Towards Green AI: Energy-Efficient State of Health Estimation and Degradation Analysis of Commercial Lithium-Ion Batteries Based on Deep Learning and Filter Technique Approach

The redundant similar data points in the large datasets increase the dataset volume, storage, memory usage, training time, and computational resources, leading to significant inefficiency in the deep learning (DL) models. These inefficiencies reduce model performance and increase energy consumption. The existing DL-based approaches for state of health (SOH) estimation of lithium-ion batteries often face challenges such as high computational demands, low accuracy, and high energy consumption. These models require high energy for accurate results and contribute to higher electricity demand and carbon footprints. Therefore, this work proposes a novel filter technique (FT) based on the redundancy reduction approach. This approach improves the quality of the dataset, i.e., reduces dataset size, memory utilization, and energy consumption. This novel FT was incorporated with the gated recurrent unit (FT-GRU) model and tested on six different types of SOH datasets. The key objective of this approach is to enhance performance and lower energy consumption. A comparison examination was performed by proposed models configured with FT and traditional models to investigate the SOH estimation effectiveness and the model's abilities and limits. The FT-GRU model was significantly tested on laboratory and publicly available NASA dynamic datasets, including B1, B2, B05, B06, B07 and B18. The B1 and B2 datasets were collected from our laboratory at room temperature. The FT-GRU approach achieved percentage improvements, 75.96% of RMSE, 99.49% of MSE, 93.07% of MAE, and computational time and energy consumption ranges from 25% to 70.76% and 69.50% to 82.38%, respectively, in comparison to the other existing approaches.