Explainable Boosting Machines for Thermal Load Prediction and Enclosure Thermodynamics of Residential HVAC Condensing Units

Abstract

The architectural integration of heating, ventilation, and air-conditioning systems in contemporary residential layouts increasingly subordinates thermodynamic performance to visual aesthetics, resulting in the concealment of outdoor split-system condensing units within louvered cabinets, service ducts, and constrained balcony alcoves. Such enclosures throttle the condenser airflow and, more damagingly, permit the exhaust plume to recirculate into the intake face, elevating the on-coil temperature above ambient and driving the vapour-compression cycle toward a high-lift, low-efficiency regime. Existing predictive models of this degradation are either computationally prohibitive, in the case of three-dimensional conjugate simulation, or opaque, in the case of deep and ensemble learners, and neither form yields the transparent, auditable relationships that building codes require. This work develops a framework based on Explainable Boosting Machines, a bagged and cyclically boosted realisation of generalized additive models with pairwise interactions. We first derive a closed-form cycle-level degradation law in which the compressor power admits a simple pole in a single dimensionless recirculation-resistance index, and we prove that the resulting physics is exactly representable as a second-order functional ANOVA decomposition in latent coordinates. This establishes that an additive learner with pairwise terms is not merely an interpretable approximation but the correct hypothesis class for the problem. A coupled computational fluid dynamics and lumped-parameter cycle model generates 15,000 parametric observations spanning frontal, lateral, and overhead clearances, louver porosity, ambient temperature, and part-load ratio. The learned model attains a mean absolute percentage error of 2.1% on held-out data, matching gradient-boosted tree and neural baselines while remaining fully decomposable. The extracted shape functions localise a sharp threshold in frontal clearance near 0.45 m, below which the contribution to compressor power accelerates in a manner consistent with the analytically predicted pole, and the leading pairwise term quantifies the porosity compensation required when that clearance cannot be met. The resulting relationships are stated as candidate clauses for residential energy codes.