Chiller systems are identified as a major source of energy consumption in commercial buildings and industrial sectors, necessitating the development of sophisticated prediction models for energy savings and efficient operation. This study proposes an ...
Chiller systems are identified as a major source of energy consumption in commercial buildings and industrial sectors, necessitating the development of sophisticated prediction models for energy savings and efficient operation. This study proposes an integrated prediction model that combines corrected performance curves from an existing physical model with a Long Short-Term Memory (LSTM)-based AI model, achieving high prediction accuracy even in data-limited scenarios. The aim is to enhance predictive accuracy by leveraging the thermodynamic reliability of the physical model and the nonlinear data-learning capabilities of AI.
In this study, the physical model was improved by optimizing key parameters of cooling towers and chillers and correcting performance curves, thus addressing limitations of the existing model. Additionally, the LSTM-based AI model was employed to learn the discrepancies between the physical model predictions and actual measurements. By utilizing time-series data, the AI model effectively captured nonlinear characteristics and complex dynamic interactions, significantly enhancing prediction accuracy.
The performance comparison revealed that the physical model with corrected performance curves (Case 2) showed improved predictive accuracy compared to the original physical model (Case 1). Case 2 recorded Mean Bias Error (MBE), Root Mean Square Error (RMSE), and Coefficient of Variation of RMSE (CVRMSE) values of 2.82%, 11.93 kW, and 4.14%, respectively, indicating a reduction in prediction error. Furthermore, Case 3, which combined the corrected physical model with the LSTM-based AI model, demonstrated even greater improvements in prediction performance. It achieved MBE, RMSE, and CVRMSE values of 0.28%, 6.43 kW, and 2.87%, respectively, marking the best performance among all cases and significantly outperforming the previous models.
In conclusion, this study confirms that the integration of performance curve correction and AI models can enhance predictive accuracy for complex systems, even in data-limited environments. This approach demonstrates the potential to effectively combine the reliability of physical models with the learning capabilities of AI, offering practical contributions to diverse applications aimed at improving energy efficiency and achieving sustainable system operations.