A Neural network based model for forecasting the power output of a commercial scale photovoltaic power plant

Abstract

Solar photovoltaic (PV) is penetrating electrical grids with a substantial growth of new additions, as a result of renewable energy policies and plans implemented locally and globally. However, the intermittent nature of the availability of solar energy brings an uncertainty into electrical power systems making it complex for power management and integrating into existing electricity infrastructure. This has been a key issue in promoting renewable energy in developing countries. Accurate solar power forecasts in different time horizons can play a vital role to bring down the uncertainty by a significant margin. In this work, a neural network (NN) model was coupled with a decomposition and transposition (D&T) model to forecast day(s) ahead hourly PV output of a grid connected 1 MW solar PV plant located in Hambantota, Sri Lanka. Historical weather and solar radiation data for last 14 years were collected from two APIs (Application Programming interfaces) for the location of PV plant and variation of global horizontal irradiation (GHI) with percentage cloud cover, rain, temperature, relative humidity, and wind speed were analysed. The selected parameters from the analysis together with day and hour numbers were fed in to the NN model through a scaling layer and trained it using Levenberg–Marquardt backpropagation algorithm. Optimum NN model was selected by changing the hidden layer sizes and calculating the mean squared error. The forecasted GHI values of the optimized NN model were decomposed to diffuse horizontal irradiance (DHI) and direct normal irradiance (DNI) using Erbs correlation, as the first step of D & T model. Then, DHI and DNI components were converted to global tilted irradiance (GTI) using HDKR correlation, in order to calculate solar PV output, including possible plant specific losses. The correlation coefficient (R) between GHI output and target values of the trained NN model for an unseen testing data set was observed to be 0.86. For final model, mean percentage forecasting accuracy was observed to be 86% with 12% standard deviation. The model could be adopted to any commercial or utility scale solar PV plant which is in a tropical climate region.