Toward Sustainable Mobility: AI-Enabled Automated Refueling for Fuel Cell Electric Vehicles

Abstract

With the global transportation sector being a major contributor to greenhouse gas (GHG) emissions, transitioning to cleaner and more efficient forms of transportation is essential for mitigating climate change and improving air quality. Toward sustainable mobility, Fuel Cell Electric Vehicles (FCEVs) have emerged as a promising solution offering zero-emission transportation without sacrificing performance or range. However, FCEV adoption still faces significant challenges regarding refueling infrastructure. This work proposes an innovative refueling automation service for FCEVs to facilitate the refueling procedure and to increase the fuel cell lifetime, by leveraging (i) Big Data, namely, real-time mobility data and (ii) Machine Learning (ML) for the energy consumption forecasting to dynamically adjust refueling priorities. The proposed service was evaluated on a simulated FCEV energy consumption dataset, generated using both the Future Automotive Systems Technology Simulator and real-time data, including traffic information and details from a real-world on demand Public Transportation service in the Geneva Canton region. The experimental results showcased that all three ML algorithms achieved high accuracy in forecasting the vehicle’s energy consumption with very low errors on the order of 10% and below 20% for the normalized Mean Absolute Error and normalized Root Mean Squared Error metrics, respectively, indicating the high potential of the suggested service.

1. Introduction

On the road toward sustainable transportation, Fuel Cell Electric Vehicles (FCEVs) have emerged as a promising solution by offering zero-emission transportation without compromising performance or range. Unlike traditional internal combustion (IC) engine cars, FCEVs operate on the principle of converting hydrogen fuel (H₂) into electricity through a chemical reaction within the fuel cell stack [1], emitting only water vapor and heat as byproducts [2]. Once fed into the fuel cell, hydrogen, containing roughly the same energy content as a gallon of gasoline per kilogram, has the same potential to power vehicles effectively without releasing harmful emissions [3]. As a result, this transformative technology presents itself as a viable alternative to fossil fuels [4] and holds the potential to redefine the transportation sector by offering zero-emission vehicles, thereby reducing greenhouse gas (GHG) emissions, combating climate change, as well as addressing the ever-growing concerns surrounding air quality and energy security.

As the transportation sector is the main contributor to GHG emissions [5], transitioning to greener forms of transportation is of utmost significance to alleviate climate change and improve air quality [6]. The European Union (EU) alone is committed to decarbonizing its energy systems by reducing its carbon emissions by 95% by 2050, focusing on hydrogen technologies [7]. Hydrogen, being the primary fuel for FCEVs, can be produced from various renewable sources [8], including water electrolysis using renewable electricity, biomass gasification and steam methane reforming with carbon capture and storage. This versatility in hydrogen fuel production constitutes FCEVs as the optimal alternative to conventional fossil fuel vehicles (gasoline or diesel), enabling them to utilize Renewable Energy Sources (RESs), thereby significantly reducing carbon emissions and dependence on finite sources [9], laying the foundations toward achieving sustainable mobility [10].

However, despite their aforementioned environmental benefits and technological advancements, FCEVs still face significant challenges, especially in their refueling infrastructure [11]. While conventional gasoline and diesel vehicles can be refueled at existing gas stations in just a few minutes, FCEVs require specialized hydrogen refueling stations with highly pressurized dispensing equipment. This limited availability of the hydrogen refueling infrastructure poses a significant obstacle to the widespread adoption of FCEVs [12], as drivers often face long wait times, reducing their practicality and convenience [13]. Finally, an important factor that needs a lot of attention from the research community is related to the degradation of the fuel cell lifetime, which is mainly affected by the varying load profiles and the complex operating conditions [14].

This article proposes an innovative solution to improve the current situation in hydrogen refueling procedures through the integration of Artificial Intelligence (AI) and Big Data (BD). More specifically, by leveraging three supervised Machine Learning (ML) algorithms for energy consumption forecasting, this work presents a state-of-the-art refueling automation service for FCEVs, particularly tailored to streamline the refueling process and enhance the overall user experience for FCEV drivers. Moreover, by integrating demand and energy consumption data from a real-world on-demand Public Transportation (PT) service in the Geneva Canton region and by utilizing open APIs for real-time traffic data, the proposed system could dynamically adjust refueling priorities based on energy consumption and traffic conditions, optimizing the refueling efficiency, and thus minimizing idle times for FCEV drivers and improving the fuel cell lifetime.

The remainder of this article is organized as follows. Section 2 focuses on the current research landscape on refueling automation approaches by methodologically examining already existing strategies and comparing them to the proposed service. In Section 3, the methods followed during the data acquisition and preprocessing are described along with the consumption simulation of the recorded trips, while in Section 4, the methodology behind both the proposed energy consumption forecasting and the refueling automation service is analyzed. Section 5 focuses on the experimental results of this work by presenting the main forecasting results of each algorithm. Finally, Section 6 summarizes the key contributions and implications of this work and proposes future work directions toward sustainable and green mobility through AI-enabled automated refueling for FCEVs.

2. Technological Background and Research Contribution

In this section, the current research landscape on both battery electric vehicles (BEVs) and FCEVs is presented, offering a comparative analysis on their working principles and current infrastructure, and highlighting the key FCEV challenges and advancements towards their widespread adoption. Finally, some key related optimization services from the current literature are presented, tailored to either BEVs or FCEVs, followed by the key contributions of this work in the field of FCEV refueling automation.

2.1. BEVs vs. FCEVs: Working Principles, Infrastructure and Challenges

On the road toward sustainability, both BEVs and FCEVs represent two promising alternatives, significantly contributing to GHG emission mitigation compared to traditional IC vehicles [15]. However, despite their environmental benefits, these two technologies are quite different in terms of their operational mechanisms, charging or refueling requirements, as well as their supporting infrastructure.

Conventional BEVs are charged using electricity and their charging times typically vary depending on both the vehicle’s battery capacity and the charging station’s battery output. For conductive charging, there are three main power levels classified from slow or “home charging” (level 1) that is typically used for overnight charging, to faster or “fast AC charging” (level 2), which is commonly found at public charging stations, to ultra-fast or “fast DC charging” (level 3) that offers charging times of 30 minutes or less [16]. Over the last few years, BEV charging infrastructures have rapidly expanded, showing an ever-growing network with the continuous development of public charging stations.

On the other hand, the FCEV refueling operation follows a completely different working principle. FCEVs operate by converting hydrogen fuel into electricity through the fuel cell on board [17]. The process is quite similar to refueling a conventional IC vehicle by dispensing hydrogen through a highly pressurized nozzle rather than fossil fuels. For this reason, refueling times are vastly different from the aforementioned BEV charging times and typically range from 3 to 5 min. However, unlike the easily accessible BEV infrastructure, hydrogen refueling stations are significantly less and primarily located in major urban centers and along transportation corridors, due to the large amounts of energy required for its production and storage.

Apart from the scarcity in hydrogen refueling stations, another significant obstacle to the adoption of FCEVs is the hydrogen production cost. Until now, the majority of hydrogen produced is derived from fossil fuels, primarily natural gas, due to its lower cost, which significantly contributes to GHG emissions by releasing large amounts of carbon dioxide in the atmosphere. Green hydrogen is a new alternative that proposes the production of hydrogen through electrolysis by using electricity to split water into hydrogen and oxygen [18]. However, although this is a more sustainable practice, the dependence on RESs and the dedicated equipment needed for this procedure increase the overall cost of the procedure.

Despite the aforementioned challenges, the field of FCEVs showcases significant technological improvements on the road toward enhancing their viability as a new sustainable transportation option. Material innovations (e.g., nanostructured catalysts and non-platinum alternatives) play a key role towards this goal, offering increased catalytic activity while simultaneously reducing the amount of precious metals required, while control strategies also showcase major improvements in enhancing the vehicles’ performance under various operating conditions [19]. Following the EU’s decarbonization initiatives, several co-funded FCEV projects have already been launched to expand the hydrogen refueling infrastructures and increase the FCEV fleets on the roads [7]. Finally, an important factor studied in the literature is related to the cost reduction efforts that are continuously being investigated [20].

The rapid evolution of AI also significantly contributes to the advancement of FCEVs by enabling more advanced and adaptive control systems compared to traditional rule-based approaches [21]. Common energy management optimization approaches opt for AI techniques, such as Artificial Neural Networks (ANNs), fuzzy logic controllers, reinforcement learning, and genetic algorithms, while approaches focusing on multi-objective optimization propose the use of swarm intelligence and physics-inspired algorithms. Finally, ML- and Deep Learning (DL)-based algorithms have also been proposed to enhance FCEV efficiency and reliability, opting for better performance predictions and fault diagnostics, making AI a considerable solution to overcoming current FCEV limitations and enabling significant improvements in efficiency and cost-effectiveness.

2.2. Related Work

Vehicle refueling and charging systems, specifically concerning FCEVs and BEVs in general, have reached the point where their technological advancement meets environmental responsibility, especially in the context of reducing GHG emissions. While the development of innovative refueling and charging infrastructures showcases great potential for improving vehicle efficiency and grid sustainability, the implications for GHG emissions are complex and require continued investigation. For this purpose, the transition toward sustainable transportation is increasingly focused on improving the efficiency and accessibility of refueling and charging infrastructures, requiring innovative approaches to improve the operation, safety and sustainability of these systems. The articles reviewed in this section provide insights into diverse approaches, ranging from hydrogen refueling station optimization to dynamic and automated EV charging systems, focusing on efforts to incorporate RESs and adopt intelligent energy management strategies.

Regarding hydrogen refueling station optimization, these approaches are mainly focused on improving operational efficiency, sustainability and safety in hydrogen refueling infrastructures. More specifically, Sun et al. [22] developed a comprehensive optimization framework for hydrogen refueling stations with on-site electrolytic hydrogen production, using simulation methods to predict hydrogen demand for effective resource allocation and scheduling. This approach not only reduces operational costs but also increases efficiency by aligning production capacity with peak demand. Yan et al. [23] expanded this scope by integrating RESs into hydrogen production and storage, aiming to create more sustainable and efficient hydrogen refueling stations that reduce their dependence on conventional energy sources while also increasing the environmental benefits of hydrogen fuel. On the other hand, focusing specifically on the safety aspects of hydrogen refueling stations, Yang et al. [24] employed advanced ML techniques to analyze and mitigate potential risks, ultimately supporting public acceptance and regulatory compliance.

Continuing with the research studies concerning EVs, and regarding the general concept of optimizing EV charging systems, the main concern is to investigate ways to improve the operational efficiency and user experience of EV charging stations by using advanced technologies for dynamic and automated recharging. A study by Cao et al. [25] proposed the development of a smart charging algorithm that employs a customized actor–critic learning method to dynamically adjust EV charging schedules based on real-time grid demands and vehicle arrival patterns. The proposed AI-driven strategy effectively manages the load on the power grid while optimizing charging times, enhancing overall efficiency. Moreover, Hirz and Lippitsch [26] also investigated automated recharging technologies for EVs, including the use of robotic systems and inductive charging techniques, to ultimately improve the user experience by simplifying the charging process and minimizing the need for physical interaction, which is intended to foster their wider adoption and regular use. Finally, Zhang et al. [27] also contributed to this context by specifically focusing on dynamic and real-time energy management strategies through using reinforcement learning (RL) to manage the load scheduling in EV charging stations efficiently. This approach adapts to real-time changes in grid demands and EV arrival patterns, optimizing charging schedules to enhance grid stability and reduce electricity costs.

However, as mentioned above, incorporating renewable energy into EV charging infrastructures is of utmost importance to attaining sustainability and enhancing grid independence. In this context, Liu et al. [28] introduced a novel approach for optimizing the integration of EVs into the power grid by utilizing RESs. Their approach is mainly focused on enhancing the reliability and efficiency of power systems while managing the variability of RESs and their impact on grid stability. Another similar approach by Chellaswamy et al. [29] explored the direct usage of solar energy to power EV charging stations, offering a sustainable and autonomous energy solution that allows EV charging infrastructures to operate more independently from the traditional power grid, thereby supporting the broader adoption of electric mobility and minimizing reliance upon conventional power sources.

Finally, additional data-driven approaches have also been proposed in relation to dynamic wireless and real-time charging by adapting to operational and grid constraints. More specifically, Palani and Sengamalai [30] examined the advancements in wireless power transfer systems, and particularly the resonant inductive charging technology that facilitates dynamic charging of EVs, allowing vehicles to be charged while in motion, thus significantly reducing charging downtime and potentially extending the vehicles’ operational range. Additionally, Tian et al. [31] proposed a real-time recommendation system for EV taxis, which leverages large-scale GPS data mining to efficiently allocate charging station resources. By predicting the operational states of EV taxis, the system was able to suggest the most strategically located charging stations in terms of timing, thereby minimizing idle times and enhancing service efficiency. Lastly, Suanpang and Jamjuntr [32] employed a multi-agent RL framework that dynamically allocates EV charging tasks across various stations to minimize congestion and balance load. Their proposed approach optimizes the distribution of EVs across available charging stations, reducing wait times and ensuring effective use of the charging infrastructure.

2.3. Research Contribution

Considering the limited research work in FCEVs for similar services and the fact that the aforementioned methodologies primarily focus on operational efficiency and safety, the proposed automated refueling service introduces a novel integration of (i) AI algorithms and specifically, three supervised ML techniques, and (ii) BD in the form of real-time data to optimize the hydrogen refueling process for FCEVs. The main purpose of this work is to propose an early-stage solution to the current literature by addressing key challenges in FCEV refueling procedures, which are significantly different than conventional BEV recharging.

Unlike already existing studies that mainly focus on EV charging or propose more generic approaches for hydrogen station optimization, as presented in

Table 1. Key contributions and related EV refueling approaches.

Work	Year	Main Focus	EV Category	AI Integration
Tian et al. [31]	2016	Dynamic Wireless Charging	BEV	Full
Chellaswamy et al. [29]	2020	Renewable Energy for Charging	N/A	N/A
Cao et al. [25]	2021	EV Charging System Optimization	BEV	Full
Liu et al. [28]	2021	Renewable Energy for Charging	BEV	Partial
Yang et al. [24]	2021	H2 Refueling Station Optimization	N/A	Partial
Zhang et al. [27]	2021	EV Charging System Optimization	EV	Full
Sun et al. [22]	2022	H2 Refueling Station Optimization	N/A	Partial
Hirz and Lippitsch [26]	2023	EV Charging System Optimization	BEV	Partial
Palani and Sengamalai [30]	2023	Dynamic Wireless Charging	BEV	N/A
Yan et al. [23]	2023	H2 Refueling Station Optimization	N/A	N/A
Suanpang and Jamjuntr [32]	2024	Dynamic Wireless Charging	BEV	Full
Proposed Service	2024	H2 Vehicle Refueling Automation	FCEV	Full

, our proposed solution specifically targets the inherent complications that are present in all FCEV refueling procedures. By leveraging real-time traffic and vehicle data and existing booking information, the suggested solution dynamically adjusts refueling priorities based on actual vehicle arrival times and traffic conditions, not only minimizing wait times but also ensuring a seamless and more efficient refueling experience, tailored to the individual needs and schedules of FCEV drivers. As a result, this service significantly improves operational throughput and driver satisfaction by optimizing the scheduling and execution of refueling operations.

3. Data Acquisition and Energy Consumption Simulation

In this section, the full methodology of the artificial dataset generation is presented in detail, focusing on both the original booking and traffic datasets and the subsequent preprocessing steps to export the necessary demand and stop probability calculations, leading to the construction of the finalized simulated FCEV consumption dataset.

3.1. Original Datasets and Preprocessing

The initial datasets included both existing trips and booking records of user pickups and dropoffs, as well as real-time traffic speed data from an open API for the Geneva Canton region in Switzerland for one month, as presented in

Table 2. Statistics of both the original and the processed (ver. II) booking and traffic datasets.

Dataset	Size (MB)	Duration (Days)	Time Frame (Hours)	Sampling Time (Seconds)	Samples
Bookings	0.506	22	14	–	1074
Traffic	0.440	30	24	1800	14,376
Bookings-II	0.393	22	14	–	840
Traffic-II	26.751	22	14	1	1,108,800

. However, unlike the initial traffic speed dataset, the original bookings dataset only included trip information for the working days (i.e., Monday to Friday) and operational hours (i.e., from 08:00 to 22:00) of each week of that month. Hence, a processed version of the traffic dataset (Traffic-II) was created to showcase the respective configurations, in which only the working days and operational hours of the week were kept as well. Therefore, the final traffic dataset contains 22 working days and 14 h intervals for each day of the week, as described in Table 2. On the other hand, the only difference between the original and the processed booking dataset (Bookings-II) lies in the exclusion of every booked trip that was cancelled either by the driver or the user.

The aforementioned datasets are eventually intended to be integrated with the Future Automotive Systems Technology Simulator (FASTSim) tool to generate the artificial FCEV consumption dataset; thus, maintaining compatibility with the tool is an essential aspect. For this reason, the following procedure was followed. More specifically, the traffic dataset collects real-time vehicle speed data at 3-minute intervals to comply with the TomTom API’s (https://www.tomtom.com/, accessed on 12 March 2024) recommended frequency for data retrieval, ensuring consistent and efficient data collection. On the other hand, FASTSim [33] requires data inputs for each second to provide realistic simulations. Therefore, to ensure a seamless integration with the FASTSim tool, the final preprocessing step involved the interpolation of the speed measurements to fill in the gaps between the original 3-minute intervals, ensuring that speed data were available for each second within the operational hours.

3.2. Demand and Stop Probability Assessment

To further improve the realism of the artificially generated dataset, additional demand and stop probability calculations have also been considered, following well-established approaches from the literature [34]. This information is integrated into the Bookings-II dataset, after filtering out any cancelled trips and segmenting the remaining data into 14 h operational intervals for each working day of the week, as mentioned in the previous section.

To effectively calculate the demand of each stop included in the Bookings-II dataset, the number of each stop’s occurrence (stop ID) in the dataset, specified as either a pickup or a dropoff location, was combined. The heatmap provided in Figure 1 depicts the demand for each stop (either pickup or dropoff) across different days of the week, with each row corresponding to a specific stop ID, while each column represents a working day of the week. From this heatmap, it is evident that certain stops, such as stop IDs 20, 42 and 43, showcase higher demand on specific days, especially on Tuesdays and Wednesdays, suggesting that they are critical points in the transportation network.

Finally, based on the aforementioned demand analysis, additional probability calculations on the minibus’s chance of actually stopping at a specific stop have also been implemented. The probability examined for the purposes of this work is derived based on the number of stops already made during the specific time interval on a particular day of the week.

3.3. Artificial Dataset Generation

The last step of this methodology contains the generation of the artificial energy consumption dataset to effectively evaluate the performance of the proposed automated refueling service. Despite the lack of real-world FCEV consumption data, the aforementioned combination of the FASTSim simulator with real-time data from the Geneva Canton reflects real-world driving scenarios for FCEVs in terms of energy usage and is in accordance with the related literature [35]. More specifically, a new simulated FCEV dataset (CERTH-FCEV) was created, incorporating vehicle telemetry data, stop information and real-time traffic conditions, as described in the previous sections, which will be later used as an input for the proposed automated refueling service.

The first step for the CERTH-FCEV dataset creation involved the specification of key vehicle parameters that affect energy consumption, such as the vehicle’s weight, battery capacity, and fuel cell stack characteristics, as well as other relevant drivetrain specifications. These parameters were carefully selected to represent the Toyota Mirai’s real configurations, ensuring that the simulation results can be applicable to actual vehicle performance [36,37,38].

After the data configuration, the simulation was run continuously for a total of 307 h, covering the entire period of the 22 days of collected data. During this period, the simulator generated detailed energy consumption data for the FCEV Toyota Mirai. After completing these simulations, the final energy consumption data were extracted, including hourly consumption metrics in kilowatt-hours (kWh), providing a precise quantification of the energy used by the Toyota Mirai for the same route, under various driving conditions. The whole procedure for the artificial dataset generation is depicted in Figure 2.

4. Methodology

This methodology section is designed to provide (i) a comprehensive analysis of the the three forecasting algorithms used for the energy consumption forecasting, along with (ii) the presentation of the refueling automation service that accounts different factors, including among others the critical output of the energy consumption forecasting algorithms.

4.1. Energy Consumption Forecasting

The utilization of predictive models in this transport sustainability context aims to address critical challenges in energy management and operational efficiency, which are crucial to increasing the adoption and sustainability of FCEVs. In this study, three forecasting models were implemented, representing three different families of ML, namely an Auto-Regressive Integrated Moving Average (ARIMA) model, representing the statistical-based ML family; a Random Forest (RF) model, representing the ensemble ML family; and lastly, a Long Short-Term Memory (LSTM) model, representing the DL family, each chosen for their specific strengths in handling time-series data and forecasting accuracy.

More specifically, the ARIMA model was used in this study because it is effective in time-series forecasting, especially for datasets with patterns and seasonality, such as the case of transportation. ARIMA models are capable of identifying basic patterns in historical data and producing accurate forecasts [39]. The model is appropriate for energy consumption forecasting because it can divide time-series data into autoregressive (AR), differencing (I), and moving average (MA) components. This allows the model to manage a variety of data characteristics. This study’s utilization of ARIMA aims to take advantage of its statistical ability in order to generate accurate and understandable forecasts, which are crucial for organizing and maximizing FCEV refueling operations.

On the other hand, the RF model was employed in this study due to its remarkable accuracy in detecting complex, non-linear relationships within the data. RF’s ensemble nature, which reduces overfitting and increases generalizations, allows it to perform better than many other traditional ML methods in both regression and classification tasks [40]. In order to improve predictive performance, this model builds several decision trees during training and averages their results. With regard to energy consumption forecasting, the RF model works especially well for short-term forecasts because it is effective at finding patterns and anomalies in huge datasets.

Finally, LSTM was the last model selected because of its proven record of success with sequential data processing, especially when it comes to energy consumption predictions. Mahjoub et al. [41] have demonstrated that LSTM models are highly effective in forecasting future power loads due to their ability to identify long-term dependencies in time-series data. This skill is essential for accurately forecasting patterns of energy consumption, which might be impacted by past usage over long time periods. Because of its architecture, which incorporates memory cells to store information for extended periods of time, the LSTM model outperforms conventional Recurrent Neural Networks (RNNs) [42] at predicting complex temporal sequences. Considering its ability to handle non-linear relationships and sequential dependencies, LSTM is especially well suited for use in energy management and smart grid applications, where accurate forecasting is crucial for maximizing resource allocation and minimizing energy waste.

It should be noted here that although in the previous paragraphs, the suitability of the selected aforementioned ML algorithms for the forecasting task is justified, a similar consideration with the work of Spanos et al. [43] was followed. Hence, the main purpose of this research work is neither to propose a novel time-series forecasting algorithm nor to compare various ML algorithms, but to emphasize the generalizability of the proposed AI-Enabled Automated Refueling service for FCEVs by adopting three well-established and stable algorithms from different ML families, such as ARIMA, RF, and LSTM.

4.1.1. Data Preparation

To ensure effective implementation and evaluation of the aforementioned forecasting models, a series of some initial data preparation steps were implemented to ensure accurate and reliable forecasting results. The first step included the effective data loading and preprocessing, where the energy consumption data, indexed by hours, were loaded into the system. In this step, initial preprocessing of the data was also performed, which involved handling missing values and ensuring data consistency. For the LSTM model, additional normalization was performed to scale the data within a specific range, ensuring efficient training and preventing bias toward larger values.

After the data loading was completed, the energy consumption dataset was then split into training (80%) and test (20%) sets, without shuffling, to preserve the temporal order. This split replicates real-world forecasting scenarios in which training includes the unknown future data points. The training set was used to build the models, while the test set was reserved for performance evaluation. For the LSTM model, an additional step was implemented that involved the data sequence generation. More specifically, the data were converted into sequences of input–output pairs, creating sliding windows over the time series to capture temporal dependencies, enabling the model to learn from the sequence of previous time steps.

4.1.2. Model Training

In this section, an overview of each forecasting algorithm, including its architecture and training process for effective energy consumption forecasting, is offered. Moreover, for the parameter fine-tuning of the three forecasting models, a typical validation procedure is followed, as described below, specifically catered to each model.

The ARIMA model is designed to predict future points in time series by expressing the data in terms of their own lagged values (autoregressive), the difference in the values (integrated), and forecast errors (moving average). For this study, in order to find the optimal ARIMA parameters (p, d, q) that minimize the Akaike Information Criterion (AIC), a grid search approach was employed, taking into consideration the dataset’s characteristics. The final model configuration includes an autoregressive order of 3 (p = 3) to capture the influence of the last three observations, one differencing step (d = 1) to make the series stationary, crucial for the effectiveness of the AR and MA components, and a moving average order of 3 (q = 3) that will help in modeling the shock effects from previous forecast errors up to three time points back, resulting in an ARIMA(3, 1, 3) model. The proposed model was trained using a rolling forecast method, which involves fitting the model on a moving window of data, updating the model with new data points and predicting one step ahead at each point. This method simulates real-time forecasting and ensures that the model adapts to new information.

On the other hand, as part of its ensemble learning process, the RF model builds many decision trees during training and outputs the average forecast of each tree, thereby increasing the model’s accuracy and reducing overfitting by using the variety of decision trees constructed on various subsets of the training data. In this study, the energy consumption data were converted into a supervised learning format, where each input includes data from previous time steps and the related output is the energy consumption at the current time step. The proposed RF model uses 1000 decision trees (n_estimators = 1000), ensuring robust performance by averaging multiple predictions with a fixed random state of 42 to ensure reproducibility of the model training process. The final RF model was trained using an ensemble of decision trees, each built on different subsets of the training data, and tested using walk-forward validation, where the model is iteratively retrained with each new data point from the test set, allowing for continuous adaptation and evaluation.

Finally, the LSTM model is a form of RNN that is particularly effective at processing and making predictions based on time-series data due to its ability to capture long-term dependencies. Unlike standard RNNs, LSTMs have a special architecture which includes memory cells that can maintain information in memory for long periods of time, being especially crucial for predicting data sequences where the context of a considerably earlier time step may still be significant. For the purposes of this study, the proposed LSTM model features units with 64 neurons each and comprises two dense layers with 32 neurons each with the ReLU as the activation function. The final layer is a dense layer with a single neuron used to predict continuous values of energy consumption, directly outputting the forecast. The proposed model was trained on partially normalized sequences of input–output pairs of historical data that were specifically created for this purpose, while its hyperparameters were tuned using the validation set to prevent overfitting and ensure that the model generalizes well to new data.

4.1.3. Model Evaluation

To ensure the efficiency of the aforementioned forecasting algorithms, a thorough evaluation procedure was carried out by employing six main evaluation metrics, including Mean Absolute Error (MAE), Median Absolute Error (MdAE), and Root Mean Squared Error (RMSE), as well as their normalized versions NMAE, NMdAE, and NRMSE, respectively. Each of these metrics is essential for measuring each model’s reliability and forecasting accuracy in real-world scenarios and is used in ML regression tasks in various sectors in the literature [44,45,46].

More specifically, MAE measures the average absolute difference between the actual and predicted values as follows:

$$\mathrm{MAE}={\displaystyle \frac{1}{n}}\sum _{i=1}^n\left|y_i-{\widehat{y}}_i\right|$$

(1)

where $y_i$ are the actual values, ${\widehat{y}}_i$ are the predictions, and n is the total number of observed data points (hours). MAE measures the average magnitude of the errors in a set of predictions, providing a straightforward representation of prediction accuracy on the same scale as the data.

On the other hand, MdAE measures the median of the absolute differences between the actual and predicted values, as explained in Equation (2), and for this reason, it is robust to outliers.

$$\mathrm{MdAE}=\mathrm{median}\left(\left|y_i-{\widehat{y}}_i\right|\right)$$

(2)

Finally, RMSE (Equation (3)) is the square root of the Mean Squared Error (MSE) metric, which, being based on the the squared errors, is eventually more strict to large errors compared to MAE.

$$\mathrm{RMSE}=\sqrt{{\displaystyle \frac{1}{n}}\sum _{i=1}^n{\left(y_i-{\widehat{y}}_i\right)}^2}$$

(3)

For the normalized versions of the three aforementioned evaluation metrics, the same consideration with the research study of Polymeni et al. [46] has been followed, where each normalized metric (NMAE, NMdAE, NRMSE) scales the corresponding one (MAE, MdAE, RMSE) by the range of actual values ($max\left(y\right)-min\left(y\right)$), thereby facilitating comparison across different datasets.

4.2. Automation of the Refueling Procedure

The automation of the refueling procedure for FCEVs is a critical component in enhancing the efficiency and sustainability of hydrogen-powered transportation. This suggested service proactively schedules refueling sessions at optimal times and locations for individual vehicles by utilizing important information from the energy consumption predictions combined them with BD such as real-time mobility data. This integration of real-time energy consumption data and predicted data aims to automate the refueling process, thereby increasing the lifetime of fuel cells and eventually enhancing the overall energy efficiency, convenience, and user experience of FCEVs.

The refueling automation algorithm includes a number of essential steps as depicted in Figure 3. Indeed, the offered service optimizes refueling schedules by taking into account multiple criteria, such as the energy consumption predictions, the positions of the vehicle and the nearby fueling stations, the current fuel level, and the booking information. Hence, considering all the aforementioned, the refueling automaton algorithm constitutes an iterative process that starts by calculating the fuel level for the next hour. According to this level and the optimal level for refueling, the algorithm decides if it goes to the start of continues to the next step. Therefore, if the next hour fuel level is below the optimum, the algorithm calculates the time needed to the nearest refueling station. After that, another decision should be performed, related to the comparison between the next hour fuel level and the critical threshold. If the next hour fuel level is below the critical threshold, then a refueling plan containing information for the nearest refueling station and the time that the driver should go is generated. Otherwise, a new decision is made from the algorithm according to the booking information. More precisely, if there is no booking for the time needed for refueling (including all the required steps, namely, arrival to the refueling station, refueling, and return), then the refueling plan is generated. Otherwise, the algorithms goes to the start.

This aforementioned service aims to reduce the need for drivers to manually search for available refueling stations, which can often be time-consuming and inconvenient, especially in areas with limited infrastructure. Instead, drivers can rely on the system to ensure that their vehicles are refueled without causing any disruptions to their schedules. By integrating BD and AI, the system can anticipate when a vehicle requires refueling and automatically schedule a refueling session at the most convenient time and location. Through this continuous coordination, refueling becomes more efficient and idle times in refueling stations decrease by more equally dividing demand.

5. Experimental Results

This section provides a thorough analysis of the predictive performance of the three forecasting models mentioned in the previous section, namely, ARIMA, RF, and LSTM. Each model was evaluated using the MAE, MdAE and RMSE metrics, as well as their normalized versions, as described in the previous section. At the end of this section, a comparison between all three of the selected models is also offered.

5.1. ARIMA Model Performance

In Figure 4, the actual and the predicted energy consumption values from the ARIMA model are depicted, reflecting its predictability. Continuing with the evaluation metrics, the ARIMA model achieved a very low MAE of 0.48 kWh, indicating that its predictions were very close to the actual energy consumed. Furthermore, ARIMA achieved an extremely low MdAE value (0.22 kWh), showing its high forecasting accuracy in this evaluation metric as well. Moreover, ARIMA has a low RMSE value (0.93 kWh), reflecting its forecasting power by not producing many significant errors. Finally, the normalized errors are 9.72%, 18.82%, and 4.39% for the NMAE, NMdAE, and NRMSE, respectively, demonstrating once again that the ARIMA model achieved its purpose of accurately forecasting hourly energy consumption.

5.2. RF Model Performance

In Figure 5, the predictions from the RF algorithm along with the actual values are shown. It is obvious from the graph that RF captured the pattern of the energy consumption. Regarding the evaluation metrics, the values of MAE (0.49 kWh), MdAE (0.2 kWh), and RMSE (0.97 kWh) are very low and similar to those of ARIMA, indicating the high accuracy of this algorithm as well. All the normalized metrics are below 20%, even for the most strict NRMSE, supporting the previous outcome of the RF’s high accuracy. Hence, it is obvious from the aforementioned that RF also achieved a very high performance in energy consumption forecasting.

5.3. LSTM Model Performance

The predictions of the last model, which is LSTM, along with the actual energy consumption values are shown in Figure 6. As with the previous two models, LSTM was able to capture the pattern of the energy consumption. Moreover, with respect to the six evaluation metrics, the results are 0.54 kWh, 0.25 kWh, and 0.95 kWh, for the MAE, MdAE, and RMSE, respectively, whereas the normalized values of these are 10.9%, 5.11%, and 19.29%. Hence, LSTM, which represents the DL family of the algorithms used, also achieved the purpose of energy consumption forecasting with high accuracy.

5.4. Model Comparison

As presented before, all models used are pretty accurate according to the results and capture the data pattern very well, although their predictions are narrower compared to the actual data. Figure 7 depicts the comparison of the three models selected for the energy consumption forecasting. Although the model comparison does not belong to the objectives of this research study, since the three algorithms were selected in order to include representative algorithms from different ML families (statistical-based, ensemble, DL-based) in the present analysis, it is more than clear from this graph and the corresponding results presented that all ML algorithms achieved comparably accurate results showing very little differences between them. Hence, considering the nature and the size of the data sample, for the energy consumption forecasting task, all three selected algorithms can achieve very accurate results. However, in ML, when different models achieve similar results, the simplest one is selected in favor of variance [47]. Therefore, in this case, the selected model will be ARIMA.

6. Conclusions and Future Work

This work proposes a refueling automation service specifically catered to FCEVs by integrating real-time speed data and simulated vehicle energy consumption, along with AI-based models, with the main objective of addressing the critical limitations currently faced by hydrogen refueling infrastructures. The proposed automated refueling service combines high-accuracy forecasting models, from different ML families, including ARIMA, LSTM, and Random Forest, to forecast the vehicle’s future energy consumption within the next hour. The accurate results presented highlighted the effectiveness of the service, which is validated through a simulated FCEV energy consumption dataset (CERTH-FCEV), created from real-time traffic data and actual booking information from an on-demand Public Transportation service in the Geneva Canton region.

The key advantages of this AI-driven approach include improved operational efficiency of hydrogen refueling stations, enhanced user experience for FCEV drivers, and potential increases in fuel cell lifetime through optimized refueling schedules, specifically tailored to each driver. In addition, the integration of real-time data and AI forecasting approaches enables dynamic decision-making, adapting to changing traffic conditions and energy demands. However, the reliance on simulated rather than real-world FCEV consumption data may impact the model’s accuracy in practical applications, which, in the case of scaling the system to accommodate a large fleet of vehicles, could introduce additional challenges in data processing and infrastructure requirements.

In the future, real-world energy consumption data from an FCEV will be used to train the forecasting models. In addition, with respect to the energy consumption forecasting, customized, more advanced, and sophisticated ML algorithms, using various factors, apart from the previous energy consumption values, will be used to improve the forecasting accuracy. Finally, this service will be showcased and evaluated in a real-world scenario, and specifically, to facilitate the energy refueling of an FCEV running on an on-demand PT platform in the Geneva Canton region. The system’s successful deployment in the pilot site in collaboration with a public transport operator authority, as part of a European research program, could potentially inform future policy decisions related to the development of hydrogen refueling infrastructures and the promotion of FCEV adoption.