Comparative Study of Optimal Long Short Term Memory Networks for One Day Ahead Solar Irradiance Hourly Forecast

Abstract

Energy sustenance is one of the key challenges India is facing in the contemporary time. Rise in global warming and the increasing need for dependency on clean energy has motivated researchers to develop novel techniques for harnessing maximum energy from renewable sources such as solar irradiance. However, one major issue which is impeding the large scale optimal implementation of solar farms is the uncertainty associated with solar irradiance. Although several statistical forecasting methods have helped in this regard, they could not contribute to efficient utilization of solar energy. In this work, long short term memory networks (LSTMs) are implemented for modelling the time series data of solar irradiance. LSTMs are deep neural networks which are proven to be extremely efficient in modelling nonlinear time series data with long term dependencies. However, LSTM networks are modelled using several heuristically governed parameters, making them an ineffective tool for time series regression. A novel multi-objective evolutionary optimization framework is proposed for optimal design of LSTM networks for emulating the real world solar irradiance data. The optimally trained LSTMs are used to forecast 1 day-ahead hourly prediction. LSTMs are compared with state-of-the-art system identification tools – Wavelet networks and feedforward neural networks through nonlinear auto-regressive exogenous modelling. LSTMs were found to be better with a root mean square error of 13% and R2 (correlation coefficient-a statistical measure of goodness of fit) value of 0.976.