Ensemble learning based multi-modal intra-hour irradiance forecasting

Abstract

Accurate intra-hour irradiance forecasting plays an important role in improving the effectiveness of photovoltaic power management. More and more sensors, for example, total sky imager and cloud sensors, are widely used in photovoltaic plants, so that various factors could be comprehensively concerned to improve the accuracy of forecast. However, the existing methods have difficulty in encoding heterogeneous multi-source data dynamically. To mitigate these problems and improve the practicality of the model, an ensemble learning based multi-modal intra-hour irradiance forecasting method is proposed to forecast the global horizontal solar irradiance in the next 10 min. First, six base learners are built by selecting different combinations of multi-modal data as inputs for considering various factors. Then, a linear ridge regressor is used to integrate the pre-trained base learners to obtain outputs. Furthermore, to improve the robustness of the ensemble models, a dynamical fine-tuning scheme based on clear sky index is proposed. The weather states are predicted by a hidden Markov model based on the clear sky index. Experiments show that the proposed forecasting method effectively improves the accuracy of intra-hour irradiance forecasting by 11.6% when compared with commonly used models.