Machine learning-based online adaptive prediction for electric vehicle energy consumption

Precisely forecasting the energy consumption of electric vehicles not only alleviates the anxiety associated with driving range but also serves as the foundation for progressive advancements, including optimizing charging strategy and energy utilization. The main challenge lies in the inaccuracy of current methods, whether they are empirical models, physics-based models, or data-driven models. Based on newly constructed and engineered physics-informed features, this paper introduces a machine learning-based prediction framework, employing a synergy of offline global models and vehicle-based online adaptation. This combination aims to elevate accuracy in point predictions and also provide valuable information on prediction uncertainties. The developed framework is trained and extensively tested using data from a fleet of real-world electric vehicles. The leading global model, quantile regression neural network (QRNN), demonstrates an average error of 6.30%. Subsequent online adaptation results in a notable reduction to 5.04%, with both surpassing the performance of existing models significantly. Concurrently, the online QRNN exhibits a strong capability in enhancing the coverage probability and decreasing the average width of prediction intervals.