Decoupling Irradiance Gain and Thermal Efficiency Loss in Photovoltaic Tracking Systems Using Explainable Machine Learning

The performance of photovoltaic (PV) generation systems is widely evaluated using physics-based simulation. However, this often provides limited insight into the interaction between the operating parameters that fundamentally govern energy outputs. In response to this limitation, this study presents an explainable machine learning framework that uses a normalized efficiency target to recover physically meaningful sensitivity coefficients directly from system-level data. The presented framework is validated on the System Advisor Model (SAM) simulated dataset for four mounting configurations: fixed-tilt, horizontal single-axis tracking (HSAT), tilted single-axis tracking (TSAT), and dual-axis tracking. The same system design parameters and loss assumptions are retained across all configurations to ensure the difference reflected in the generated dataset is due to the tracking modes. To capture the nonlinear input–output relationships, an XGBoost surrogate model is trained, and SHapley Additive exPlanations (SHAP) are subsequently applied to quantify the global importance of individual parameters. To investigate the interaction between the irradiance gain and temperature-induced efficiency losses at the system level induced by PV tracking, two complementary prediction targets are employed: raw system power output and a normalized efficiency-like metric. The results demonstrate that plane-of-array irradiance dominates PV power generation across all tracking configurations, while module temperature governs variations in normalized performance. Thermal sensitivity analysis under high-irradiance conditions reveals a weakly configuration-dependent slope of approximately −5.63 × 10−4 to −5.85 × 10−4 °C−1 (R2 ≈ 0.99). However, the relative spread among the slopes is only approximately 3.6%, showing that tracking systems increase energy yield primarily through enhanced irradiance capture while the temperature-induced efficiency penalty remains similar in engineering magnitude across configurations. The proposed framework extends the role of machine learning from prediction to physically meaningful interpretation and increased transparency.