Prediction of Energy Consumption of Supermarket Refrigeration Systems Using DMD and LSTM-NN Modelling

The energy consumption of the refrigeration systems can account to over 20% of global electric energy consumption. It is crucial to optimal design and control of this kind of energy intensive system to improve its energy efficiency and thereby also contribute to climate protection. By focusing on a class of supermarket refrigeration systems, this work investigates how to use the emerging data-driven modeling methods, i.e. the Dynamic Mode Decomposition (DMD) method and LSTM-NN method, to make a dynamic prediction of the energy consumption of these systems. The relevant data is produced from an industrial digital twin system. After analyzing and treating the original data, both the simple DMD modeling method and the sophisticated LSTM-NN method are applied to obtain a prediction model to forecast the compressors' energy consumption subject to different ambient air temperature conditions. It has been observed that a single DMD model trained cannot perform well even for limited range, due to the nonlinearity inherent in the concerned systems. Thereby, the LSTM-NN model is applied for better prediction performance. After enhancing the standard LSTM-NN into an observer-type of LSTM-NN architecture (i.e., introducing output feedback into the standard LSTM-NN), the LSTM-NN model exhibits a precise prediction performance even with smaller set of training data, which indicates a promising potential to extend this type of solution into real-life industrial applications.