State-dependent efficiency estimation in electric vehicles using an artificial neural network approach

Accurate modeling of powertrain efficiency is essential for optimizing energy management and range prediction in electric vehicles. This is particularly important under varying real-world driving conditions. To address the limitations of fixed efficiency assumptions in conventional models, this study proposes a hybrid approach combining experimental data with physics-based simulation. A feedforward artificial neural network (ANN) is trained to predict powertrain efficiency dynamically using real-world data collected from a prototype electric vehicle. The ANN utilizes four input variables—motor torque, motor speed, battery temperature, and state of charge—selected through a combined physical and experimental data-driven relevance analysis. The trained model is integrated into a longitudinal vehicle simulation framework, enabling dynamic efficiency estimation and energy consumption analysis. The validation was performed by comparing the ANN predictions against a separate set of experimental measurements. Compared to a baseline linear regression model, the ANN demonstrated a 95.2% lower mean squared error (MSE) and 80.4% lower mean absolute error (MAE) during efficiency interpolation, with a coefficient of determination ( R 2 ) of 0.995. Simulations were conducted on both long-haul and city drive cycles, validating the model’s adaptability in diverse scenarios. These results support its application in predictive energy control, route-specific planning, and on-board performance evaluation.