Remaining Useful Life Estimation of Fouled HVAC Condensers via a Physics-Constrained Temporal Convolutional Network with SHAP Interpretation

Abstract

The condenser is one of the most fouling-prone parts of an HVAC system, and as deposits build up the compressor draws more power while efficiency falls. Predictive maintenance tries to catch this decline early, but it needs a dependable estimate of how much useful life the component has left. As a preliminary, simulation-based feasibility study, this work examines whether an explainable and physically consistent model can supply that estimate for an air-cooled condenser. Using the asymptotic Kern-Seaton model, a degradation dataset covering one complete fouling cycle (180 days sampled every minute) was generated, and from it six thermal and electrical features were derived and checked against the underlying physics. A Temporal Convolutional Network (TCN) was then trained with a physics-informed penalty that prevents the predicted life from rising over time, and SHapley Additive exPlanations (SHAP) were used to expose the reasoning behind each prediction. On a quartile-stratified test set the unconstrained TCN obtained a mean absolute error of 2.63 days and an R2 of 0.994, against 3.43 days for an LSTM baseline. Adding the physics penalty raised the error only slightly, to 2.88 days, while cutting non-monotonic predictions, a deliberate trade-off of a small amount of pointwise accuracy for physically consistent RUL trajectories. SHAP ranked the approach temperature as the most influential feature, which matches the way fouling degrades heat transfer. The predicted RUL and a derived Health Index were finally translated into a four-level maintenance decision scheme. The results indicate that the framework is promising; validation on real sensor data is the necessary next step.