A wavelet-based method for high resolution multi-step PV generation forecasting

Forecasts play a vital role in maintaining power system stability and maximizing economic benefits of distributed energy resources. The issue with PV generation forecasting is that it relies on forecasts of solar irradiation which, due to the complex, nonlinear relationship between humidity, pressure, temperature, and cloud transients, can be quite difficult to model. Two important decisions in the forecasting process are selection of the forecasted variable (model output) and selection of explanatory variables (model inputs). This paper proposes a new method to forecast PV generation using wavelet based input selection and an output variable that directly represents clouds transients. We model this cloud effect by first determining a clear sky model (CSM) and forecasting the difference between the CSM and actual measurements of global horizontal irradiance (GHI). Potential model inputs are first decomposed using wavelet multi resolution analysis and final input selection is based on the correlation between the inputs and output at various timescales. Two separate neural network structures are designed to separately forecast sunny and cloudy days. Using the high resolution forecast of GHI (20 min increments), the next day´s PV generation is determined. This method improves on the persistence method by 69% on sunny days, 26% on cloudy days.