State of Charge and State of Energy Estimation for Semi-Solid Lithium-Ion Batteries Based on Multi-Scale Temporal Convolutional Network

Against the backdrop of the energy crisis and environmental pollution, semi-solid lithium-ion batteries have been widely applied in energy storage and electric vehicles. Accurate joint estimation of state of charge (SOC) and state of energy (SOE) is key for the safe and efficient operation of battery management systems (BMSs). To solve the problems of existing methods, such as open-loop error accumulation, strong dependence on mechanistic models, insufficient multi-scale temporal feature extraction and poor adaptability of traditional attention mechanisms, this paper proposes a hybrid deep network model named CNN-PMSTCN-BASE-GRU. The model uses a one-dimensional convolutional neural network (1D CNN) to extract preliminary features, and a Parallel Multi-Scale Temporal Convolutional Network (PMSTCN) to capture both short-term transient characteristics and long-term evolution rules of battery data. The Branch-Aware-Squeeze-and-Excitation (BASE) attention mechanism is introduced to realize adaptive weighting of multi-branch features, and dual-path gated recurrent unit (GRU) is adopted for collaborative SOC and SOE estimation. An unscented Kalman filter is further utilized to suppress noise and smooth the output. The experiment was conducted under different temperatures and driving cycle conditions. The results show that the proposed model reduces the average mean absolute error of SOC estimation by 30.32% compared to the baseline model under variable temperature conditions, providing an effective solution for accurate multi-state estimation of lithium-ion batteries. Cross-material validation was carried out based on the publicly available INR 18650-20R Battery dataset from the University of Maryland, and it demonstrated excellent cross-material performance.