Abstract
Accurate ultra-short-term forecasting of photovoltaic output is essential for improving the reliability and stability of solar energy systems. However, many existing forecasting methods rely primarily on irradiance-based features and do not fully exploit the information contained in the solar spectrum. To address this limitation, this paper proposes a two-stage approach to ultra-short-term forecasting of PV output current using solar spectral information. In the first stage, solar spectral irradiance is forecasted, while in the second stage, the PV array current is estimated to use the predicted spectra and a multilayer perceptron (MLP) regression model. The proposed method is compared with three alternative forecasting schemes. Firstly, a model based on statistical features of solar irradiance, including the moving-average clear-sky index (CSI ma ) and moving-average variance (σ ma ). Secondly, a model combining CSI ma , σ ma , irradiance, and spectral data. Thirdly, a model using irradiance, the white-to-total ratio ( wtr ), and sky images. Experiments were conducted using data collected in October 2025, and forecasting performance was evaluated for horizons of 1, 5, 10, 15, 30, and 60 minutes using R² score, MAE, MSE, and RMSE. The results show that forecasting accuracy decreases as the prediction horizon increases, with R² decreasing and error metrics increasing across all models. Nevertheless, the proposed spectral two-stage MLP-based approach achieves the best performance, with an R² of 0.971 for a 1-minute horizon and strong explanatory power (R² ≈ 0.87) for horizons up to 15 minutes. These results indicate that incorporating solar spectral information can significantly enhance the accuracy of ultra-short-term PV current forecasting.