Read-First LSTM model: A new variant of long short term memory neural network for predicting solar radiation data

The prediction of solar radiation data is important for countries to reduce their dependence on fossil fuels. Since the development of solar energy systems relies on an accurate prediction of solar radiation data, this study is conducted to predict monthly and daily solar radiation data and contribute to the development of solar energy systems. The current study develops a long short term memory (LSTM) model that can extract temporal features more efficiently than other deep learning models and predict solar radiation data. The new model is called the Read-first LSTM (RLSTM) model. The gate units of the LSTM model are independent, so they may not fully extract the features of long time series. Thus, the current study is conducted to address the limitations of the LSTM model for predicting solar data. The main innovation of this study is to develop an improved LSTM model to predict solar radiation data and establish a collaborative process between gates. While recent studies focus on optimizing LSTM parameters, the current research improves the efficiency of LSTM gates. Since there is a collaborative process between the gates of the RLSTM, correlation values ​​, and temporal features can be captured effectively. Climate data are used to predict solar radiation in two basins of Iran country, including the Kashan Plain and the Sefidorod Basin. The Boruta-Random Forest (BRF) feature selection algorithm was used to determine the best input scenario. The RLSTM model was compared with the LSTM model, recurrent neural network (RNN), radial basis function neural network (RBFNN), and Bidirectional LSTM (BILSTM) model. The RLSTM model could successfully predict the monthly solar radiation data in the Kashan plain. The RLSTM decreased the testing mean absolute error (MAE) of the other models by 5.8%-42%, respectively. The RLSTM model also accurately predicted daily data in the Sefidrood basin. The RLSTM improved the testing index of agreement (IA) of the BILSTM, LSTM, RNN, and RBFNN models by 5.2%-18%. The RLSTM enhanced the Nash–Sutcliffe efficiency of the other models by 5.2%-18%. The R2 values of RLSTM, BILST, LSTM, RNN, RBFNN, Prescott model, Ogelman model, Bakirci model, Rietveld model, and Almorox model were 0.9988, 0.9812, 0.9811, 0.9703, 0.9698, 0.9514, 0.9489, 0.9399, 0.9322, and 0.9284, respectively. The study demonstrates that RLSTM outperforms other models in predicting monthly and daily solar radiation data. The results provide insights into the limitations of existing LSTM models in predicting solar radiation and the importance of studying correlations between gate units. The study contributes to renewable energy development by providing a more reliable method for predicting solar radiation. The new model enhances the efficiency of empirical models for predicting solar radiation data.