Hyper-heuristic optimized Bi-LSTM approach for accurate state-of-charge estimation of Ternary and LiFePO₄ batteries

The State of Charge (SOC) is an essential metric in battery management systems for electric vehicle applications. However, the nonlinear and time-varying characteristics of these systems pose significant challenges for achieving high estimation accuracy. Alternatively, data-driven models, such as feedforward neural networks (FFNNs) and Long Short-Term Memory (LSTM) networks, have proven effective for SOC estimation; however, they are associated with high computational costs and sensitivity to hyperparameter settings, often necessitating trial-and-error methods. Traditional optimization techniques are impeded by an extensive hyperparameter search space. To address these challenges, we propose a Bidirectional Long Short-Term Memory (BiLSTM)- based data-driven estimation method, in which hyperparameters are automatically tuned using the Bayesian Optimization algorithm. This approach employs hyperparameter optimization to identify the model parameters that minimize the predefined loss function for a given dataset. The efficacy of the proposed model was demonstrated using three types of battery datasets at 25 °C: NMC-18650, NCA-18650, and LFP-26650. The primary objective of this study was to minimize SOC estimation error by implementing hyperparameter optimization, aiming to achieve an error of less than 1%. Ultimately, using the NMC-18650, NCA-18650, and LFP-26650 datasets, the mean absolute errors (MAE) were 0.1603%, 0.3120%, and 0.4064%, and the root mean squared errors (RMSE) were 0.2128%, 0.4252%, and 0.5663%, respectively.