Electric Vehicle Energy Management via Traffic Light Detection and Segmental Velocity Forecasting

As one of the most promising approaches to address the climate change caused by massive transportation, lithium-ion battery-based vehicle electrification has emerged, developed, and advanced with an overwhelming speed during the past decade. However, there still exist several technical barriers preventing its further expansion for electric vehicles (EVs) such as battery safety, battery lifespan, and driving range anxiety. A safer battery pack with higher energy density, larger capacity, and higher reliability has always been a consistent aspiration for the fast-growing EV industry. Theoretically, these concerns can be potentially fully mitigated as battery technologies advance. At the current stage, multiple feasible solutions have been developed from cell to pack simultaneously, such as seeking high energy density electrode materials and solid state electrolytes, employing integrated cell geometry designs like cylindrical type-4680 and prismatic blade battery, and utilizing space-efficient pack assemble methods like cell-to-pack and cell-to-chassis [1]. Additionally, based on these readily available systems, it is achievable to alleviate the range concerns by improving system-level energy efficiency via predictive thermal control and optimal power allocation.

Extensive studies have shown that the performance of lithium-ion batteries is highly influenced by temperature variations. High temperatures can lead to faster degradation and potential thermal runaway, while exposure to temperatures below $0∘C$ can cause irreversible damage due to lithium plating, resulting in a significant reduction in power and capacity. State-of-the-art BTMS use a variety of heat transfer mediums to maintain appropriate temperatures, including air, liquid, phase changing materials (PCMs), and heat pipes. Early BTMS focused on developing methods to remove the waste heat from a single battery system, with active air cooling commonly used in early hybrid EV models such as the Toyota Prius and Nissan Leaf. However, after Tesla successfully integrated its liquid-based battery thermal system with the vehicle air conditioning (AC) system, research on BTMS broke into two different scopes and developed in parallel.

Regarding component-level research, the relationships between the cell and thermal medium have been mainly investigated in three topics. First, novel thermal materials like PCM, heat pipe [2], and hybrids among them have been evaluated with improved performance, but the applications of these materials are still limited to laboratory level due to their intrinsic complexity. Second, ventilation arrangements like Z, U, and J-type [3], spoiler [4], and their further advancements [5,6] have been designed and proven to have better thermal performances in terms of maximum temperature control and uniformity. Similarly, three major liquid-based structures have been developed based on different battery geometries. Third, multiple linear and nonlinear algorithms have been developed to constrain battery temperatures for daily driving. Additionally, model predictive control (MPC) that foresees several steps ahead and yields an optimized control solution has also been employed for real-time component-level dynamic controls.

At the system level, robust battery thermal control requires strong heat dissipation capacity to remove waste heat under heavy working conditions, which typically employs liquid as the heat transfer medium and is coupled with an onboard AC system or radiator. Optimizing battery and cabin thermal controls simultaneously at the system level has become a trend in EV industrial applications in recent years. Multiple control strategies have been successfully developed for liquid cooling configurations, but the complexity of multisystem modeling remains a major concern in implementing MPC for thermal control. The reported literature tends to develop nonlinear MPC models or establish hierarchical structures. For example, Amini et al. [7], Tang et al. [8], Park and Ahn [9], and Zhao and Mi [10] have proposed various MPC strategies to regulate battery and cabin thermal control. These strategies have achieved higher efficiency and better regulation of thermal requirements based on driving cycle information. The discussed predictive strategy also has the potential to be integrated with the whole power users to achieve better energy performance.

Besides thermal dissipation, thermal preservation in extremely low temperatures has garnered interest. Traditional methods like resistance heaters and heat pumps are used to maintain temperatures, but there is an increasing interest in using PCMs as thermal reservoirs, in which the heat storage enables passive thermal preservation for short-time parking and even overnight parking with higher ambient temperatures [11,12].

Minimizing energy consumption through joint optimization of vehicle energy systems, including thermal systems, is essential for EVs. Besides driving motors, the air conditioning system for both the cabin and battery is the second highest power-consuming system for EVs, followed by entertainment and other auxiliary systems such as liquid pumps for BTMS. Dynamic load shifting is a potential solution to prevent high power demand overlap, which has two benefits: (i) lower power output density mitigates cell swelling and prolongs battery cycle life [13–15] and (ii) high power density generates more waste heat, which negatively affects thermal control and energy savings. Therefore, optimizing regenerative charging and discharging sequences in conjunction with thermal control systems is a reasonable approach.

Similar to the thermal control for multiple cooperated systems, besides the rule-based deterministic methods and offline optimization algorithms like Pontryagin’s minimum principle and dynamic programming, MPC is an effective online optimal control method for coordinating load shifting with battery and cabin thermal control. For example, hierarchical MPC has been used for eco-cooling of connected and autonomous vehicles [16], optimizing battery operations and discharging schedules [17], and for vehicle-level and electric powertrain-level optimizations simultaneously [18]. Learning-based algorithms, particularly reinforcement learning-based methods, have also been successfully implemented in plug-in hybrid EVs [19]. However, studies focusing on pure EVs are rare due to the complexity of nonlinear multisystem optimization [20,21]. In addition, applying these predictive algorithms in vehicles requires robust onboard computational resources for real-time processing.

In order to achieve optimal control in energy management algorithms, accurate velocity forecasting is essential. In particular, short-term local intersection prediction requires high accuracy to avoid potential instability among multiple systems. Individual vehicle velocity prediction has become increasingly important with the advent of big data and machine learning-based forecasting technologies [17,22].

Compared with network traffic forecasting that emphasizes more on traffic information in networks and aims to provide further insights for transportation management and policy-making, there are three major discrepancies between individual and network vehicle velocity prediction in general. (i) Individual vehicle velocity prediction uses floating speed trajectories as a data source rather than network traffic records. (ii) Velocity prediction for individual vehicles necessitates prediction horizons at seconds, which is much shorter than timescales in minutes or hours for network traffic prediction. (iii) For computing platforms, individual velocity forecasting has no alternative but to directly implement onboard computing tools due to real-time requirements, while network traffic forecasting can utilize local/cloud-based platforms with higher computing capability.

As discussed earlier, there are two broad categories of algorithms for individual traffic forecasting: stochastic and deterministic, with deterministic approaches being more suitable due to computing limitations. The hidden Markov model (HMM) chain algorithm is a widely used stochastic method for annotating sequential data with underlying hidden structures. Several modifications have been proposed to adopt HMM for vehicle velocity forecasting, such as using a fuzzy logistic model by Jing et al. [23] to estimate individual vehicle speed 8 s ahead, a self-learning multistep Markov chain model proposed by Zhou et al. [24] based on simulated data, and a second-order HMM model for dynamic forecasting model selection in segment-based vehicle velocity forecasting by Liu and Zhang [25]. However, studies have shown that deterministic approaches have higher prediction accuracy, especially with limited datasets. For example, Sun et al. [26] compared stochastic (HMM) and deterministic algorithms (support vector machine, radial basis function neural network, and back-propagation neural network) for short-term vehicle velocity forecasting, and the results showed that deterministic algorithms outperformed HMM in terms of prediction accuracy. Liu et al. [27] also validated similar results by evaluating long short-term memory network, auto-regressive integrated moving average, and HMM for 10 s ahead speed forecasting on a real urban driving dataset. By leveraging the strengths of both methods, it is practical to combine the deterministic and stochastic approaches. For example, Shen et al. [28] developed a long-term velocity forecasting method by integrating a transformer network with the Markov chain Monte Carlo algorithm, achieving enhanced performance compared to the particle filter and long short-term memory (LSTM) methods. Moreover, Lemieux and Ma [29] investigated a deep belief network with a stacked auto-encoder for highway speed forecasting.

It is worth noting that individual vehicle velocity forecasting can be potentially improved by considering surrounding network traffic situations such as image detection of front vehicles and reported traffic accidents [30,31]. However, the policy-making and construction of intelligent transportation systems are currently falling behind the trend of electrification and intellectualization. Vehicles have limited access to real-time traffic information given the cost of data communication, but approaching traffic information can be collected via existing vehicle installations such as radar and camera systems, and traffic light detection technologies could be used to further improve velocity forecasting. Unlike traditional approaches that treat vehicle velocity prediction as a standalone component, our study proposes a comprehensive framework that directly integrates velocity forecasting into the energy management systems of EVs. This integration enhances the coordination between vehicle-level and component-level (battery, air conditioning) frameworks, improving the responsiveness of EVs to dynamic driving conditions and contributing significantly to energy optimization.

Aiming to enhance the vehicle energy efficiency, in this article, a hybrid two-stage localized model selection framework is developed for short-term vehicle velocity forecasting. Real-time detected traffic light signals are incorporated to boost velocity forecasting accuracy and optimize power allocation at intersections. Compared with the nonlinear MPC approach, the real-time traffic light detection-based energy management method achieves similar level of energy efficiency, but requires significantly less computational resources. Furthermore, a novel sandwich-like cell-to-chassis battery cooling prototype is developed and validated using transient computational fluid dynamics (CFD) simulations, in which we exploit the potentials of using PCM to store heat under cold environments. Overall, this study seeks to develop alternative energy management methods with lower computational cost rather than predictive algorithms to further enhance vehicle energy efficiency. Our contributions are threefold: (i) propose a two-stage localized model selection-assisted vehicle velocity forecasting framework integrated with image-based traffic light detection as an indicator for intersection velocity forecasting; (ii) develop a liquid-based plate cooling structure with PCM as the thermal buffer and heat reservoir; and (iii) explore the feasibility of leveraging traffic light detection for real-time intersection energy management.

The remainder of the article is organized as follows. In Sec. 2, a short-term velocity forecasting framework is developed with the integration of traffic light detection. The vehicle dynamic model and battery electric-thermal model are established in Sec. 3. Comparative studies between the MPC based and traffic signal light detection-assisted real-time energy management are conducted in Sec. 4. Finally, Sec. 5 summarizes the concluding remarks and the potential future work.

From the perspective of vehicle energy management, compared to a discrete inconsecutive driving dataset with multiple participants, a repeated driving cycles on a fixed route with the same driver has huge advantages serving as a fundamental spatial-temporal basis for algorithm development, emulation, and validation. Though several repeated driving datasets have been investigated in the literature [27,32], there are few datasets publicly available at the current stage due to data privacy and project constraints. To this end, we have successfully generated a Dallas repeated driving (DRD) cycle dataset² with dozens of driving cycle tests performed on a fixed route in the Dallas area. Its main purpose is to simulate a typical commuting route for passenger vehicles, which consists of a 5-km expressway test and an urban local road test of 20 km, as illustrated in the dataset module in Fig. 1. According to the regional traffic retiming program [33], it is worth noting that a large portion of the traffic signals is timing controlled in a daily manner, meaning that a traffic light tends to follow the same patterns at the same time of a day.

Prior to forecasting, a piece-wise approach that divides the whole cycles into segments is implemented in this study, in which the intersections and casual locations with high probability of stopping are primarily identified based on the driving data rather than directly using labeled data from public maps. By analyzing the DRD cycle dataset, it is found that nearly a quarter of those intersections should not to be treated as regular intersection due to a relatively low stopping probability, while multiple locations tend to result in indirect hidden stops under heavy traffic conditions. Once the intersections or stops are located, the routes in between are defined as independent segments. An intersection segment consists of a deceleration, a waiting, and a reacceleration process, while a normal road segment refers to a continuous moving at a steady speed. The details of the DRD cycle dataset and the preprocessing can be found in our previous study [25].

Building on the segmentation approach, a forecasting framework is constructed with a two-stage structure to perform the short-term velocity forecasting. In Stage I, a forecasting submodel pool that consists of a collection of stochastic and deterministic methods are established based on their popularity, i.e., LSTM, artificial neural network (ANN), support vector regression (SVR), HMM trained with augmented data, and a similarity-based method. Their kernels, training algorithms, and hyperparameters are tuned after training and validation. This process utilizes 23 cycles (cycle 1–23) for testing and 5 cycles (cycle 24–28) for validation.

In stage II, the hybrid approach integrates the ensemble forecast from stage I with two different models: the offline probability-aided ensemble model and the online Markov chain model with dynamic model selection. This is done to mitigate potential fluctuations and uncertainties, thereby improving the accuracy of short-term velocity forecasting.

During the training process, eight scenarios are explored and three conditions for dynamic model selection are taken into account, followed by an ensemble step that is similar to the offline approach. A larger weight factor is assigned to the chain sequences with top models. It is also worth noting that due to the intrinsic localized characteristics, unique transition matrices are established for each segment rather than using a universal transition mechanism for the whole driving cycle. The biggest difference between these two hybrid models lies in the required model numbers for each segment, in which the universal transition mechanism needs a real-time online ranking among all the models, while the unique transition matrices can update the models offline based on the recorded velocity data. Stage II reuses the stage I validation dataset combining a collection of new data (cycle 29–31) for training, while the rest cycles (cycle 32–34) are used for testing.

Performance comparisons among the forecasting methods are presented in Table 1, where the mean absolute error (MAE) and root-mean-square error (RMSE) matrices are employed for evaluation.

It can be observed from Table 1 that the proposed hybrid models show promise for improving short-term forecasting. The Markov chain model-based method with online dynamic model selection tends to yield slightly better results than the offline probability method, but the latter is more likely to be embedded into real industrial applications, especially for a shorter lead time, due to model update frequency and onboard computational resources. In the following sections, the probability-based method will be employed.

Besides the hybrid model attempts, we also recognize that the velocity profiles at intersections can be generally classified into three groups, i.e., moving forward at a constant velocity, passing through with deceleration, and completely stop, as illustrated in the intersection velocity model in Fig. 1. As a popular data mining technique, extensive surveys have implied that forecasting can be significantly enhanced via accurate classification. A traditional physical model-based classification method is developed in this study to divide the velocity sequences with unequal length, which directly utilizes the predefined deceleration, reacceleration, and stop processes as a classification threshold. We have also tested the hierarchical classification algorithm integrated with dynamic time warping and obtained similar results. Here, the prediction window is set as 5 s to align with the intersection dynamics like traffic light settings, average speed, and traffic volumes at Dallas area. Another consideration is to cooperate with control interval of potential predictive algorithms. Using the base models to forecast the velocity 5 s ahead, an averaged improvement of $0.11m/s$ regarding MAE is observed for all the intersection segments compared to an original MAE of $1.80m/s$⁠. Possible reasons of achieving such a limited enhancement can be attributed to the dataset size and the sample size unbalance among varying groups, which also means that the limitations can be mitigated by enlarging the dataset. Owing to the energy management focus in this study, we only present the basic flow path and summarize the takeaways that may be beneficial to future in-depth studies. Substantial explanations can be found in our previous study [25].

As discussed earlier, intersection velocity classification is a promising effort to improve velocity forecasting. However, when it comes to real forecasting practice, it is extremely challenging to classify the unknown future velocity sequence by merely using the velocity data. For the majority of situations, we notice that the aforementioned three scenarios at intersections are strongly associated with the traffic light signals, as presented in intersection velocity model. Moreover, it is observed that a green light usually leads to a steady running with high velocity, while the vehicle tends to stop at red/yellow lights. For the deceleration and reacceleration scenario, it can either be a green light at heavy traffic conditions or an ending red light followed by a green light. Given this consideration, in this study, we aim to develop an image-based indication framework for velocity forecasting by detecting and identifying the extra traffic control light signals via object detection. Building on this, a cornerstone of our proposed energy management strategy is the innovative use of traffic light detection. This feature significantly refines our vehicle velocity forecasts, particularly at intersections, a critical point for energy consumption in urban driving. By integrating real-time traffic light data, our system offers more precise control over the EV’s energy resources, leading to marked improvements in efficiency and a reduction in unnecessary power usage.

Extensive studies have been conducted on traffic light detection by leveraging novel and more effective convolutional neural networks (CNNs) in the emerging field of autonomous driving, covering a broad range of deep learning structures such as the R-CNN family (fast R-CNN, faster R-CNN, and mask R-CNN), the YOLO family (v1-v5), the single shot detection family, and the Retina-net family [34]. Since the motivations of this study emphasize on the impact of traffic light detection to vehicle velocity forecasting, a one-stage YOLO-v2 network with pretrained network structures is directly employed here after modification. Compared with other networks, YOLO-V2 possesses an effective classification backbone with 19 convolutional layers and 5 max-pooling layers, which provides an accurate detecting precision while maintaining a high processing rate.

For this DRD cycle dataset, all the traffic lights are horizontally installed with the same light arrangements for different colors, making it possible to directly detect the traffic light and its corresponding colors. An approximate of 1100 images are extracted from the driving cycles for labeling. To prevent potential recognition errors during the cycle, we not only label a single group of lights but also combine and label two nearby groups of lights as a whole. As an outcome, a traffic light and its color are recognized only when both objects have positive feedbacks. Our methodology employs a YOLO-V2-based object detection model, which stands out for its compatibility with standard automotive camera systems. This choice strategically avoids the high costs associated with LiDAR technologies, aiming to democratize advanced energy management capabilities in EVs. The practicality of this approach not only reduces the implementation costs but also facilitates easier adoption across various EV platforms. The model received an averaged precision of 0.845 and an averaged recall of 0.567. Compared to the typical accuracy around 0.9 reported in the literature [35], this basic model here in this study still needs further improvements.

The possible reasons and potential approaches to further improve the traffic light detection accuracy are twofold: (i) We empirically labeled all the vague images for the forecasting-oriented purposes to identify traffic lights as earlier as possible, which brings in a large amount of misleading noise. (ii) The images in this study were taken in a 3X optical zoom by a household camera device, making the images in an undesirable low quality. However, we still confidently received some inspiring results in traffic light detection: the model is able to detect the traffic light with its correct color at an approximately 100 m away in a straight road with no slopes, which is also 5–6 s ahead the intersection at a constant cruising speed. It is worth noting that the confidence score threshold in this study is set as 0.38 as a tradeoff between detection accuracy and exploration. Although there may be potentially misidentifications, i.e., detect other objects as a single group of traffic light, the object detection model is still able to yield desirable outcomes with the implementation of labeling near groups of lights.

The forecasting differences between the probability-based hybrid method and traffic light detection-assisted method are compared in Fig. 3. By comparing the 5 s ahead forecasting, it is seen that the traffic light detection-assisted method results in a significant improvement from $1.56m/s$ to $0.78m/s$ regarding MAE for this specific intersection. The improvements mainly come from the deceleration process where the red light signal acts as an indicator to directly determine the ongoing stopping scenario, while there are still unavoidable forecasting delay in the reacceleration process. For other scenarios, only small enhancements are observed by adopting the traffic light detection-assisted method, leading to a limited overall improvement when it comes to whole driving cycle, i.e., an averaged $0.02m/s$ improvement for MAE. Note that there is no improvement or even worse outcome for the 10 s ahead forecasting using the detection-assisted approach, since most of the deceleration and reacceleration processes occur within or around 10 s. Overall, as a promising indicator, the image-based traffic light detection can be leveraged to improve the EV energy efficiency in a twofold manner: (i) traffic light detection tends to increase the forecasting accuracy for short-term velocity forecasting and (ii) it also has the potential acting as a mode trigger to activate or terminate the functions and operations of devices in advance in a predictive energy management strategy. Overall, the multiperiod short-term velocity results shown in Fig. 4 are sufficient for energy management.

There are various systems integrating and working as a whole in an EV from the perspective of electric, thermal, and energy control, including the main battery system, vehicle motor system, air conditioning system, battery thermal control system, and the cooling functions for varying components, i.e., AD–DC bidirectional inverter, DC–DC converter, and driving motor. The section attempts to develop control-oriented models for the aforementioned systems, aiming to provide a comprehensive overview for further energy management.

The regenerative energy may either be utilized directly for AC and BTMS or be recharged back to the battery pack. How and when to optimally distribute the regenerative energy needs prompt solutions, which is also one of the major contributions in this study.

Compared to conventional internal combustion engine (ICE) vehicles, the air conditioning system in EVs differs in multiple perspectives: (i) in winter conditions, a heat pump-based AC system for EVs has predominant advantages on energy efficiency toward conventional AC with positive temperature coefficient (PTC) heaters, while ICE can directly utilize the waste heat from the engine without any auxiliary devices or equipment; (ii) compared with ICE, there are several extra thermal loads for EVs, such as battery thermal control, motor cooling, and converter cooling; and (iii) the AC compressor in ICE is propelled by the engine, while it is powered by the battery pack for EVs. All these unique characteristics lead to a heat pump-based AC with more complicated structures and larger cooling/heating capacity for EVs, as illustrated in Fig. 5. Note that here are two types of heat sinks designed for different operating and environmental conditions separately, i.e., the AC-based chiller deals with intense driving and charging conditions under high temperatures, while the radiator usually works at low ambient temperatures. This study mainly focuses on the impacts of AC operations.

Multiple parameters are required to present a specific dynamic state of the battery pack system, such as the temperature of battery bricks [$Tb1$⁠, $Tb3$⁠, $Tb5$], the temperature of PCM sections [$Tpcm1$⁠, $Tpcm3$], and the temperature of the cooling plate [$Tplate$], as indicated in Fig. 7. Three parameters are regarded as the system inputs, including the battery inlet temperature $Tcool_btm_in$⁠, the mass flowrate $m˙cool$ of the coolant, and the volumetric heat generation rate $G˙$⁠. The system state updates by iteration following new system inputs, as illustrated in Fig. 8.

The system state at each time-step $Xk$ and the BTMS outlet coolant temperature $Tcool_btm_out$ are predicted directly through the neural network models $N1−1$ and $N1−2,$ respectively. These models incorporate inputs such as battery inlet temperature, coolant mass flowrate, and heat generation rate to accurately simulate dynamic responses.

Based on the 5–15 s ahead velocity forecasting for a daily commute route, the initial system parameters are set with an exterior temperature of 310.15 K and a battery SoC of 0.95. The cabin temperature is targeted at 294.15 K, while the battery control temperature is aimed at 313.15 K. The upper bound of the battery is set as 317.15 K, to allow for a large margin of thermal impacts considering the usage of PCM.

The simulation results of the real-time control and the MPC-based approach are presented in Figs. 9 and 10, respectively. The PCM remains in fluid status during the driving cycles, since the temperature has been well constrained around the target value. The results show that the final SoC decreased from 0.95 to 0.8818 with MPC, while the real-time control yields an SoC of 0.8806. Compared with the real-time control, only a limited improvement is obtained via MPC for this specific driving cycle, i.e., less than 2% regarding the energy efficiency calculated by the SoC. The results are also verified based on real velocity data rather than forecasting. It is known that the MPC strategy is less effective for a steady driving stage with cruising speed, but performs well for varying conditions. Given this consideration, four potential reasons may account for this observation: (i) Compared to the hybrid driving cycles that consist of the Urban Dynamometer Driving Schedule (UDDS), Worldwide Harmonized Light Vehicles Test Cycle (WLTC), and the Highway Fuel Economy Driving Schedule (HWFET) cycles, these Dallas driving cycles are considerably smoother, with fewer intersections and complete stops. (ii) The waste heat dissipation from the battery system relies on the AC system, and when combined with the cooling demand from the cabin, the AC system runs at a high load ratio. The AC load is also very close to the regenerative power, leaving limited space for load shifting. (iii) The control intervals for AC power and BTMS power allocations are set at 5 s, making the AC system less responsive to varying power-train demands, and so does the BTMS coolant pump with on–off control. (iv) Given the time required for data telemetry and velocity forecasting in real practice, the computational time for solving the MPC problem is capped at 4 s. The solutions are local optima rather than global optima for a nonlinear nonconvex problem due to the very limited computational time.

As discussed earlier, the MPC-based energy management has its drawbacks during steady driving conditions, especially when using nonlinear and nonconvex models for complex systems. It is also observed that the load shifting mainly occurs under changing conditions, i.e., the deceleration and reacceleration processes at intersections. Given these considerations, we aim to develop a real-time energy control framework to avoid the overlapping among peaks and to reuse the regenerative power instead of recharging back to the battery pack. For instance, if a deceleration is anticipated, the system is motivated to decrease the power demand of the AC system by defining a higher target temperature beforehand, allowing the AC system to be powered by regenerative energy during the deceleration. On the other hand, when a reacceleration is predicted, the system tends to lower down the manipulated temperature to prevent overlaps among different loads.

The image-based traffic light detection method developed in Sec. 2, in conjunction with real-time acceleration signals, is utilized as a mode indicator to update control parameter settings. At an intersection, the vehicle is able to activate the low-demand mode based on its location. Given the largest detection threshold of 100 m in our model, the distance for low-demand mode activation is set as 250 m, which is approximately 12–13 s prior to an intersection at a cruising speed. It is worth noting that accurate traffic light recognition can be achieved as far as 130–150 m away as reported in the literature [35], which is about 7–8 s ahead for a cruising scenario and more than 20 s ahead for a completely stop scenario.

As described in the pseudo-code of Algorithm 2, based on the detected traffic light signals, two major scenarios are predefined with a sequence of energy allocation actions: (i) For a green light signal, the vehicle switches back to the normal control mode and deactivates the low-demand mode. (ii) For a red or yellow light signal, the AC system prioritizes the use of regenerative power as much as possible during deceleration, instead of recharging the battery system. The system also determines an average AC power demand based on the estimated waiting time at the intersection. The AC system operates at a relatively higher power level to cool down the vehicle when a vehicle stops before the intersection, avoiding power demand overlap during reacceleration. The system switches back to the normal control mode upon leaving the intersection, as indicated in Fig. 11. This approach relies on the recognitions of traffic light signals, incorrect recognition may bring down the overall energy efficiency, but can not jeopardize the safety of battery system by setting up prioritized temperature thresholds. It is worth noting that the traffic light can act as an early termination signal, i.e., a green light detected in front of an intersection suggests switching back to normal control directly.

In this study, the AC energy consumed by the battery thermal system remains unchanged due to its limited total volume. Instead, the AC energy for the cabin is adjusted based on the aforementioned principle. The simulation results of the modified control are presented in Fig. 12. As compared to real-time energy distribution, the main differences are observed at intersections. As a trade-off, the final stage SoC is improved by approximately 2.5% to 0.8823, at the cost of introducing more fluctuations in the cabin temperature. In comparison to the MPC-based energy management, the traffic light detection-based energy management has very similar performance for a smooth driving cycle. Given its computational efficiency, it is anticipated that the traffic light detection-based method could be a potential alternative for vehicle energy management. Our simulations were conducted in matlab, utilizing validated models of vehicle dynamics, battery systems, and thermal management, which are based on real-world parameters. The traffic light detection algorithm and model predictive control strategy were evaluated against realistic urban commuting scenarios to ensure the results closely mimic actual driving conditions.

Despite the promising simulated results, it is important to note that these simulations do not include tests with real vehicles or hardware-in-the-loop simulations. Future work will focus on validating our approaches under actual operating conditions to confirm their effectiveness and practical applicability.

This study contributes to the development of efficient and accurate energy management solutions for electric vehicles. The proposed YOLO V2 framework for traffic light detection achieved high accuracy in detecting traffic lights, and the probability-based hybrid model showed promising results in short-term velocity forecasting, particularly for 5 s ahead. Moreover, the MPC-based energy management approach optimized the energy consumption of the battery thermal management system, while the traffic light detection-based approach improved the final stage SoC by approximately 2.5% to 0.8823, with computational efficiency. These findings highlight the effectiveness of the proposed methods in improving energy efficiency and reducing emissions in urban commuting routes. The results indicate that using traffic light detection for real-time control could be a viable alternative for managing energy consumption in urban commuting routes.

While the MPC-based approach was used to optimize the energy consumption of the battery thermal management system, it served primarily as a benchmark in this study. We recognize that NMPC applications in real-world scenarios may yield only local optima due to computational constraints. However, our innovative approach of integrating real-time traffic light detection for energy management is both practical and efficient. This method is currently underutilized in the automotive industry, demonstrating its potential for significant impact and further exploration.

Potential future work includes (i) enhancing the localized model selection and averaging framework by collecting more data and investigating reinforcement learning for model selection and (ii) identifying and detecting traffic signals through CNN-based image identification techniques to further improve the waiting time estimation.

There are no conflicts of interest.

The data and information that support the findings of this article are freely available.³