Dual-gate Temporal Fusion Transformer for estimating large-scale land surface solar irradiation

Abstract

An accurate estimation of land surface solar irradiation (LSSI) is crucial to address the solar intermittency for optimizing solar photovoltaic (PV) installation and mitigrating PV curtailment. This involves enhancing solar photovoltaic (PV) system efficiency by optimizing layout and maximizing solar energy capture and conversion. While deep learning methods have significantly improved the rapid and accurate estimation of solar irradiation, they face challenges in handling geographical heterogeneity and providing interpretable results. To address these challenges, this study proposes the Dual-gate Temporal Fusion Transformer (DGTFT), a novel interpretable deep learning network, to improve LSSI estimation. By integrating the Temporal Fusion Transformer with the Dual-gate Gated Residual Network and Dual-gate Multi-head Cross Attention, the optimal network achieved $R^2$=0.93, MAE=0.022 ($kWh/m^2$), RMSE=0.038 ($kWh/m^2$), rRMSE=0.13, and nRMSE=0.048 through ablation experiments. When applied to datasets observed from Australia, China, and Japan, DGTFT outperformed traditional machine learning methods with a minimum $R^2$ increase of 23.88%, MAE decrease of 43.18%, RMSE decrease of 9.09%, rRMSE decrease of 32.25%, and nRMSE decrease of 62.79%. Furthermore, the interpretability results of the DGTFT model indicate that clear-sky solar irradiation significantly contributed to the model’s performance from Australia and Japan; and the maximum temperature and humidity were the largest importance variables in the Chinese dataset. Accurately estimating LSSI, providing interpretable results, and generating continuous solar irradiation maps for large-scale areas, this study aids in quantifying solar potential and offers scientific guidance for the PV industry’s development.