Optimized Support Vector Regression models for short term solar radiation forecasting in smart environment

High penetration of intermittent and uncontrollable renewable energy sources necessitates smarter and fast grid control mechanisms for maintaining system security. Evolution of smarter grids in such environment requires accurate short term power forecasting for optimum power dispatch, spinning reserve planning, stability analysis and security evaluation. A variety of models, such as numerical weather prediction, artificial neural network, machine learning algorithms and Bayesian approaches are used for solar radiation forecasting. Processing time of these models is quite high for an accurate prediction. This paper proposes a Support Vector Regression (SVR) model without hyper parameter optimization and two optimized SVR models, support vector regression with optimized hyper parameters using Genetic Algorithm (SVRGA) as well as Particle Swarm Optimization (SVRPSO) for solar radiation forecasting. These models use similar day approach for prediction considering that position of sun and earth is same on a similar day in previous years, albeit with the difference of cloud cover, cloud movement, wind speed and temperature. When dependent factors on similar day of previous year remain same, solar radiation would be similar to previous years similar day values. Results obtained from these models show that these models have strong potential towards short term prediction, and out of these SVRPSO gives better results compared to SVR and SVRGA.