Deep Learning-Based Battery Voltage Forecasting Using Current as Input: A Multi-Stage Approach for Time Series Prediction

This paper introduces a novel approach for forecasting the voltage of the battery by learning the deep neural networks, based on leveraging the concurrent battery current as an input variable. The methodology employs the Long Short-Term Memory model, named LSTM, which is known for its capability to handle long-range dependencies in sequential data. The approach begins with the data pre-processing, where an in-depth analysis of the battery's Equivalent Electric Circuit (EEC) parameters is performed. These parameters are crucial for accurately predicting the battery's voltage response under various conditions. Subsequently, the data undergoes normalization, feature importance determination, windowing, and is split into training and testing sets. The training subset is applied to develop the LSTM-based forecasting model. A key innovation of this approach is the use of a multi-stage forecasting technique, where the LSTM model is trained to predict one step at a time, and then feed the output back into the model for the next prediction. This approach allows the model to dynamically adjust its predictions based on the feedback received from previous steps, enabling more accurate and robust forecasting. The experimental framework evaluates the proposed multi-stage LSTM forecasting model against single-stage LSTM, and other neural networks including a dense layer model, as well as a dense layer model with activation function. The results demonstrate that the multi-stage LSTM approach outperforms other models in terms of Mean Absolute Error (MAE), indicating its superiority for forecasting the voltage of the battery.

I. Introduction

Battery voltage prediction is a critical aspect of battery management systems of different applications such as electric vehicles, and renewable energy systems. Accurate voltage forecasting can significantly improve battery performance, efficiency, and longevity. Traditional methods for time series forecasting are based on statistical techniques or linear models, which can not extract the complex dynamics of battery behavior, completely.