Battery SOH Estimation Using LSTM with Genetic Algorithm-Based Hyperparameter Optimization

The effective operation and management of battery-based energy storage devices hinge upon timely assessment of battery health. This paper introduces an optimized deep learning approach utilizing Artificial Neural Networks (ANN) for assessing the State of Health (SoH) of lithium-ion batteries, leveraging on battery aging datasets from NASA's Open Data Portal. An innovative methodology for optimizing Long Short-Term Memory (LSTM) hyperparameters using Genetic Algorithm (GA) is also proposed. Traditional approaches rely heavily on trial and error for hyperparameter determination, facing challenges in selecting optimal parameters due to an expansive search space. The proposed systematic methodology not only identifies the best architecture but also introduces novel criteria for evaluating ANN performance and complexity. The novel optimized GA-LSTM model is trained using the curated NASA data, and its accuracy evaluated by fine-tuning its parameters such as batch size, unit size, activation and loss functions. The results demonstrate that the proposed optimized GA-LSTM model exhibits better performance in estimating battery State of Health (SoH) across diverse conditions, achieving significantly improved Root Mean Square Error (RMSE) values compared to the Particle Swarm Optimization algorithm aided LSTM model, namely PA-LSTM, introduced in recent research. The improvement in RMSE ranges from 12.4% to 76.79% when trained with 70% of discharge cycle data. This represents a notable enhancement over previous methods, highlighting its efficacy as a more accurate prediction model for battery health assessment.