Energy consumption prediction for water-based thermal energy storage systems using an attention-based TCN-LSTM model

Predicting energy consumption in energy storage systems is crucial for optimizing system operation and promoting building energy sustainability. Nevertheless, due to the periodicity, intermittency, and strong nonlinearity of energy consumption in storage systems, conventional deep learning models often fail to fully capture the characteristics of energy consumption fluctuations, leading to an inability to balance both single-step and multi-step prediction accuracy. To address this issue, this paper improves the conventional Long Short-Term Memory (LSTM) model by integrating Temporal Convolutional Networks (TCN) with LSTM to optimize 1-hour single-step predictions and incorporates an attention mechanism to enhance 8-hour and 24-hour multi-step predictions, thereby proposing an Attention-TCN-LSTM prediction model. Utilizing data from an operational water-based thermal energy storage system for model training and testing, the model was compared with three neural networks and two machine learning benchmark models across three prediction time scales. The results demonstrate that, compared to the five benchmark models, the Attention-TCN-LSTM model reduced the Root Mean Square Error (RMSE) by 17.74 %-34.26 %, the Mean Absolute Error (MAE) by 19.68 %-38.60 %, the Mean Absolute Percentage Error (MAPE) by 0.12 %-2.65 %, and the Coefficient of Variation of RMSE (CVRMSE) by 9.65 %-17.27 %, while increasing the R-squared (R²) value by 3.54 %-9.36 %. These results highlight the model's higher prediction accuracy and stability, as well as shorter training times compared to conventional neural networks, underscoring its potential to enhance energy prediction and system performance.