Finding the Optimum Horizon for Short-Term Solar Irradiance Forecasting Using Long Short-Term Memory (LSTM) Network

The emerging negative effects of non-renewable fuels have compelled many nations to promote using Renewable Energy Sources (RES). Because it is affordable and readily available, solar power is among the RES that has achieved acceptance globally. Solar energy is, however, infamously unpredictable. This is why research fields that involve predictive analysis continue to have a considerable demand for improving the accuracy of global solar radiation prediction. Although Artificial Intelligence (AI) models have been quite helpful, researchers are still developing algorithms that can produce the least amount of prediction errors. This paper presents a successful improvement in estimating solar irradiance using the Long Short-Term Memory (LSTM) network. The input dataset used consisted of historical Cape Town meteorological data collected at five different horizons. Results obtained show that LSTM performed better than the Support Vector Regression (SVR) model, with an nRMSE value of 2.4 percent at 5-minute intervals of prediction. It is suggested that the real-time application of these findings to the operation of solar power plants in the study location produce reliable electricity for the consumers.