Velocity Prediction Based on Map Data for Optimal Control of Electrified Vehicles Using Recurrent Neural Networks (LSTM)

Abstract

In order to improve the efficiency of electrified vehicle drives, various predictive energy management strategies (driving strategies) have been developed. This article presents the extension of a generic prediction approach already proposed in a previous paper, which allows a robust forecasting of all traction torque-relevant variables for such strategies. The extension primarily includes the proper utilization of map data in the case of an a priori known route. Approaches from Artificial Intelligence (AI) have proven to be effective for such proposals. With regard to this, Recurrent Neural Networks (RNN) are to be preferred over Feed-Forward Neural Networks (FNN). First, preprocessing is described in detail including a wide overview of both calculating the relevant quantities from global navigation satellite system (GNSS) data in several steps and matching these with data from the chosen map provider. Next, an RNN including Long Short-Term Memory (LSTM) cells in an Encoder–Decoder configuration and a regular FNN are trained and applied. The models are used to forecast real driving profiles over different time horizons, both including and excluding map data in the model. Afterwards, a comparison is presented, including a quantitative and a qualitative analysis. The accuracy of the predictions is therefore assessed using Root Mean Square Error (RMSE) computations and analyses in the time domain. The results show a significant improvement in velocity prediction with LSTMs including map data.

Keywords

• artificial intelligence
• recurrent neural networks
• long short-term memory (LSTM)
• electrified powertrains
• model predictive control
• global navigation satellite system (GNSS)