Paper No 231 – Frost Detection with Neural Networks: Determining Necessary Sensors to Predict Optimal Defrost Initiation Time for Air Source Heat Pumps – 14th IEA Heat Pump Conference, Chicago, USA
Air Source Heat Pumps (ASHPs) are the most common heat pump type in Europe's residential buildings. To increase the energy efficiency of ASHPs, a main research field focuses on defrosting management. Currently, researchers showed that optimal defrosting initiation time (ODT) exists, which exhibits great potential to improve operational efficiency. However, ODT depends on multiple factors such as ASHP operation (e.g., compressor RPM) and ambient conditions (e.g., relative humidity). While mapping all correlations between ODT and all relevant factors can be accomplished with artificial neural networks (ANN), gaining sufficient test-bench data is time-consuming. When combining ANNs with reinforcement learning (RL) the data can be automatically generated on-site. A key aspect for the successful realization of RL is the determination of necessary sensors to detect frost under dynamic ASHP operation and varying ambient conditions. This work studies the applicability of different sensor sets to predict frost. Therefore, we use a heat pump model with valid frosting and defrosting behavior. The model is calibrated with test bench data. The results indicate that commonly available sensors in heat pumps are suitable for robust frost detection. Using only the ambient and evaporation temperature, the RL agent can separate frosting behavior from heat pump control and improves energy efficiency by up to 9.4 % compared to conventional time-controlled defrosting.