Remaining useful life prediction for lithium-ion batteries incorporating spatio-temporal information

Accurately predicting the remaining useful life (RUL) of lithium-ion batteries (LIBs) is crucial in the field of industrial intelligence. However, existing generalized models face challenges in battery life prediction due to the presence of prevalent noise and limited degradation data. To address this issue, we proposed a spatiotemporally integrated RUL prediction model, namely, the DAE-MSCNN-LSTM model, which leverages the combination of multiscale convolutional neural network (MSCNN) and long-short-term memory network (LSTM). This approach can be used to effectively extract feature information from limited available data. Moreover, the model incorporates a denoising autoencoder (DAE) to refine the raw data by mitigating noise and handling outliers. Subsequently, the processed data are fed into parallel MSCNN and LSTM networks, allowing us to capture both spatial and temporal information. The fused features from these networks are subsequently input into a multilayer perceptron (MLP) for RUL prediction. Additionally, we introduced a unified learning framework to address data denoising and model prediction simultaneously. Finally, the optimal hyperparameters were determined using the grid search algorithm. Through extensive experiments on the NASA and CALCE datasets, the superiority of our proposed LIB RUL prediction for model is demonstrated.