Detection of the low ∆T syndrome using machine learning models

Abstract

The low ∆T syndrome has been a prevalent issue in many chilled water systems, leading to an increase in the pump energy consumption, increase in the chiller energy consumption, and/or failure to meet the cooling loads. It is therefore important to detect the low ∆T syndrome using suitable fault detection and diagnosis methods. One such fault detection method is the data-based approach using machine learning algorithms. The main signs indicating the low ∆T syndrome include a reduced return water temperature from the cooling coil and an increased mass flow rate through the cooling coil. Since the mass flow rate of water is not measured in all chilled water installations, the cooling coil valve position is measured instead. This research aims to compare the performance of different machine learning regression models which predict the return water temperature and the cooling coil valve position, based on the R2 score and root mean square error. The different machine learning algorithms compared for the study include Support Vector Regression, Artificial Neural Network and eXtreme Gradient Boosting. The data required for the analysis was obtained from fault-introduced experiments conducted in an office building. The different fault cases include stuck cooling coil valve at 50%, stuck cooling coil valve at 75%, reduced supply air temperature by 2K and reduced supply air temperature by 1K. The regression models are expected to predict the fault-free data (Xpredicted) of the system such that faulty data (Xactual) can be identified with residuals (Xpredicted – Xactual). The results showed that XGBoost was the best performing algorithm in terms of model accuracy. The XGBoost based prediction models for return water temperature and cooling coil valve position were able to successfully detect anomalies for 3 out of the 4 fault cases.