Real-time frosting detection and deep learning-based defrosting control for air-source heat pumps with optimized discharge pressure

초록

Frost formation in outdoor heat exchangers poses a critical challenge for air-source heat pumps (ASHPs) under frosting conditions. Conventional temperature-time (T-T) defrost control frequently results in performance degradation owing to premature or delayed defrosting. This study proposes a deep learning-based defrost control logic (DL-DCL) for ASHPs with real-time discharge pressure optimization under frosting conditions. The proposed DL-DCL simultaneously optimizes the discharge pressure in real time and predictively determines the defrosting initiation timing for maximizing the cumulative coefficient of performance (COP). DL-DCL uses a deep neural network to dynamically regulate the discharge pressure in response to instantaneous thermal loads, whereas a long short-term memory network forecasts 60 min-ahead trajectories of key thermodynamic variables. Defrosting is initiated when the predicted cumulative COP is maximized, thereby enabling proactive intervention before the frost-induced degradation becomes critical. Experimental validation of ASHPs with the DL-DCL under different operating conditions demonstrates cumulative COP improvements of 5.8%-7.5% over the T-T baseline. Adaptive model updating across successive cyclic operations reduces the indoor temperature deviation by 33.3%, thereby enhancing thermal comfort. Furthermore, its adaptive self-learning capability refines control performance over successive cycles using only built-in sensors, and its stable adaptability is verified under a stepwise time-varying indoor load. Seasonal simulations for six cities confirm cumulative COP gains of 5.5%-6.3% in cold and humid regions. DL-DCL offers a practical solution that requires no additional sensors during operation to maximize the performance and comfort of frost-prone ASHPs.

키워드