An improved SOC estimation method for lithium batteries based on dynamic graph convolutional neural network and bidirectional fusion Q-Learning framework

• This method leverages graph convolutional neural network (GCN) models with dynamic adjacency matrix and battery state attention mechanism to model battery state data, capturing the complex correlation and temporal features between battery states through node feature learning and convolutional layers. It effectively addresses the problem of inaccurate equivalent models in dynamic battery scenarios. • By introducing the Q-Learning framework with a multi-objective reward function (integrating SOC error and SOH decay), the SOC estimation problem is transformed into a reinforcement learning task. The agent selects actions based on the current state to estimate SOC, and optimizes via rewards based on the error between prediction and actual SOC, solving the problem of cumulative measurement error and considering battery lifespan simultaneously. • This method features online learning, adaptive updating, and dynamic parameter adjustment (e.g., GCN node weights adjusted by SOH and temperature, Q-learning rate adjusted by GCN feature variance). It can continuously learn and update model parameters from actual data, thereby adapting to the changes and evolution of battery temperature and aging, and providing more accurate and robust SOC estimation results. With the popularization of new energy vehicles and renewable energy, accurate estimation of battery SOC is crucial for vehicle performance and safety. Traditional methods have problems such as inaccurate models, accumulated errors, and susceptibility to temperature effects. This article proposes an improved method for the fusion of dynamic GCN and BF-QL: dynamic GCN optimizes node weights through temperature decay sensing adjacency matrix, while BF-QL integrates SOC error and SOH decay reward signal to optimize feature selection and learning rate. The experimental results of HPPC, BBDST, and DST operating conditions show that the RMSE of the algorithm is 0.35 %, 0.86 %, and 0.52 %, respectively, with a maximum reduction of 87.7 % compared to BP and 71.9 % compared to EKF, verifying its high accuracy and convergence.