Dynamic Estimation of Representative Backsheet Temperature for Rooftop Photovoltaic Modules Under Extreme Heat: Thermal Inertia, Temporal Resolution, and Machine Learning

Abstract

Accurate module-temperature estimation supports assessment of rooftop photovoltaic (PV) systems, but minute-scale meteorological fluctuations and thermal memory challenge steady-state models. This study used 44,638 synchronized one-minute records from a rooftop PV platform in Chongqing, China, during August 2025. The representative backsheet temperature was the arithmetic mean of five simultaneously valid sensors. Physical, static XGBoost, and thermal-history-aware XGBoost models were compared. Lagged values, strictly prior rolling means, and increments were constructed uniformly for four meteorological variables at 5, 10, and 20 min windows. Three expanding-time validation folds and a one-standard-error parsimony rule selected eight features: four current variables and their strictly prior 10 min means. During the retrospective 27–31 August held-out comparison, the dynamic-model MAE was slightly higher than the static-model MAE (1.088 versus 1.081 °C). RMSE decreased from 1.757 to 1.437 °C and P95AE from 3.907 to 3.049 °C. Paired day-level bootstrap intervals supported the observed tail-error reduction more clearly than the RMSE difference within this five-day record. MAE was about 40% lower during severe ambient heat and 45% lower during rapid irradiance decrease but increased at night. Daily peaks were underestimated by 2.574 °C on average; absolute peak-timing error had a mean of 26.8 min and a median of 1 min. The findings support site-specific thermal-history features for estimating the monitored-point mean, not a universal thermal timescale or uncalibrated safety-related peak thresholding. Prior inspection of the comparison dates, limited temporal coverage, and absent external validation restrict transferability.