Fractional-order long short-term memory network for forecasting of solar irradiance

The unpredictable nature of solar energy presents a significant obstacle to its effective incorporation into current grid systems. Global Horizontal Irradiance (GHI) is a critical factor in solar energy technology, as it directly influences the effectiveness of photovoltaic systems and solar thermal plants. Precise GHI forecasts are essential for this challenge and facilitate prompt and efficient involvement in the energy market. However, traditional neural network models often struggle to accurately predict GHI due to their time series data's nonlinear and nonstationary nature. Thus, this research proposes fractional-order LSTM (FOLSTM) models by accurately incorporating the fractional activation functions to predict GHI using the National Renewable Energy Laboratory data. The fractional activation functions, including sigmoid, tanh, and rectified linear units, have been developed using the fractional derivatives. The performance of FOLSTM models will be compared with the conventional LSTM models during training, testing, and validation in mean square errors and R-squared values. The results demonstrate the proposed FOLSTM models' effectiveness compared to conventional networks.