Soft measurement and prediction of refrigerant leakage based on SVR-LSTM

Refrigerant leakage is a frequently-occurring fault that may have serious influence on the performance and environment, and is quite hard to be identified especially at an early stage. The traditional ways mainly concentrated on pattern recognition or classification, which is valid only when the leakage reaching serious level. This study proposes an SVR-LSTM model for real-time measurement and future prediction of refrigerant leakage with support vector regression (SVR) for leakage (soft) measurements based on four characteristic temperatures carefully screened via correlation analysis, and long short-term memory networks (LSTM) for leakage prediction based on soft measurement. Periodic optimisation is proposed for the prediction model using the constantly acquired running data. Leakage experiments on a 409.5-ton screw chiller with a rated charge of 330 kg were carried out and employed for verification. The results imply that the SVR soft measurement obtains an outstanding performance with R2 for training and testing reaching 95.80% and 99.98%, respectively, and is applied as the basis for subsequent prediction research. The RMSEs of the leakage prediction by the LSTM model established on 7000 s data are 0.323, and 3.3 kg, respectively, for 10 and 60 min, illustrating a difficult pre-diction for longer time. The optimised prediction model achieves higher performance with over 60% reduction in RMSE for 10-min prediction if 25.3% extra data is added. For longer-time prediction up to 40 min, the model with 20 min extra data performs best so far, demonstrating an effective prediction period of nearly 1/3 of the training duration.