRU–LSTM(P) for bridge damage	Accuracy, F1-score	Accuracy 93.32%, F1 93.3	Z24 bridge damage vibration data
	Parida et al. [47]	LSTM, CNN	Hybrid CNN–LSTM	RMSD	RMSD 5.3% at 10 mm displacement	Embedded sensor data (vibration & ultrasonic)
Urban waste management	Lin et al. [48]	CNN-LSTM-Attention	Attention–LSTM-CNN	R2, MSE	R2: 95.31% MSE: 0.032	Municipal solid waste & socioeconomic data
	Lilhore et al. [49]	CNN, TL, CNN-LSTM	CNN-LSTM+TL	Accuracy, Precision	Accuracy 95.7% Precision 95.4%	TrashNet image
	Fatovatikhah et al. [50]	LSTM, SVM, Hybrid LSTM–SVM	LSTM–SVM	RMSE, MAPE	RMSE 4.48, MAPE 0.26	Effective for forecasting flood-related waste surges; strong multivariate pattern capture
	Xie et al. [51]	TCN, LSTM, TCN-LSTM	TCN-LSTM	Accuracy	Accuracy↑ 33.1%	Robust model for WWTP inflow prediction

Note: ↓ indicates a decrease in performance of the proposed model compared to the conventional model; ↑ indicates an increase; → indicates a change from A to B, meaning the value changed from A to B.

. Based on this study comparison, analysis in relation to the different domains was carried out to identify domain-specific insights including trends, trade-offs, and the strengths/weaknesses of various hybrid LSTM architectures.

In traffic and energy management, models that integrate traditional time-series techniques—such as SARIMA, ARIMA, and Prophet—with LSTM have shown consistent improvements in capturing periodic trends and short-term fluctuations. These combinations leverage the interpretability of statistical models and the non-linear learning power of LSTM. For example, SARIMA–Bi-LSTM and Prophet–LSTM hybrids have demonstrated significant error reductions in vehicular flow and energy demand forecasting. The structure of these models is particularly suitable for domains where historical data patterns repeat with clear seasonal or cyclical behaviors.

Spatial–temporal domains, such as air quality monitoring and water demand prediction, benefit from hybrid architectures that incorporate CNNs alongside LSTM. These CNN-LSTM models effectively extract and preserve spatial correlations before passing sequences to the temporal model, thereby improving the forecasting of pollution levels or environmental variables that vary across both time and space. Furthermore, models that incorporate signal decomposition techniques—such as Ensemble Empirical Mode Decomposition (EEMD) or Extended Stationary Wavelet Transform (ESWT)—combined with LSTM have shown promise in reducing noise and isolating predictive components in environmental data.

In domains involving more heterogeneous and irregular data, such as public health and cyber-physical safety, ensemble models and parallel LSTM architectures have proven to be especially effective. For instance, GRU-LSTM or DNN-LSTM hybrids have achieved high levels of accuracy in tasks like COVID-19 hotspot prediction and IoT-based intrusion detection. These models improve generalization and robustness by mitigating individual model biases, though they often increase computational and interpretive complexity. The trade-off between performance and transparency becomes especially critical in domains where decision-making carries ethical or operational implications.

Urban infrastructure monitoring presents a unique challenge due to the need for real-time feedback and integration with physical sensor systems. In this context, hybrid LSTM models that include uncertainty quantification techniques or are optimized through evolutionary algorithms (e.g., GRU–LSTM with metaheuristic tuning) are well suited to handling the dynamic and localized nature of infrastructure data. However, their deployment often requires domain-specific customization and calibration with real-world systems. Across all domains, it is evident that the most effective hybrid configurations are those that align not only with data characteristics but also with the operational context in which predictions are applied.

4. Future Perspectives

This section presents a forward-looking perspective on hybrid LSTM models, focusing on technological advancement trajectories, application domain expansion, market growth projections, and implementation challenges.

The technological evolution of hybrid LSTM models is expected to advance primarily through architectural sophistication and improvements in computational efficiency. New hybrid structures incorporating transformers are likely to emerge, significantly enhancing long-term dependency learning and parallel-processing capabilities. Recent studies have demonstrated that adding self-attention layers to LSTM–CNN hybrids can further improve multi-step forecasting accuracy on urban traffic and energy datasets [52,53]. Recent research has identified limitations in RNN-based attention mechanisms for seasonal modeling, prompting the development of novel architectural solutions. A bibliometric survey of AI applications in smart cities, covering over 4400 publications between 2010 and 2023, shows exponential growth in neural network-based approaches, particularly in the transportation and energy domains [54,55,56,57,58,59]. While these bibliometric trends indicate rising academic interest, they do not yet confirm large-scale industrial adoption.

In particular, promising advancements are anticipated through CNN integration. Combinations with extended causal convolutions and Temporal Convolution Networks (TCNs) that effectively capture long-range dependencies in the temporal dimension show significant potential for time-series forecasting. Integration possibilities with quantum computing merit attention, whereas convergence with federated learning is expected to enhance model performance while maintaining data privacy protection.

The predictive capabilities expand beyond univariate time-series forecasting to probabilistic and multivariate predictions. Complex forecasting scenarios, such as those in retail, require sophisticated hybrid LSTM architectures capable of modeling interdependent relationships between different variables. Enhanced modeling capabilities for big data environments with multiple seasonality patterns are likely to gain prominence.

Innovative hybrid models that integrate global and local variables are designed to reflect specific time-series characteristics by combining hierarchical models with ensemble techniques. These advancements will enable more accurate predictions and decision-making across diverse smart city applications.

The utilization scope of hybrid LSTM models is projected to extend beyond the current transportation, environmental, and energy sectors to broader urban problem-solving domains. New applications are expected in disaster prediction and response, urban crime prevention, and social infrastructure life cycle forecasting. This technology can be applied to behavioral pattern analysis and demand forecasting for personalized citizen services.

Convergence with existing fields will accelerate. For instance, integration with digital twins could help develop tools that support optimal decision-making through various scenario simulations in virtual urban environments. Metaverse applications represent another emerging possibility.

The global smart city market was valued at USD 820.7 billion in 2023 and is projected to grow at a compound annual growth rate (CAGR) of 22.1% from 2024 to 2030 [60]. Other forecasts predict an even faster expansion—from USD 877.6 billion in 2024 to USD 3.76 trillion by 2030—reflecting a CAGR of approximately 29.4% [61,62] and indicating strong demand for AI-based technologies such as hybrid LSTM models. Patent landscape analyses also reveal a cumulative increase of 22,655 smart-city-related patents over the past 16 years, with annual filings peaking in recent years—underscoring heightened innovation and industry interest in this sector [63]. Amid this growth, smart city technologies increasingly rely on hybrid LSTM architectures to enhance real-time data processing and prediction accuracy. The Asia–Pacific region, driven by rapid urbanization and active government-led smart city initiatives, is expected to see explosive growth in demand for such technologies.

By industry, the smart mobility and smart energy sectors forecast the highest demand, reflecting the prioritization of urban traffic solutions and energy efficiency improvements. The smart healthcare and smart building sectors will also experience rapid growth as IoT and AI technology adoption expands. Both government and private sector investments are expected to increase continuously, particularly aligned with ESG (Environmental, Social, Governance) management strengthening and sustainable urban operations.

Hybrid LSTM technologies provide a significant competitive advantage in terms of market growth. Real-time data processing and predictive capabilities are essential for maximizing the operational efficiency of smart cities and have become fundamental elements in various applications, including energy management, waste prediction, and traffic flow control. These technological advancements will further refine the smart city ecosystem and contribute to addressing the diverse challenges faced by global cities.

Several challenges must be addressed for the widespread implementation of H-LSTM models. Interpretability issues arising from model complexity require further improvement. Integrating explainable AI (XAI) technologies to create transparency in decision-making processes is crucial. Data quality management and standardization are urgent priorities.

Additionally, the development of unified meta-analytic frameworks and standardized benchmarking systems represents a critical future research priority to enhance the comparative validity of hybrid LSTM performance across different urban contexts. Such frameworks would address current limitations in generalizing model performance given the heterogeneous nature of datasets, temporal resolutions, and evaluation metrics used across various cities.

In response to these challenges, international standardization organizations have established data quality criteria and model evaluation metrics. Government initiatives are emerging to develop data governance frameworks and refine relevant regulations.

Hybrid LSTM models are expected to enhance the sustainability of smart cities significantly. They create environmental value through energy efficiency optimization and carbon emissions reduction while realizing social value by improving citizens’ quality of life and urban service accessibility.

From an economic perspective, these models positively impact urban operational cost reductions and the creation of new business models. Predictive maintenance and optimized resource management can extend the lifespan of urban infrastructures and enhance operational efficiency. Public–private sector collaboration will maximize the potential of the hybrid LSTM model and promote sustainable smart city development.

This developmental trajectory faces several critical challenges. Therefore, the increasing computational time and cost requirements of more complex models must be addressed. Data quality issues when processing large volumes and difficulties in explaining model prediction processes also require resolution. Continuous research and innovation to overcome these technical limitations are necessary to provide opportunities to further enhance the practicality and reliability of hybrid LSTM models.

5. Conclusions

Hybrid LSTM models serve as powerful tools for effectively learning and predicting complex patterns in time-series data, thereby demonstrating their successful implementation across diverse smart city domains. These models contribute to urban problem-solving and quality-of-life improvements by delivering superior predictive performance compared with single models in various sectors, including traffic management, energy consumption forecasting, environmental monitoring, and public health and safety.

By combining the strengths of different algorithms, hybrid LSTM models simultaneously address data linearity and non-linearity while providing integrated spatiotemporal analysis. This capability is particularly significant in complex multidimensional data environments such as smart cities. Integrated analytical capacity is expected to play an increasingly critical role in sophisticated smart city decision-making.

Several challenges must be addressed for the successful implementation of these models. These include increased computational resource requirements and training time owing to model complexity, substantial data demands with overfitting risks, and limited model interpretability. In addition, data quality management, standardization, and privacy protection are important considerations.

Hybrid LSTM models have advanced significantly by integrating diverse techniques such as attention mechanisms, transformers, and optimization algorithms. These developments enhance the models’ ability to capture long-term dependencies, spatial–temporal patterns, and complex data features. Techniques like ensemble learning, model compression, and Bi-LSTM further contribute to improved prediction accuracy and efficiency. As a result, hybrid LSTM models have broadened their applicability and demonstrated strong performance in sub-areas of the smart city.

This review has shown that hybrid LSTM models consistently outperform single models in specific smart city applications by leveraging complementary strengths across algorithms. However, their performance is context-dependent, often relying on historical datasets with limited attention to implementation feasibility, real-time constraints, or standard evaluation frameworks. Furthermore, a lack of model interpretability and insufficient validation under extreme or non-stationary conditions remain key methodological limitations across the literature.

Ongoing research and technological development are necessary to overcome these challenges. Examples include model compression and optimization, the implementation of explainable AI technologies, and the establishment of data governance frameworks. When governments, industries, and academia collaborate in these efforts, hybrid LSTM models will further evolve as core smart city technologies and contribute to sustainable urban development.

In conclusion, hybrid LSTM models will play a crucial role in addressing complex smart city challenges through technological advancements and application-domain expansion. They are expected to enhance the efficiency of urban management, deliver improved citizen services, and contribute to sustainable urban development. Concurrently, data-driven decision-making systems are becoming more sophisticated, significantly enhancing the operational efficiency of smart cities.