Artificial Neural Network-Based Day-Ahead Solar Irradiance Forecasting for Unit Commitment Applications

The increasing penetration of solar photovoltaic (PV) generation introduces significant operational uncertainty due to the intermittent nature of solar irradiance, making accurate day-ahead forecasting essential for reliable power system scheduling. Although artificial neural networks (ANNs) have demonstrated promising forecasting capabilities, existing studies primarily emphasize aggregate accuracy metrics while providing limited assessment of forecasting robustness and operational applicability. This study proposes an ANN-based framework for day-ahead solar irradiance forecasting using historical solar irradiance, ambient temperature, wind speed, and calendar features obtained from the NASA POWER database. The model utilizes meteorological observations from the preceding seven days to predict the 24-hour irradiance profile of the following day, which is subsequently converted into PV power using a temperature-adjusted model. Forecasting performance is evaluated through statistical accuracy metrics, representative best-, median-, and worst-case weather scenarios, and hourly error analysis. The proposed ANN achieved an RMSE of$62.69 \text{Wh} / \mathrm{m}^{2}$, an MAE of$3 0. 7 5 ~ \text{Wh/m}^{2}$, and an$\mathrm{R}^{2}$of 0.9426. Across the representative daily scenarios, RMSE ranged from 12.96$\text{Wh}/\mathrm{m}^{2}$for the best-performing day to${253.27} \text{ Wh} / \mathrm{m}^{2}$for the worst-performing day, while the median day recorded an RMSE of 45.69 Wh/m${ }^{2}$. By integrating representative weather scenarios and temporal error analysis with PV power estimation, the proposed framework provides a comprehensive assessment of forecasting robustness and establishes reliable PV power profiles for day-ahead unit commitment and other renewable-integrated power system applications.

Hali tarjima qilinmagan