Deep learning solution for intra-day solar irradiance forecasting in tropical high variability regions

Located on the Equator, Singapore has one of the most challenging climate for solar irradiance forecasting. The tropical rainforest climate in the area demonstrates high variability in solar irradiance due to the dynamic and unpredictable cloud formation. To provide a solid solution for intra-day (1-6 hours) solar irradiance forecasting in the area, we design and implement deep learning solutions including the state-of-the-art machine learning models: deep neural network, extreme gradient boosting, random forests, extremely randomized trees and adaptive boosting. By using stacked generalization, the individual machine learning models can be combined to improve the forecasting accuracy further. To appropriately design and implement these models in intra-day solar irradiance forecasting, input features are carefully prepared and processed. After proper feature selection, the machine learning models are implemented and optimized specifically for our application. Then the models are combined using stacked generalization to achieve the optimal forecasting accuracy. For each forecasting horizon separated by one hour, a specific deep learning structure is proposed.