Predicting state of charge in latent thermal energy storage systems using artificial intelligence modelling techniques

Accurate state of charge (SoC) estimation in latent thermal energy storage (latent TES) systems is critical for optimizing renewable energy integration and ensuring efficient energy storage operation. Traditional physics-based methods, though accurate, were computationally intensive and required complex parameter calibration, limiting their applicability for real-time applications. This study introduced an artificial intelligence (AI) framework for real-time SoC estimation in shell-and-tube latent TES systems, integrating long short-term memory recurrent neural networks (LSTM-RNN) and extreme gradient boosting (XGBoost) through an optimised weighted ensemble (0.3 LSTM-RNN, 0.7 XGBoost), with weights determined using grid search. The framework was trained and validated on 605 charging and discharging profiles from a previously experimentally validated model, significantly expanding prior studies. The dataset covered a broad range of operating conditions, including charging (140–150 °C) and discharging (100–110 °C) cycles with mass flow rates between 0.8 and 1.2 kg/s. The ensemble model achieved superior accuracy (RMSE: 1.7%, MAE: 0.9%, R 2 >99.7% for charging; RMSE: 1.4%, MAE: 0.8%, R 2 >99.8% for discharging) compared to individual LSTM-RNN and XGBoost models, while reducing computational costs by 80%, enabling inference within 2.6 milliseconds per sample. This significant reduction in computational requirements transforms SoC estimation from a computational bottleneck into an enabling technology for real-time control integration, allowing direct deployment in industrial control systems where traditional physics-based methods remain impractical due to their computational demands.