Auxiliaries’ power and energy demand prediction of battery electric vehicles using system identification and deep learning

Abstract

The energy demand of the auxiliaries of battery electric vehicles can account for a significant share of the total energy demand of a trip and must be taken into account for the prediction of the vehicle's remaining driving range or the implementation of predictive driving functions. This paper investigates a method that uses system identification and neural networks with bidirectional long short-term memory layers to predict the power requirements of the auxiliaries depending on information that is known prior to the trip. By using a self-learning, data-driven approach as well as data that can be measured without additional instrumentation, a prediction is made possible without the need to design detailed physical models in advance. Additionally, a rule-based allocation of the training data based on environmental conditions is implemented, which serves to adapt individual models to different climatic modes of the thermal system. The potential of the method is demonstrated for three different systems showing a prediction accuracy of on average 3% to 8% in terms of energy, while the deviation of the predicted power consumption is on average about 500 watts. Due to the complete automation of the process, a further increase in prediction accuracy can be expected.