Probabilistic Solar Power Forecasting Based on Bivariate Conditional Solar Irradiation Distributions

Abstract

We propose a two-stage probabilistic solar power (SP) forecasting algorithm to utilize the solar irradiation (SI) observations measured from different locations. In the first stage, we predict the SI based on the numerical weather prediction (NWP) after interpolating SI observations. Since the SI on the target location is not measured, we interpolate it using the spatio-temporal Kriging technique based on the SI observed from nearby weather stations. In the second stage, we forecast the SP based on the SI predictions after training the SI and SP observations. The model is trained by observations, but it forecasts based on predictions. Furthermore, in the two-stage model, forecasting errors can propagate across stages. We overcome these problems by using probabilistic forecasting. We design distributions of SI predictions through the probabilistic graphical model. Then, we extract SI scenarios from the distributions and predict SP scenarios based on these SI scenarios. We also group the NWP with respect to its prediction time, and we subdivide these groups as subgroups with respect to weather conditions. Furthermore, we propose a changeable ensemble model, where we have different weights for each weather condition. We verify our algorithm based on the data from the Korea power exchange renewable energy forecasting competition 2019. We finished the competition in 2nd place among a few hundred participants.