Contributions to the determination of optimized driving strategies for electric vehicles using artificial intelligence based methods

Abstract

This thesis presents a hybrid and modular framework for optimizing energy-efficient speed profiles in battery electric vehicles (BEVs), particularly under realistic road conditions and travel time constraints. With the global acceleration of electric vehicle adoption, the demand for intelligent driving strategies that minimize energy consumption while remaining compliant with road dynamics and timing requirements becomes increasingly critical. The proposed methodology integrates Simulink-based vehicle simulation, machine learning-based surrogate modeling, and metaheuristic optimization into a unified and flexible architecture capable of adapting to various road geometries and driving scenarios. At the core of the framework lies a detailed MATLAB/Simulink model representing the electric powertrain of a BEV, including a battery model, a DC-DC buck converter, an inverter, a brushless DC (BLDC) motor with sector-based commutation logic, and a longitudinal vehicle dynamics module. This model, designed to reflect energy consumption, speed tracking, and component-level electrical and dynamic response with high fidelity, serves as the reference model for both training data generation and validation. Given any candidate speed profile, the model outputs energy consumption and trip duration, enabling reliable evaluation of system performance. To construct a robust training dataset, thousands of randomized yet physically plausible speed profiles were simulated on a composite road containing flat, uphill, and downhill segments. For each simulation, key outputs such as total energy usage and travel time were recorded. These data points were used to train surrogate models based on Gaussian Process Regression (GPR), enabling fast and accurate prediction of energy and time outcomes without relying on time-consuming simulations during the optimization phase. The Particle Swarm Optimization (PSO) algorithm was employed to identify the optimal speed profile that minimizes energy consumption while satisfying a strict 300-second travel time constraint. A tailored cost function was used to penalize deviations from this target, guiding the search process effectively. The optimized profile, as estimated by the surrogate model, showed a small energy difference of approximately 0.4% when compared to the simulation-based validation result, confirming the accuracy and reliability of the surrogate modeling approach. Compared to a constant-speed baseline scenario covering the same 700-meter route within the same time constraint, the optimized profile achieved approximately 12% lower energy consumption. This result underscores the advantage of terrain-aware, adaptively optimized driving strategies even when the overall trip time remains unchanged. A notable strength of the proposed framework lies in its modular and extensible structure. Users can modify road definitions, adjust segmentation resolution, or substitute the GPR surrogate with other machine learning models. Similarly, the PSO algorithm can be replaced by alternatives such as Genetic Algorithms or Bayesian Optimization. Moreover, the framework supports the integration of user-defined constraints such as local speed limits, zero-speed targets for certain segments, or bounded acceleration profiles directly into the cost function, enhancing its applicability for real-world deployment. Although the current implementation focuses solely on longitudinal dynamics, the overall architecture is open to future extensions, including the incorporation of lateral vehicle behavior and regenerative braking systems. These additions could be integrated into both the simulation model and the surrogate modeling layer, further enhancing the model's predictive accuracy and its ability to represent realistic driving behavior. The framework can also be configured with real-world vehicle parameters, enabling field testing or hardware-in-the-loop validation. In conclusion, this thesis introduces a robust, scalable, and data-driven approach for optimizing electric vehicle energy consumption within defined travel time limits. By combining simulation, machine learning, and optimization in a modular architecture, the system delivers accurate, adaptive, and efficient speed profiles. The demonstrated energy savings, validated against simulation results, reflect the practical relevance and technical robustness of the proposed method, marking a significant contribution toward advanced energy management in electric mobility.