Impact of Training Strategies on Satellite-Based Solar Irradiance Forecast
Abstract
This article investigates the impact of training strategies on the performance of a satellite-based model with an artificial neural network (ANN) for solar irradiance forecasting. The methodology analyzes the influence of: (i) cyclic variables as model inputs; (ii) global horizontal irradiance (GHI) versus the clearness index (Kc) as target variables; (iii) ANN specialization for individual forecast horizons versus generalization across horizons; and (iv) hyperparameter optimization using the Optuna framework. The temporal analysis covers hourly forecasts from 1 to 9 hours ahead. The main findings indicate that using Kc as a target variable eliminates the need for cyclic input variables, compared to using GHI directly. This transition in feature dominance was also demonstrated: satellite PCs are dominant in horizons of up to 2 hours, while seasonal features become more prominent after 3 hours. The increase in complexity of the model did not translate linearly into accuracy gains. The generic ANN 9-MLP demonstrated the best cost-benefit ratio for the 1 hour ahead horizon. However, for subsequent horizons, the 45 Specialized MLPs architecture presented the best overall error behavior and the best accuracy-cost trade-off. The analysis reinforces the nRMSE as the most reliable metric for performance comparison across different horizons.
Keywords
ANN, hourly, Kc, MLP, Optuna