Predicting battery health and remaining useful life by integrating hybrid supervised machine learning for energy storage systems

The rapid global adoption of Battery Energy Storage Systems (BESS) has amplified the need for accurate battery health management. Predicting the State of Health (SoH) and Remaining Useful Life (RUL) is critical for ensuring operational safety, optimizing performance, and extending battery lifespan. A significant challenge in this field is the discrepancy between predictive models trained on pristine laboratory data and the noisy, stochastic nature of real-world battery signals, which can obscure subtle degradation trends and lead to unreliable predictions. Current approaches often treat signal denoising and predictive modelling as disconnected stages, which can result in suboptimal performance. This research addresses this gap by proposing and systematically evaluating an integrated hybrid framework designed to function robustly under a realistic denoising condition. To create a realistic testbed, this study utilized the benchmark NASA PCoE battery dataset (B0005, B0006, B0007) and injected controlled additive Gaussian noise specifically into the voltage signals to imitate real sensor errors. Then, this study systematically evaluated noise data in nine distinct hybrid prognostic models, where the signal processing techniques were applied independently to voltage and current signals. These models were created by combining three signal processing techniques Wavelet Transform (WT), Kalman Filter (KF), and Empirical Mode Decomposition (EMD) with three supervised machine learning algorithms: Support Vector Machine (SVM), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU). From all of this, we developed the combination of WT-SVM, WT-LSTM, WT-GRU, EMD-SVM, EMD-LSTM, EMD-GRU, KF-SVM, KF-LSTM, and KF-GRU. The performance of each combination was rigorously assessed under noisy conditions using a suite of metrics, including Root Mean Squared Error (RMSE), Regression, and the accuracy of RUL predictions. The comprehensive analysis unequivocally identified the WT-SVM combination as the superior methodology. This model demonstrated exceptional robustness and accuracy under noisy conditions across all batteries and noise levels, achieving the lowest overall prediction error. For the challenging B0007 battery near its end-of-life, the WT-SVM model yielded a remarkable RMSE of just 0.0022 and a perfect RUL prediction. The success of the WT-SVM framework establishes a new state-of-the-art performance benchmark and provides a clear, evidence-based methodology for developing highly reliable battery management systems for real-world applications.