Artificial neural network-based efficiency estimation and modeling approach in electric vehicle powertrains

Abstract

The accelerating transition toward sustainable passenger transportation and road logistics has been marked by a rapid increase in the adoption of electric vehicles across the globe. While passenger cars initially proved to be dominant in this trend, recent years have witnessed the electrification of commercial vehicles for both light and heavy duty operations, as well as electric buses and public transportation fleets. This trend reflects a broader transformation in mobility. The transition into electric drive is mostly motivated by environmental concerns, newer regulations, and the growing awareness towards energy efficiency and sustainability. On the other hand, the widespread deployment of electric road vehicles also presents new technical challenges where concerns related to energy efficiency and energy management are among the foremost. This setback underscores the importance of modeling efforts focused on energy efficiency, as they play a crucial role in optimizing the performance of electric road vehicles. The ability to model efficiency under real-world conditions directly influences fundamental concepts such as vehicle energy management, route optimization, and range estimation which are essential to build user confidence and maximize the benefits of electrification. Traditional approaches for efficiency modeling typically rely heavily on static efficiency maps or model based simulations. While these methods provide useful insights, they often tend to fall short when applied to dynamic operating conditions where different system parameters become variables. For example, efficiency maps are typically constructed from experimental measurements under steady-state conditions. However, this approach may not fully capture the complexities of transient driving or environmental variability. Similarly, a physics-based approach can easily become computationally expensive and is prone to lacking the flexibility to adapt to diverse vehicle configurations or battery chemistries. Moreover, their high computational demand and reliance on detailed parameterization make them less suitable for real-time implementation in embedded vehicle systems, since processing power and memory are limited. As a result, there is a need for more adaptable, data-driven methods that can provide reliable predictions in real time in order to support both research and practical vehicle control systems. This thesis addresses these challenges by developing a machine learning approach for powertrain efficiency prediction based on artificial neural networks (ANNs). The main idea is to leverage the ability of ANNs to capture nonlinear relationships among system variables in order to provide accurate efficiency estimation across a wide range of conditions. The model is trained on key input parameters that were systematically identified and evaluated for their contribution to EV efficiency, namely torque, rotational speed, battery temperature, and state of charge (SoC). These variables reflect the electromechanical loading of the motor and the electrochemical behavior of the battery and provide a comprehensive description of the factors most critical to efficiency. These state variables are normalized in order to ensure consistent scaling throughout the analysis and allow the network to learn effectively from diverse operating conditions without bias toward variables of larger magnitude. The normalized dataset is then used to train a feedforward ANN architecture, which is designed with a single hidden layer of neurons and activation functions chosen to strike a balance between model complexity and generalization capability. The training process is carried out using supervised learning with experimental and simulated datasets. The network iteratively adjusts its weights to minimize the error between predicted and actual efficiency values. The data is divided into training, validation, and test sets in order to ensure robustness, which also allows the assessment of generalization performance and reduces the risk of overfitting. Once trained, the ANN is integrated into a longitudinal vehicle model where it operates as a real-time efficiency estimator. In this framework, the trained ANN replaces traditional static maps by dynamically estimating efficiency based on real-time operating conditions. This integration of ANN allows for the evaluation of vehicle energy consumption in various drive cycles, which include both urban stop-and-go traffic and long-haul highway scenarios. This approach not only provides more realistic estimates of vehicle range and consumption but also makes it possible to comparatively analyse different routes, driving styles, and operating environments. The evaluation of the proposed ANN within the longitudinal vehicle model demonstrated clear improvements over the baseline linear fitting approach. The ANN reduced the mean squared prediction error of efficiency by approximately 95\% and enhanced the correlation coefficient by approximately 5\% across the operating points tested, and maintained stable accuracy under both urban and long-haul drive cycles. In particular, the model captured transient variations in speed and load that the linear method failed to represent, which enables applying more reliable energy analysis. These results confirm that the ANN framework not only improves accuracy, but also provides the computational efficiency required for practical deployment. The ANN shows a markedly improved ability to capture nonlinear relationships among torque, speed, SoC, and temperature when compared against a baseline linear fitting approach. This allows the model to adapt more effectively to variations in operating conditions such as speed fluctuations, load changes, and battery states. This performance demonstrates the suitability of the method for embedded vehicle control systems, enabling rapid and accurate energy assessment to support advanced functions such as predictive energy management and eco-routing. Overall, the contributions of this thesis lie in advancing the application of data-driven techniques to the modeling of EV efficiency. The study provides a scalable framework that can be adapted to different vehicle types and technologies by combining machine learning methods with system-level vehicle simulations. While the present study establishes a solid foundation for data-driven efficiency modeling, several aspects for future research remain. Expanding the input vector to include environmental variables such as road gradient, weather conditions, or traffic level could further improve robustness under diverse conditions. Similarly, extending the framework to multi-objective applications, such as joint optimization of efficiency, battery health, and thermal management, would enhance its relevance for next-generation electric vehicle energy management systems. In summary, this thesis demonstrates the potential of machine learning applications to enhance the energy management and adaptability of electric vehicle powertrains. The proposed ANN-based framework effectively captures complex nonlinear relationships among system variables, enabling more accurate predictions on vehicle performance under diverse driving conditions. The results highlight the advantages of data-driven approaches over simpler linear methods, offering improved adaptability, computational efficiency, and suitability for real-time implementation. Overall, this work establishes a scalable and robust methodology that can inform both research and practical strategies to improve efficiency, reliability, and sustainability in electrified transportation systems.