Robust state of charge estimation for lithium-ion batteries using a fully entangled temporal convolutional network with particle swarm optimization

Abstract

Lithium-ion batteries are central to modern transportation, particularly electric vehicles (EVs). Accurate state of charge (SOC) estimation is vital for safety, efficiency, and extended battery life. While many methods exist, few address the challenges of SOC estimation at low temperatures, where battery behavior becomes highly nonlinear and measurement noise increases. To overcome this, we propose a fully entangled temporal convolutional network (FE-TCN) for SOC estimation across a wide temperature range, including low-temperature conditions. The model integrates the sequence learning ability of temporal convolutional networks with a quantum-inspired entanglement mechanism. Four parallel data flow paths, connected through Kronecker-product-based skip connections, enable effective information exchange and capture complex feature interactions. To ensure stable training under temperature-induced noise, the log-cosh loss function is employed, while particle swarm optimization (PSO) is used to optimize hyperparameters such as block depth, kernel size, and learning rate. Experimental validation demonstrates that the FE-TCN achieves high accuracy across -20 °C to 25 °C and under four driving cycles, with minimum mean absolute error and root mean squared error of 0.14% and 0.21%, respectively. Furthermore, the model maintains stable performance under different initial SOC conditions, demonstrating strong robustness.