<p>State of charge (SOC) is the key parameter in battery management systems (BMS). To improve the accuracy of SOC estimation for lithium-ion batteries under complex temperature conditions and to address issues in traditional data-driven models such as the difficulty of manually optimizing hyperparameters and significant result variability, this study proposed an HDFO-CBLA-KF hybrid neural network framework for estimating the SOC in lithium-ion batteries. The framework employs a convolutional neural network (CNN) to extract local time-domain features from externally measurable electrical data. Subsequently, the bidirectional long short-term memory (BiLSTM) neural network with attention mechanisms is constructed to enhance sequence modeling capabilities and focus on key information. Then, the Kalman filter (KF) is introduced at the end of the framework to perform posterior correction and noise suppression on the SOC estimation results, thereby achieving the integration of data-driven estimation and filter correction. Finally, at the network hyperparameter optimization level, an advanced collaborative search framework combining Harris hawks optimization (HHO) and differential evolution (DE) is proposed, achieving adaptive exploration and development throughout the entire process and eliminating the need for tedious manual parameter tuning. Results under various operating conditions at different temperatures indicate that the proposed model demonstrates good SOC estimation accuracy and stability across a range of temperatures and operating conditions. Specifically, the maximum estimation error is kept within 5% under positive temperature conditions and within 10% at − 15&#xa0;°C. The model exhibits favorable adaptability to continuous dynamic loads and can serve as a reference for SOC estimation in dynamic electric vehicle operation scenarios, as well as for BMS state management.</p>