Lithium-Ion Battery Remaining Useful Life Prediction Based on MCNN-BiLSTM

Precisely estimating the Remaining Useful Life (RUL) of lithium-ion batteries is crucial for maintaining the reliability and safety of energy storage systems, especially in new energy vehicles. This paper proposes a hybrid deep learning model, MCNN-BiLSTM-MAF, combining a MultiScale Convolutional Neural Network (MCNN), Bidirectional Long Short-Term Memory (BiLSTM), and a Moving Average Filter (MAF) to enhance RUL prediction. The model first extracts health indicators from the CALCE public dataset based on features closely correlated with battery degradation. MCNN captures multi-scale local patterns, while BiLSTM effectively models temporal dependencies in both directions. To further improve stability, MAF is applied to smooth the prediction results. Experimental evaluations show that the proposed model achieves superior performance with an average RMSE below 0.014 and MAPE below 0.013, outperforming conventional LSTM-based methods. This demonstrates the model’s effectiveness in delivering accurate and stable RUL predictions for lithium-ion batteries.