Predicting electric vehicle performance metrics using a convolution neural network-gated recurrent unit-attention based deep learning architecture

The indicators of electric vehicle performance such as state of charge (SOC), remaining useful life (RUL), and charge demand need to be accurately forecasted to ensure maximum energy control and battery life. The models used are usually not able to capture the spatial and temporal correlation of battery data and be robust to the presence of noisy measurements. In this study, we model a sequential attention-based deep learning structure with convolutional neural networks, gated recurrent units, and an attention mechanism that can ultimately understand the local features, temporal relationships, and dynamic significance of various features in sequential battery data. The hybrid architecture of this model allows it to extract local spatial features, long-term sequential dependencies and dynamically find the importance of the critical time steps. We also develop a hybrid loss that is an accumulation of Huber loss and Mean Squared Error, which is much more resilient to outliers and at the same time has high prediction accuracy. It is experimentally proven that the proposed model has R2 values of 0.9575, 0.9558, and 0.9199 on SOC, RUL, and charge demand, respectively, which are better than the current single-architecture methods.