A lithium battery SOC prediction method using TCN during charging

To address the problem of a sudden increase in State of Charge (SOC) prediction error during the constant current-constant voltage transition period caused by the fixed receptive field of Temporal Convolutional Networks (TCNs), which is difficult to adapt to the switching of charging stages, this paper proposes an improved TCN prediction method based on a dynamic dilation mechanism.This mechanism determines the transition between charging stages by monitoring the rate of change of current in real time and adaptively adjusts the dilation factor of each convolutional kernel, thereby dynamically expanding the receptive field of the model during stage switching and achieving accurate modeling of long- and short-term dependent features. The combination of gated causal convolution and multi-scale residual connections enhances the model's nonlinear representation capability. The output layer incorporates a sigmoid activation function and an L1 regularization term to strengthen the SOC monotonicity constraint and improve physical consistency.Experimental results show that the proposed method has a mean absolute error of 1.13 ± 0.15% during the transition period, an average root mean square error of 1.08 ± 0.13% over the entire process, and an average inference latency of 10.3 ms, balancing high accuracy and real-time performance, and is suitable for deployment in battery management systems under complex operating conditions.