Data driven fault detection and diagnostics for hydronic and monitoring systems in a residential building

Buildings are responsible for 40% of the global energy use and associated with up to 30% of the total CO2 emissions. The drive to reduce the environmental impact of the built environment was the catalyst to the increasing installation of meters and sensors to monitor the energy use and environmental monitoring. This is key to cost effective Fault Detection and Diagnostics (FDD) which guarantees enhanced thermal comfort for the occupants and reduction in energy use. Most of FDD research work in buildings was focused on the commercial buildings due to the higher consumption and the higher saving potential, while limited work was directed towards residential buildings. This paper investigates the usage of two supervised machine learning algorithms, namely Random Forest, K nearest neighbour, to detect and diagnose twelve faults in both the monitoring system of the indoor/outdoor conditions, and the hydronic circuit inside an apartment located in Milan using minimal features that are easy to access and inexpensive to monitor to cut down in both computational and financial costs. The thermal zones are being conditioned using an electric air to water heat pump connected to fan coils for cooling and radiant floor for heating. The faults include valve leakage, faulty temperature sensors and recirculating pump’s inadequate flow rate. The faults were modelled in a Modelica based detailed model of the apartment. After tuning the hyper-parameters of all three algorithms, the Receiver Operator Characteristics curve for each fault were compared for each algorithm to compare the optimal one to be used. The Random Forest algorithms showed the highest accuracy with almost 89% across the twelve faults. Generalization of the trained algorithm across different weathers were tested but the results were not promising.