ABSTRACT

Renewable energy sources like solar are getting popular as a viable alternative to fossil fuel energy sources. The stand-alone system is popularly used for residential purposes while the grid tied is used mostly for power generating companies. For better management and planning, it is important to investigate the power variation with weather conditions and be able to predict the power that can be generated from a PV system. This project aimed to build a neural network based model to predict power from a PV system.

In implementing the model, a two-step approach was used which featured an auxiliary and a main model. The input to the auxiliary model was a weather forecast set; the auxiliary model gave Diffuse Horizontal Irradiance (DHI) and Global Horizontal Irradiance (GHI) as outputs. The inputs to the main model were weather forecast dataset, output of the auxiliary model and historic power data values. The output of the main model is forecasted/predicted power. Among the algorithms used were support vector machines, random forest regression, linear regression and recurrent neural networks. The model was built, trained and validated using ANN and statistical tools.

The model was found to have an MSE of 0.02387 and an MAE of 0.10493. The result showed that the predicted power closely followed the generated power and the ANN based two-step approach was effective in predicting the power of a PV system.