Real-time prediction of battery remaining useful life using hybrid-fusion deep neural networks

Accurate prediction of battery remaining useful life (RUL) is crucial for their reliable and efficient use. Traditional methods struggle with the nonlinear and complex nature of battery aging, which is compounded by high computational demands, lengthy model development times, and issues such as data inconsistencies and noise. To overcome these, this study presents a hybrid neural network model integrating a 2D convolutional neural network (2D-CNN) and self-attention mechanism. The 2D-CNN extracts local features from interpolated voltage and capacity data, crucial for addressing diverse battery usage patterns, while the attention mechanism enhances prediction accuracy and efficiency by focusing on key features across regions and time. Experimental evaluations were conducted using two extensive datasets comprising over 240,000 cycles from 201 LFP batteries, as well as 41 NMC and NCA ternary batteries, the model utilizes data segments (80 % SOC to 3.6 V) and a sliding window technique to reduce data volume and computational overhead, achieving training convergence in 18 min and inference speeds of 4.86 ms per battery across the entire lifecycle. The proposed approach attained root mean square errors (RMSE) of 109 and 65 cycles on the two LFP datasets, respectively, and RMSEs of 26 and 12 cycles on the NMC and NCA datasets, respectively, demonstrating its robustness and adaptability across different battery chemistries. This development offers a dependable, real-time predictive tool for battery diagnostics and management, effectively bridging significant gaps in RUL prediction.