Predicting the Surface Temperature of Radiative Cooling Coatings with Time-Series Forecasting: Validation of the Small-Batch Training Dataset and Implementation of a Hyperparameter Optimization Strategy

Abstract

Radiative cooling coatings can maintain their surface temperature below ambient air temperature under solar radiation. The hemispherical emissivity model (HEM) is a widely used physical model for calculating the surface temperature of radiative cooling coatings; however, its reliance on fixed radiative properties limits its long-term accuracy. To solve this problem, this study used bidirectional long short-term memory (Bi-LSTM) and encoder-Transformer (E-T) models to capture the dynamic changes in the cooling performance. Subsets containing 20%–50% of the original training dataset’s size were used to validate the impact of the small-batch training dataset. An automatic hyperparameter optimization strategy was also proposed to determine the optimal hyperparameter combination. The results demonstrated that across all training ratios, both the Bi-LSTM and E-T outperform HEM, with E-T providing the highest accuracy. The E-T model, within the 20%–50% training ratio range, exhibited the root mean squared error (RMSE), mean absolute error (MAE) and coefficient of determination (R²) ranging from 1.24°C to 1.58°C, 0.94°C to 1.26°C, and 0.92 to 0.95, respectively. Comparatively, the HEM showed the RMSE, MAE and R² ranging from 1.46°C to 1.76°C, 1.18°C to 1.39°C, and 0.91 to 0.94, respectively. A roughly 15% prediction accuracy improvement was hence achieved.

Keywords: Radiative cooling coating, Time-series forecasting model, Automatic optimization strategy, Small training dataset