Hybrid Fuzzy Long Short-Term Memory Framework-Based Accurate State-Of-Charge Estimation for Future-Oriented Battery Management in Vehicle-To-Grid System

Abstract

Accurate state-of-charge (SOC) estimation is critical for vehicle-to-grid (V2G) systems, where bidirectional power flow and dynamic operational transitions impose severe challenges on conventional estimation methods. While long short-term memory (LSTM) networks have shown promise in capturing temporal dependencies, their fixed activation functions limit adaptability under rapidly varying V2G conditions, particularly during charging/discharging mode transitions and irregular load profiles. This paper presents a fuzzy-modulated LSTM (FM-LSTM) framework that dynamically adjusts LSTM sigmoid activation functions in real-time using a fuzzy inference system driven by battery current and voltage measurements. Unlike black-box LSTM implementations, this approach integrates domain knowledge through linguistic rule formulation, enabling the network to adaptively recalibrate its temporal processing based on instantaneous operating conditions. Validated on standardized drive cycles—Urban dynamometer driving schedule (UDDS) and Highway fuel economy test cycle (HWFET)—and a comprehensive V2G dataset encompassing grid-charging, vehicle-discharging, and bidirectional transitions, the FM-LSTM demonstrates substantial performance improvements over conventional LSTM: 44.81% root mean squared error (RMSE) reduction for UDDS, 23.64% for composite V2G profiles, and 12.66% for HWFET. The proposed method achieves estimation accuracy consistently below 1.5% RMSE across all V2G test scenarios at nominal (25°C) operating conditions and remains below 3% RMSE in the evaluated temperature conditions between 0°C and 45°C, including testing at an unseen intermediate temperature, with the adaptive sigmoid modulation effectively addressing the nonlinear dynamics and operational mode transitions that challenge fixed-architecture networks.