Development of a Decision-Based Neural Network for a Day-Ahead Prediction of Solar PV Plant Power Output

Abstract

Day-ahead photovoltaic power prediction is vital for policy making and providing necessary backup capacities. Previous researchers include the implementation of time series, auto-regression and Soft computing techniques like Artificial Neural Networks and Fuzzy Logic. Artificial Neural Networks provides a better flt to complex, non-linear and error-prone data. The paper shows a comparative study of a Radial Basis Neural Network Schema (exact fit), a ‘k-means’ Radial Neural Network, and a Feed Forward Neural Network with Levenberg-Marquardt error backpropagation designed for the prediction of power output at an hourly resolution. The ability of the Neural Network to be trained to adapt to a previous set of data and then interpolate or extrapolate to the new data set has been exploited. The proposed model uses five meteorological variables and uses recorded data collected from the SN Mohanty PV Power Plant. Training of neural network is done on a monthly basis so that normalization constants of variables can be lower and better mapping can be produced. An improveddecision-based schematic using Neural Networks is proposedwhich combines the advantages of both Radial Basis Function (exact fit) and FFNN.