Enhancing Fuel Cell Hybrid Electric Vehicle Energy Management with Real-Time LSTM Speed Prediction

Abstract

This paper presents an innovative approach to optimize real-time energy management in fuel cell electric vehicles (FCEVs) through an integrated EMS (iEMS) framework based on a nested concept. Central to our method are two LSTM-based speed prediction models, trained and validated on open-source datasets to enhance adaptability and efficiency. The first model, trained on a 27 h real-time database, is embedded within the iEMS for dynamic real-time operation. The second model assesses the impact of incorporating external traffic data on the prediction accuracy, offering a systematic approach to refining speed prediction models. The results demonstrate significant improvements in fuel efficiency and overall performance compared to existing models. This study highlights the promise of data-driven AI models in next-generation FCEV energy management, contributing to smarter and more sustainable mobility solutions.

1. Introduction

The automotive industry operates within a complex landscape shaped by economic, technological, and environmental factors [1,2,3,4,5]. Economic aspects, such as consumer demand and interest rates, influence sales, while technological advancements in safety and fuel efficiency drive innovation. Additionally, environmental concerns, including climate change and air pollution, play a crucial role in shaping regulations and consumer preferences [6,7].

In response, fuel cell hybrid electric vehicles (FCHEVs) emerge as a promising solution, offering zero emissions, a longer driving range compared to battery-electric vehicles, rapid refueling, a reduced dependence on fossil fuels, and greater energy efficiency than internal combustion engine vehicles [8,9,10]. However, integrating fuel cell systems (FCSs) into electric vehicles presents several challenges, such as system control complexity, cost, and long-term durability [11]. Fuel cells gradually lose performance over time due to material degradation, catalyst loss, mechanical wear, and variable operating conditions. This decline reduces efficiency, shortens lifespan, and raises maintenance and replacement costs [12].

To address these challenges, an energy management strategy (EMS) can be implemented to optimize the performance of fuel cell systems, enhancing their competitiveness with other powertrain technologies [13,14,15]. EMS plays a crucial role in managing hydrogen storage, distribution, control, and optimization while also improving durability and reliability, resulting in increased efficiency, reduced energy waste, and lower operating costs for fuel cell vehicles [15,16]. Various EMS frameworks have been developed, ranging from rule-based systems to global optimization algorithms such as Dynamic Programming and the Pontryagin Minimum Principle [17,18,19,20]. Although such approaches provide theoretical optimality, their dependence on prior knowledge of driving cycles limits real-time applicability.

Recent advancements in deep learning have opened new avenues for predictive energy management [21]. It aligns with the growing integration of intelligent transportation systems (ITSs) [22]. For instance, ref. [23,24] explores the use of deep reinforcement learning (DRL) algorithms to develop sophisticated EMS for hybrid electric vehicles (HEVs). Innovative forecasting methods, such as those discussed in [25], highlight the potential of deep learning architectures in capturing complex spatiotemporal traffic patterns.

In our context, the accuracy of speed prediction is crucial, as it directly impacts optimization outcomes. In [26], the authors demonstrate that an optimization algorithm can achieve global optimality with a sliding window speed prediction of 7 s or more. However, accurately predicting the stochastic nature of complex real-time driving conditions remains a significant challenge. Multiple factors influence speed prediction. In [27], the authors classify these factors based on whether they are internal or external and their dynamic or static characteristics (see

Table 1. Influencing factors of speed prediction.

Data Location	Data Group	Single-Vehicle Data Examples
Internal: measurable by the powertrain sensors	Driving behavior (dynamic)	Driver’s behavior (e.g., calm, sportive…)
	Vehicle information (static)	Vehicle weight, maximum speed…
	Vehicle state (dynamic)	Vehicle speed, available fuel, battery SOC…
External: cooperative environment data	Traffic flow state (dynamic)	Front vehicle speed, distance…
	Weather conditions (dynamic)	Fog, rain, snow…
	Road and traffic rules (static)	Slope, speed limits, type of road (e.g., urban)…
	Traffic signals and events (dynamic)	Traffic light state and distance, accidents…

). Each factor influences speed differently depending on the vehicle referential (i.e., single-, double-, or multi-vehicle scenarios). Learning-based techniques, trained on a wide range of data, have shown the best results for predicting the speed of a single vehicle [28,29,30]. In [31,32], the authors demonstrate that the Long Short-Term Memory (LSTM) model achieves the highest accuracy for speed over a prediction horizon of 1 to 10 s. Moreover, the challenge of accurate speed prediction in dynamic traffic conditions is being tackled through innovative models such as the D-LSTM. Ref. [33] presents a model that combines Dynamic Time Warping with LSTM, significantly improving traffic speed predictions, particularly under varying conditions.

In this work, we adopt a hybrid energy storage system (HESS) with a parallel architecture (FCS and battery), incorporating two converters to maximize controllability (see Figure 1) [34,35]. We incorporated an LSTM module for speed prediction to improve our energy estimation. Our work differs in its integration strategy and application scope. First, we introduce a real-time Integrated Energy Management System (iEMS) that incorporates LSTM-based speed predictions within a hybrid rule-based, optimization-based, and learning-based framework. Second, we examine not only the base LSTM model but also evaluate the influence of external data (e.g., traffic signals and leading vehicle behavior) on prediction accuracy. Third, we implement online filtering and validate the predictive results within an iEMS architecture in real-world scenarios. This approach enables smoother control transitions and more efficient fuel usage, as validated through performance comparisons with a reference iEMS model.

The remainder of this paper is organized as follows: Section 2 provides an overview of our iEMS framework, which integrates three key strategies: optimization-based, rule-based, and learning-based, previously detailed in [36]. Section 3 introduces and validates our speed prediction model, a crucial component for improving the adaptability of iEMS to real-time driving conditions. Finally, Section 4 presents the results of integrating the new LSTM speed prediction model into our iEMS and evaluates its impact on system performance.

2. Integrated Energy Management Strategy Model

The proposed system uses a Hybrid Energy Storage System (HESS) composed of a fuel cell system (FCS) and a high-voltage battery connected in a parallel configuration, as shown in Figure 1. Each source is managed by a dedicated DC/DC converter, enabling independent power control. The FCS supplies steady power for long-range driving, while the battery handles transients and peak loads. This setup improves flexibility, supports regenerative braking, and reduces stress on the fuel cell. The dual-converter design allows the EMS to optimize power flow in real time based on predicted speed and state of charge (SOC), making the architecture adaptable to various vehicle types.

The EMS model, illustrated in Figure 2, incorporates three key strategies: an optimal online On-Board Strategy (OBS) for fuel consumption optimization, a Rule-Based Strategy (RBS) for managing electrical energy and power assistance, and a Learning-Based Strategy (LBS) to adapt the iEMS to the vehicle’s real-time environment.

The OBS relies on a dynamic convergence condition (target SOC) over a future horizon. At each iteration of the Pontryagin’s Minimum Principle (PMP), it computes the optimal control sequence to reach the target SOC at t + Hp, where Hp is the prediction horizon. This approach minimizes real-time energy consumption, thereby enhancing the overall performance of the iEMS. As part of the PMP framework, the co-state variable is introduced; it acts as a dynamic weighting factor that reflects the relative cost of using battery energy versus fuel cell energy, and helps guide optimal energy-splitting decisions over time.

The OBS requires two key inputs. The first is the target SOC, calculated by the RBS block, which leverages expert knowledge of vehicle states and driving patterns to formulate a battery usage strategy. It also incorporates fuzzy C-means clustering for Driving Pattern Recognition (PDR). The second input, crucial to this study, is the predicted speed at the fixed horizon, which enables accurate system optimization. This aspect, handled by the LBS block, will be detailed in the next section.

The iEMS model is designed to adapt dynamically to real-time driving conditions, integrating additional constraints from the cooperative driving environment and the challenges posed by the fuel cell system in electric vehicles.

The FCHEV system model considered in this study is described in detail in [36]. To evaluate hydrogen consumption (our cost indicator), we introduced a second-order polynomial interpolation of the FCS fuel power, $P_{fuel}$, as illustrated in Figure 3. This interpolation enables the real-time application of PMP by expressing the cost function (i.e., $P_{fuel}$) in terms of the control variable, namely the FCS net power $P_{fc}$. The FCHEV model is designed for adaptability across a wide range of sizing applications, offering an intelligent trade-off between computational efficiency and modeling accuracy for real-time iEMS implementation.

This iEMS model represents a significant advancement in the development of Energy Management Systems (EMSs), providing a structured and systematic approach to designing complex EMS solutions. Its adaptability to real-time operating conditions, coupled with its emphasis on improving fuel efficiency, makes it a promising framework for future research and development in the field.

The next section explores the development of a speed prediction model based on a Long Short-Term Memory (LSTM) neural network. This model aims to further enhance the iEMS’s ability to adapt to real-time driving conditions, thereby improving the overall performance of the FCHEV. As a key innovation in EMS, the speed prediction model holds great potential for increasing the adaptability and efficiency of electric vehicles.

3. Speed Prediction

A more precise speed prediction model enhances the efficiency of the iEMS by enabling more informed power distribution between the fuel cell and the battery. Accurate forecasts of future speed allow the iEMS to minimize energy losses and improve overall system performance. In our case, as discussed in Section 1, the predicted speed is directly utilized by the OBS (see Figure 2). LSTM networks are particularly well-suited for vehicle speed prediction, as they effectively capture long-term dependencies and temporal dynamics within input sequences.

3.1. Long Short-Term Memory Neural Network

LSTM is a type of recurrent neural network (RNN) architecture specifically designed to address the vanishing gradient problem in traditional RNNs. This issue makes it challenging for standard RNNs to learn from data where outputs depend on long-term dependencies [37,38]. The key difference between LSTMs and conventional RNNs lies in the structure of their hidden layer. In a traditional RNN, the repeating module typically contains a single activation function. In contrast, LSTMs incorporate three distinct gates, each with its own activation function, which interact in a structured manner:

• The input gate determines which values from the input should update the cell state.
• The forget gate decides which information should be discarded from the cell state.
• The output gate controls what portion of the cell state should be passed forward.

These gates regulate information flow using a sigmoid neural net layer and pointwise multiplication operations (see Figure 4 and reference [39] for further details). By enabling selective memory retention and removal, LSTMs are particularly well-suited for tasks requiring the comprehension of long data sequences, such as speed prediction.

3.2. Database Preprocessing

Database preprocessing is a critical step in deep learning model design, ensuring data quality, consistency, and compatibility. It involves data integration, feature extraction, data augmentation, normalization, standardization, and data splitting, all of which significantly impact model performance, generalization, and reliability.

• Datasets

For this study, we use an open access, real-time database that records 20 driving cycles over 20 laps along the same route, with a resolution of 10 Hz. This dataset includes external factors such as the front vehicle’s behavior and traffic light data [40]. We selected this database to demonstrate how integrating external data can enhance speed prediction accuracy.

• Handling Missing Data

Missing data is a common challenge in data analysis and machine learning. Addressing missing values involves identifying them and choosing an appropriate handling strategy, such as deletion, imputation, prediction, or the use of indicator variables. The chosen method depends on the quantity and pattern of missing data, the type of analysis, and the study objectives. In this work, since the number of missing values is negligible compared to the total dataset, we opted for a simple deletion approach.

• Normalization

Before training machine learning models, data normalization is essential to improve model performance, accelerate training, enhance interpretability, and prevent numerical issues due to varying feature scales.

Min–max normalization, also known as feature scaling, is a standard technique that rescales all input features to a range between 0 and 1. This method ensures uniformity across features, contributing to a more stable and efficient learning process.

Min–max normalization can be calculated as

$$X_{norm}={\displaystyle \frac{X-X_{min}}{X_{max}-X_{min}}}$$

(1)

where $X$ is the original value of an input feature, $X_{min}$ is the minimum value of that feature, $X_{max}$ is the maximum value of that feature, and $X_{norm}$ is the normalized value of that feature.

• Features selection

Feature selection is a crucial step in the preprocessing phase of machine learning. It involves identifying and selecting the most relevant and informative features (i.e., variables or attributes) from the dataset to be used in the model. This process plays a key role in model development by enhancing predictive performance, reducing the risk of overfitting, improving training efficiency, and increasing model interpretability.

To achieve these objectives, feature selection methods such as the filter method help reduce the number of input variables by retaining only those most predictive of the target outcome. This refinement ensures that the model focuses on the most meaningful data, leading to better generalization and overall efficiency.

We employ a filter method to select our features, assessing their relevance based on statistical properties such as correlation with the target variable and variance (see Figure 5). Since this dataset corresponds to a specific track, applying this feature selection step to any other real-time database remains appropriate.

Notably, we do not expect the LSTM model trained on the 20-lap dataset to generalize directly about real-time traffic conditions. Instead, our goal is to establish a precise methodology that can be adapted to any real-time traffic database.

Histogram analysis reveals that speed, acceleration, front vehicle distance, and traffic light distance are the most predictive features. Accordingly, these variables are selected as input candidates for the LSTM-based speed prediction model. Each feature will be incorporated individually and in combination to evaluate its respective contributions, with the goal of determining the optimal input configuration for speed prediction.

3.3. LSTM Model Training

• Dataset Splitting

Dividing the dataset into training, validation, and test sets is a crucial step in machine learning to effectively evaluate model performance. The primary objective of this split is to ensure that the model is trained on one subset of the dataset, validated on another, and tested on an independent set. This approach helps to evaluate how well the model generalizes to unseen data while mitigating the risk of overfitting.

• Sliding Window and Prediction Horizon

The concepts of the sliding window and prediction horizon play a fundamental role in training LSTM models, especially for time series data analysis.

The sliding window technique involves segmenting a time series dataset into multiple shorter sequences of fixed length (as illustrated in Figure 6). For instance, if speed measurements are recorded for over one hour, the dataset can be divided into smaller sequences of fixed intervals. This method helps capture temporal dependencies and patterns within the data, which are critical for making accurate predictions.

The prediction horizon refers to the time interval into the future that we aim to predict. In the context of speed prediction, our objective is to forecast vehicle speed at a future point based on historical speed measurements. The prediction horizon is a critical parameter, as it defines the number of time steps the LSTM model must anticipate, directly influencing the complexity of the forecasting task.

By integrating the sliding window technique with the prediction horizon, we generate a sequence of input–output pairs for training and testing our LSTM model. Each input sequence consists of a window of previous feature measurements, while the corresponding output is the predicted speed at a future time step.

After several tests and comparisons, we found for the iEMS framework a 7 s prediction horizon and a 7 s sliding window. This configuration ensures temporal coherence between input sequences and output targets, and it aligns with previous work [32]. This ensures effective training and evaluation of our LSTM model, optimizing its accuracy and reliability for real-time speed prediction.

3.4. LSTM Architecture

The LSTM architecture consists of several key layers, each playing a distinct role in the network:

• Sequence Input Layer: This layer receives sequential data, such as time series measurements, serving as the entry point for the network.
• LSTM Layer: This core layer processes the input sequence, capturing long-term dependencies and temporal patterns critical for accurate predictions.
• Fully Connected Layer: Positioned after the LSTM layer, this densely connected layer transforms the LSTM outputs into a fixed-size feature vector, enabling higher-level pattern recognition.
• Regression Layer: The final layer processes the feature vector to generate a continuous output signal, making it well-suited for predicting real-valued data, such as vehicle speed.

In our case, we opted for a single-layer LSTM architecture. This decision was based on a trade-off between model complexity and real-time applicability. Preliminary testing with deeper stacked LSTM networks did not yield substantial improvements in prediction accuracy, while significantly increasing computational load and training time. Given the relatively compact size and consistency of our dataset, a single-layer structure proved sufficient for capturing temporal dependencies effectively. The defining feature of the LSTM architecture is its ability to capture and predict long-term dependencies within sequential data. This is made possible through a memory cell, which retains information from past inputs, enabling the network to remember relevant patterns and incorporate them into future predictions. The options used to train the models are summarized in

Table 2. Options configuration of the Matlab training algorithm.

Options	Parameters	Description
Training solver	Adam	Optimization algorithm
InitialLearnRate	0.001	Initial learning rate
ValidationData	{Xval; Yval}	Validation data
MaxEpoches	100	Training maximum number of epochs
ValidationFrequency	10	Frequency between validations
ValidationPatience	5	End of the network validation
MiniBatchSize	256	Midi-batch size used
Shuffle	Never	We never shuffle the data because our data are time-dependent
Plots	Training progress	Display the training progression on a plot

. Regarding hyperparameter selection, we conducted a manual grid search using validation RMSE as the selection criterion. The selected configuration provided a favorable balance between prediction accuracy and training efficiency.

• Model validation

During training, validation is a crucial step in LSTM modeling. It involves splitting the dataset into training and validation sets to evaluate the model’s performance on unseen data. This ensures that the model generalizes well and helps detect potential issues such as overfitting or underfitting.

To enhance validation accuracy, we applied k-fold cross-validation, a technique that divides the dataset into k equally sized folds. The model is then trained and evaluated k times, each time using a different combination of folds for training and validation. This approach offers a more robust assessment of the model’s performance.

In the next section, we present results from our speed prediction model and its integration into our iEMS.

3.5. Post-Processing

To ensure accurate real-world applicability, the LSTM-based speed prediction model must be refined prior to its integration into our iEMS framework. This section outlines the necessary post-processing steps to enhance prediction accuracy and system performance.

• Identifying the Issue

Raw LSTM speed predictions often exhibit oscillations compared to actual values (Figure 7). These fluctuations can significantly impact acceleration calculations, which are crucial for FCHEV control and health management.

• Enhancing Integrated Energy Management with Online Exponential Filtering

To mitigate these oscillations, we incorporate an online exponential filter into our iEMS framework. This filter smooths raw predictions using the following formula:

$$V_{filtered}\left(t\right)= \alpha \cdot V_{pred}\left(t\right)+ \left(1 - \alpha \right)\cdot V_{filtered}(t - Ts)$$

(2)

Here, $V_{filtered}$ represents the filtered speed prediction, $V_{pred}$ is the LSTM raw predicted speed, $\alpha$ is the smoothing filter factor, and Ts is the speed prediction vector sample time. α is set to 0.02 through tuning, and the filter’s last-known speed initializes V_filtered.

This filtering process reduces noise, ensuring smoother and more reliable predictions, which ultimately enhances the efficiency of iEMS decision-making. Results showed that applying the filter reduced the RMSE from 0.135 to 0.124, confirming that the filter not only smooths oscillations but also improves prediction accuracy. It proves to be a highly effective solution for smoothing LSTM speed predictions, offering several key advantages:

• Preserves Data Integrity: Unlike traditional filtering techniques, it reduces oscillations without introducing undesirable signal shifts.
• Computational Efficiency: The lightweight nature of the filter makes it well-suited for real-time iEMS applications, ensuring fast processing without excessive computational overhead.
• Improved Accuracy: By refining speed predictions, the filter enhances acceleration calculations, directly contributing to better fuel efficiency and power management within the FCHEV system.

With the successful implementation of this exponential filter, our LSTM speed prediction model is now fully integrated into the iEMS framework. This integration results in a more stable and reliable EMS, ultimately optimizing FCHEV performance in real-world driving conditions.

4. Results and Discussion

To validate our innovative iEMS, we conducted a rigorous assessment using a cycle test partition from our original open-source database. Our evaluation followed a two-step process:

-

Speed Prediction Model Assessment

• We first examined the impact of internal and external inputs on the LSTM speed prediction model.
• Performance was assessed across four test cycles, highlighting improvements in accuracy with additional input features.

-

Comparative Analysis of the Real-Time iEMS

• We systematically compared our newly integrated real-time iEMS with the reference iEMS from our previous work [37].
• A key part of this evaluation included investigating the influence of initial state-of-charge (SOC) on energy management.
• Two initial SOC conditions (35% and 65%) were analyzed for each test cycle, providing insights into how different SOC levels affect energy distribution strategies.

This comprehensive assessment provides a data-driven validation of our iEMS framework, demonstrating its ability to optimize fuel cell hybrid electric vehicle (FCHEV) performance under real-world conditions.

4.1. Real-Time Speed Prediction Model

To ensure a robust comparison between different LSTM model configurations, we employed a comprehensive methodology involving ten training and evaluation iterations per configuration, resulting in average RMSE values as the primary evaluation metric (see

Table 3. Traffic Lights Database LSTM models comparison.

Inputs	Mean RMSE
Speed	0.1537
Speed, acceleration	0.1310
Speed, acceleration, traffic light distance	0.1294
Speed, acceleration, traffic light distance, followed vehicle distance	0.1247

). In addition, the MAE is estimated between 0.123 for the speed-only model to 0.935 for the full-feature model.

A consistent trend emerged: As more input features were integrated, the mean RMSE decreased, indicating improved prediction accuracy. The best performance was achieved with four key features (speed, acceleration, traffic light distance, and followed vehicle distance), and resulted in the most accurate speed predictions. These results highlight the importance of considering both vehicle dynamics and external traffic context, incorporating proximities to other vehicles and traffic lights significantly enhances the model’s ability to precisely predict vehicle speed in real-time conditions.

This confirms that a multi-feature LSTM model is crucial for optimal speed prediction in an integrated energy management system (iEMS).

Figure 7 illustrates the outcomes of our optimal model, incorporating four inputs and the previously discussed filtering mechanism. We see a strong match between the predicted and actual signals. The model prediction aligns well with high and low speeds. It also responds well to changes from low to high speed and vice versa, indicating it has learned to use acceleration changes effectively. This configuration leads to stable, realistic speed predictions and moderate acceleration, aligning seamlessly with our iEMS framework.

Figure 7 shows the performance of our optimal LSTM model, which integrates

• Four key input features (speed, acceleration, traffic light distance, and followed vehicle distance).
• An exponential filtering mechanism for refining predictions.

We can summarize our findings as follows:

• Strong correlation between predicted and actual speed signals.
• Accurate predictions at both high and low speeds.
• Effective response to acceleration changes, ensuring smooth transitions from low to high speed and vice versa.
• Stable and realistic speed predictions lead to moderate acceleration values, a critical factor for FCHEV energy efficiency.

This refined speed prediction model seamlessly integrates into our iEMS framework, significantly enhancing real-time adaptability and energy management performance.

4.2. Speed Prediction Model iEMS Integration and Comparison

After integrating our LSTM speed prediction model, we evaluated its impact on real-time energy management.

In Figure 8, both costate trajectories exhibit oscillations within the range of 1.5 to 2, often stabilizing at 1.5, indicative of minimal FCS power usage. This observation underscores the substantial similarity between the two trajectories. Additionally, it is noteworthy that under the initial SOC condition of 35%, co-states exhibit higher values in the initial segment of the cycle, typically up to around the 200th second, in contrast to the rest of the cycle or to the corresponding portion of the cycle with an initial SOC of 65%. This disparity reflects the battery usage strategy’s successful efforts to replenish the battery when SOC is low and to maintain it when SOC is within an acceptable range, provided that power demand remains moderate. Consequently, we expect the SOC to remain within the 65% configurations and to increase within the 35% configurations, an aspect to be elaborated upon in the ensuing section.

However, a discernible distinction arises as our real-time iEMS demonstrates a marginally lower hydrogen consumption rate, resulting in a slightly improved fuel economy compared with our reference iEMS. This achievement is attributed to our LSTM model combined with our filtering method, enabling a slight underestimation of acceleration demands, thereby placing less strain on the FCS. Subsequently, we anticipate observing a lesser extent of battery stack replenishment by the real-time iEMS.

This achievement is attributed to our LSTM model combined with our filtering method, enabling a slight underestimation of acceleration demands. This results in less strain on the FCS, optimizing energy efficiency. Our enhanced real-time iEMS achieves better fuel economy and smoother energy management compared to the reference model.

Figure 9 consists of two subfigures for each test configuration, offering a comprehensive view of our findings. The left-hand side subfigures showcase the outcomes of our real-time iEMS, displaying power profiles for $P_{bus}$, $P_{batt}$, and $P_{fc}$, the system’s trajectory (i.e., SOC), and fuel consumption. On the right, subfigures offer a direct comparison between trajectories and battery energy consumption of our real-time iEMS and the benchmark iEMS, which operates with an ideal speed prediction model.

To thoroughly evaluate our real-time iEMS, Figure 9 presents two sets of subfigures for each test configuration. The left-hand side subfigures showcase the outcomes of our real-time iEMS, displaying the following:

• Power Profiles:

  ○

  $P_{bus}$: Power demand on the bus;

  ○

  $P_{batt}$: Battery power output;

  ○

  $P_{fc}$: Fuel cell system power output.

• State-of-Charge (SOC) Trajectory: Reflecting battery charge levels over time.
• Fuel Consumption: Evaluating hydrogen efficiency.

On the right-hand side, subfigures offer a direct comparison between the following:

• SOC trajectories of our real-time iEMS vs. benchmark iEMS (ideal speed prediction).
• Battery Energy Consumption, showing differences in energy usage.

Globally, the battery successfully manages high-frequency power demands, preserving the fuel cell system’s longevity and health. Also, the SOC is effectively regulated, increasing from 35% to approximately 45% or maintaining a satisfactory range close to 65%.

Nonetheless, distinctive differences emerge in the SOC trajectory between our real-time iEMS and the reference iEMS. The reference iEMS tends to use more battery energy, which was not anticipated based on the previous fuel consumption comparison, as this strategy tends to consume slightly more hydrogen. Once again, the considered speed prediction model compensates for a drawback in our reference iEMS by effectively mitigating the issue of over-reliance on the FCS. The real-time iEMS successfully balances energy sources, optimizing hydrogen efficiency and battery usage, proving its practicality for real-world deployment.

The battery effectively handles high-frequency power demands, reducing strain on the fuel cell system (FCS) and extending its lifespan. Also, the SOC regulation is a success with

• The 35% SOC scenario: Gradual increase to ~45%, ensuring stability.
• The 65% SOC scenario: Maintains a satisfactory range, balancing energy sources efficiently.

From a comparative point of view, the reference iEMS unexpectedly consumes more battery energy while also slightly increasing hydrogen consumption. Thus, our real-time iEMS counterbalances this issue by optimizing speed predictions, reducing over-reliance on the FCS, and improving overall fuel efficiency. Moreover, iEMS achieves smoother SOC regulation, more balanced power distribution between the battery and fuel cell, and reduced hydrogen consumption—up to 7.3% lower than the reference model. These results demonstrate that the real-time LSTM-based iEMS achieves near-optimal performance with improved energy efficiency under real-world prediction constraints. LSTM-based speed prediction model plays a crucial role in enhancing energy efficiency, demonstrating its potential as a pivotal component in iEMS.

While the proposed real-time iEMS framework shows promising results, several limitations should be acknowledged. First, the training and validation of the LSTM-based speed prediction model were based on a specific open-source driving cycle database. Although diverse, this dataset may not fully capture the variety of real-world driving behaviors, road conditions, and vehicle types. Future studies should incorporate additional datasets, including those from different geographic regions and real traffic recordings, to enhance model generalization and robustness. We also need, in this case, to use more rigorous statistical approaches for feature selection. Second, the current study focuses primarily on software architecture and algorithmic validation through simulation. While the results show clear improvements in fuel efficiency and power management, further work is needed to integrate the model into a hardware-in-the-loop (HIL) platform for real-time testing. Such experiments would help verify the responsiveness, latency, and stability of the control strategy under realistic computation and sensor conditions. Additionally, fuel cell degradation and thermal behavior were not explicitly modeled in the current system. These factors can significantly affect long-term energy management decisions and system health. Incorporating aging models and thermal constraints into the iEMS could further improve reliability and real-world applicability.

The final section will synthesize the study’s contributions, outlining its impact and future directions for optimizing real-time energy management in FCHEVs.

5. Conclusions and Future Work

This paper presented a real-time Integrated Energy Management Strategy (iEMS) for Fuel Cell Hybrid Electric Vehicles (FCHEVs), enhanced by an LSTM-based speed prediction model and a lightweight exponential filtering method. The proposed framework leverages speed predictions to anticipate future power demands and optimize energy distribution between the fuel cell and battery. Comparative evaluations with a reference iEMS using ideal speed profiles demonstrated that our approach achieves near-optimal control performance, while offering improved hydrogen efficiency, smoother SOC regulation, and more stable power sharing. The exponential filter not only reduced prediction oscillations but also improved overall predictive accuracy. Our real-time architecture achieved significant improvements while maintaining a practical computational load suitable for embedded control environments.

To estimate the practical impact of improved hydrogen efficiency, we translated the fuel savings into economic terms. Based on current hydrogen prices (e.g., $12–$16 per kg in most regions), a 7.3% reduction in hydrogen consumption would yield approximately $250–$300 in annual savings for a commercial FCHEV covering 20,000 km per year. This represents a tangible reduction in operating cost. However, it is also important to note that as prediction accuracy improves beyond a certain point, the marginal cost savings may diminish. A detailed life cycle cost (LCC) or total cost of ownership (TCO) analysis, following the SAE J2908 standard, could more precisely evaluate this trade-off and guide controller design decisions from an economic perspective.

Despite these promising results, some limitations remain. The training and evaluation of the LSTM model were based on a single open-source dataset, which may not fully capture the diversity of real-world driving conditions. In future work, we plan to train and test the prediction model on broader datasets, including real-world driving logs and varying geographic environments, to improve model generalization. We also intend to deploy the iEMS framework on a real EMS test bench for hardware validation. This will allow us to assess system responsiveness, latency, and stability under actual runtime constraints. Further research will also explore the integration of fuel cell aging models and thermal constraints into the control logic, enabling better long-term management of system health.