HOURLY SOLAR RADIATION FORECASTING USING ARTIFICIAL NEURAL NETWORK MODEL FOR COLOMBO, SRI LANKA

Abstract

Sri Lanka is a tropical country located close to the equator with abundant sunlight throughout the year. For efficient utilization of this solar resource for power generation in photovoltaic (PV) systems and agricultural modelling, prior knowledge of global solar radiation (GSR) in the future is important. Limited availability of onsite GSR data and the high cost are the main barriers in forecasting GSR for Sri Lanka. As a solution this study suggests an artificial neural network (ANN) model to forecast hourly solar radiation using weather data and solar angles to forecast GSR in Colombo, specifically using feedforward neural network (FFNN) trained with Levenberg- Marquardt (LM) back propagation algorithm. Hourly weather data for 6 weather variables and two solar angles from 1st of March 2017 to 14th of February 2018 were used for training, validation and testing the network. Input parameters and training parameters were adjusted to identify the most accurate network configuration and the performance of the network was measured using normalized mean squared error (NMSE). Coefficient of determination (R2) measured to identify the appropriateness of using weather variables and solar angles to forecast solar radiation. The final hourly FFNN model consists of 2 hidden layers and there are 5 neurons and 3 neurons in each layer respectively. This model was able to forecast hourly solar radiation with 0.0961 NMSE and the R2 was 90.39%. This implies the capability of this model for prediction of global solar radiation when unseen weather data input supply to the model and ensure the accuracy of the result.