Hour-Ahead Global Horizontal Irradiance Forecasting Using Long Short Term Memory Network

Energy management is an evolving challenge these days and deployment of renewable energy sources is an effective way out. It is imperative to specifically recognize the quantity from diverse resources and at distinct horizons: minutes, hours, and days, in order to achieve an effective deployment. Based on horizon, the approaches needed to be utilized for solar irradiance forecasting can be classified in two main categories: statistical time series prediction techniques for short to medium prospects and numerical weather forecasting techniques for medium to long-term prospects. This study mainly accentuates on statistical time series forecasting techniques for short term horizons (1 hour). A forecasting model has been introduced in this paper for solar irradiance depending upon the meteorological variables and historical data analysis. In the anticipated prototype, the Recurrent Neural Network (RNN) variation, specifically Long Short-Term Memory (LSTM) is employed. In order to develop best models that suit the global horizontal irradiance data, hyper-parameter tuning is being employed for achieving optimal parameters. After this, the quantity of data is being enhanced to discover the impact of increased data on this optimized model. The experiment is being performed on dataset gathered in Meteorological Station of National University of Sciences and Technology (NUST) Islamabad, Pakistan. Finally, the results of these models are compared by calculating errors.