Short-term solar irradiance forecasting based on a novel Bayesian optimized deep Long Short-Term Memory neural network

Accurate forecasting is indispensable for improving solar renewables integration and minimizing the effects of solar energy's intermittency. Existing research on time series solar forecasting confronts challenges such as determining the accurate hyperparameters and flexibility in considering meteorological parameters. This study proposes a novel deep learning model, namely an optimized stacked Bi-directional Long Short-Term Memory (BiLSTM)/ Long Short-Term Memory (LSTM) model to forecast univariate and multivariate hourly time series data by integrating stacked LSTM layers, drop out architecture, and LSTM based model. The performance of the model is enhanced by Bayesian optimization with the tuning of six relevant hyperparameters. To evaluate the model, standard Global Horizontal Irradiance (GHI) and observed Plane of Array (POA) irradiance with meteorological real-world solar data from Sweihan Photovoltaic Independent Power project in Abu Dhabi, UAE, and NREL solar data for year-round data are forecasted. Furthermore, the performance of the proposed algorithm is also evaluated under weather uncertainty for different climate types. The forecasting accuracy is evaluated based on various performance metrics and it is observed that the proposed model offered the best R-2 values, 0.99 for univariate as well as multivariate models using GHI data and 0.97 using POA data. The findings suggest that the proposed model is a reliable technique for solar prediction due to its comparable performance with both GHI and POA in terms of accuracy.