Solar Radiation Forecasting Using Hybrid Convolutional Long Short Term Memory Neural Network

Abstract

With the quick advancement of solar PV based power invasion in the cutting edge electric power framework, appropriate anticipating of solar irradiance take up a significant role in planning, operation, unit commitment of power system. However, the noteworthy historic solar irradiance data demonstrates uncertain and non-straight patterns, which makes it hard to foresee the radiation precisely. Taking these challenges into account, this paper proposes a deep learning based hybrid model named as Convolutional Long Short Term Memory (CLSTM) neural network to fore castsolar radiation in half an hour ahead. The Convolutional part of the model extracts the features (time series patterns) form the data set and the Long Short Term Memory part processes the data to overcome the long term dependency problem and stores the useful information in its memory part. The proposed model has been trained with two years historic data that includes global horizontal radiation (W/m^2), temperature (0C) and relative humidity (%). The comparative (with recurrent neural network and convolutional neural network) results (Mean Absolute Error, Root Mean Squared Error, Mean Absolute Percentage Error) shows that the proposed model performed better in learning the uncertainties in the dataset and therefore can be useful in predicting the solar irradiance in solar power plants. Along these lines, the proposed model has potential in harvesting power from inexhaustible solar energy resource to contribute in electric power framework.