A Lightweight Informer-LSTM Hybrid Model for Real-Time Lithium-Ion Battery State of Charge Estimation on Edge Devices

Accurate state of charge (SOC) estimation is vital for ensuring the safety, efficiency, and longevity of lithium-ion batteries, particularly in electric vehicles and distributed energy storage systems. However, the nonlinear and time-varying characteristics of batteries across varying operating conditions greatly hinder the robustness and accuracy of conventional estimation approaches. To address these challenges in a resource-constrained edge environment, this paper proposes a lightweight and SOC prediction approach based on a hybrid architecture combining Informer and long short-term memory (LSTM) networks. The proposed Informer-LSTM model leverages the Informer’s efficient self-attention mechanism to capture long-range temporal dependencies and global patterns, while the LSTM component focuses on modeling short-term dynamics and local nonlinear transitions. For the purpose of edge deployment, the model is optimized for real-time inference on embedded platforms such as the NVIDIA Jetson Nano. Experimental evaluations on battery data from real-world usage demonstrate that the proposed approach not only outperforms existing methods in prediction accuracy and generalization capability but also satisfies the latency and computational constraints of edge devices. The lower inference overhead and higher throughput confirm the model’s suitability for battery management systems in edge computing scenarios.