The state of charge (SOC) estimation of lithium battery exhibits nonlinear and time-varying characteristics. Traditional BP neural networks for SOC estimation suffer from slow convergence speed and reduced accuracy. A lithium battery SOC estimation method was proposed based on the TLPSO-BP neural network algorithm. The method first enhanced the particle swarm optimization (PSO) algorithm by Tent chaotic mapping and Levy flight strategy, thereby improving exploration and convergence behavior. These enhancements were then applied to optimize the initial weights and thresholds of the traditional BP neural network, which increases the accuracy and stability of SOC estimation. Finally, the proposed algorithm was applied to SOC estimation of lithium battery under aggressive driving conditions (US06) and compared with PSOBP and BP algorithms. The experimental results indicate that the root mean square error of SOC estimation using the TLPSOBP algorithm is reduced by 14.5% and 32.4% compared to the PSO-BP and BP algorithms, respectively. These improvements verify the good generalization capability and high estimation accuracy of the algorithm.