Abstract:

This heat pump study introduces a Resonance-Based Sensitivity Analysis (RBSA) framework, which was inspired by the resonant characteristics of LSTM networks to visualize and interpret correlations between output features. First, we developed an LSTM network that predicts the time-series distribution of refrigerant within the system, focusing on refrigerant migration and its nonlinear dependency on the initial distribution in startup operation. A total of nine different datasets were employed, structured as a 3x3 matrix combining three levels of charged refrigerant, incrementing approximately 10wt% of system refrigerant, and three levels of initial refrigerant in evaporator from 30wt% to 70wt%. The prediction by the network achieved a coefficient of determination exceeding 95% in refrigerant distribution against validation data. Subsequently, targeted noise was applied to specific outputs of the trained network to analyze the intensity of inter-feature dependencies, demonstrating the utility of the RBSA approach in capturing causal relationships within the system. We investigated using both spike noise and persistent Gaussian noise in a comparative analysis to evaluate their distinct effects. During sensitivity evaluation with spike noise, we examined noise propagation between features using cross-correlation functions. The analysis revealed that the relationships between parameters maintained physical plausibility, even without an explicit physical model. We then introduced continuous white noise into the refrigerant distribution to examine its propagation effects and map how distribution fluctuations affected system operating parameters. The findings revealed that variations in refrigerant distribution substantially affect operating parameters such as mass flow rate, compressor input and condenser and evaporator capacity.