SVM-based Multi-point Prediction System of Solar Irradiance using DTW-based Cluster Analysis

Abstract

Unexpected short-term fluctuation of solar irradiance may negatively affect an electric power system; hence, predicting the solar irradiance is necessary. Considering that large quantities of solar power will be introduced across wide areas in the future, this paper addresses multi-point predictions of solar irradiance. Support vector machines (SVMs) perform well for predicting solar irradiance, but involve relatively high computational complexity. In addition, SVMs have to be repeatedly computed at multiple points because prediction models need to be updated regularly to adapt to the season. Here, we introduce a multi-point prediction system that considerably reduces the amount of calculation. Instead of constructing prediction models at all grid points, this system establishes clusters based on similarities of time series data by executing the dynamic time warping algorithm, and then constructs a small number of models representing each cluster. Further, it automatically tunes the number of clusters by inspecting prediction accuracy. Simulation results reveal that the system saves a considerable amount of calculation while maintaining high prediction accuracy.