Solar PV Power Forecasting Using Modified SVR with Gauss-Newton Method

Abstract

Globally among 28% of renewable energy in total generation mix, solar photovoltaic (PV) contributes more than 23% and it is growing rapidly in recent years. Solar PV power forecasting plays an important role in grid integration of solar PV plants, optimum solar PV power utilization, energy market trading, ancillary services and maintaining power system stability and control. In last decade, machine learning methods, especially Support Vector Regression (SVR) are widely used in solar PV output power forecasting. SVR requires less amount of data for training, has less computational complexity and better forecasting accuracy. However, performance of SVR based forecasting models depends on the initial guess values provided in the training phase. Finding appropriate initial values using optimization methods increase computational time and complexity of these models. This paper proposes a simple, quick learning and robust method to modify the SVR model using Gauss-Newton method. The proposed model is tested to forecast power on sunny, partly cloudy and mostly cloudy day. It is compared with forecasted power by SVR and actual power output of a real solar PV plant. The modified model of SVR with Gauss-Newton method successfully converts nonlinear data of solar insolation and PV power output into linear data and achieves better forecasting accuracy than traditional SVR model in all three weather conditions.