Effect of parameterized activation functions on the performance of LSTM models for state of charge (SoC) estimation in EV batteries
State of charge (SoC) represents the remaining capacity of a battery relative to its fully charged state and is a critical indicator of electric vehicle (EV) performance and driving range. Accurate SoC estimation is essential for effective battery management systems. Conventional methods such as Coulomb counting, open-circuit voltage, and Kalman filtering often suffer from reduced accuracy under varying operating conditions. long short-term memory (LSTM) networks effectively capture the nonlinear and temporal characteristics of battery behavior; however, the impact of activation function parameters within LSTM cells has not been thoroughly investigated. In this work, the impact of parameterized activation functions on the performance of LSTM models for SoC estimation in lithium-ion electric vehicle batteries is investigated. By introducing a scaling factor λ to the standard log sigmoid and tan sigmoid activation functions, the gradient profiles during training are modified. The empirical results demonstrate that modulating the scaling parameter λ directly reshapes the internal gradient landscape of the recurrent cells, fundamentally enhancing sequence tracking precision. The proposed approach achieves its peak SoC estimation performance at an optimal scaling value of λ= 0.5, yielding a maximum coefficient of determination (R2) of 0.9103 and a significantly reduced tracking error compared to baseline unscaled configurations. Furthermore, the parameterized gating architecture accelerates training convergence velocity by up to 45% while maintaining strict stability boundaries within a range of. These key numerical findings prove the structural effectiveness and universality of the proposed framework, establishing it as a highly accurate, computationally efficient solution suitable for online deployment in real-time electric vehicle battery management systems.