Effective Implementation of Convolutional Long Short-Term Memory (ConvLSTM) Network in Forecasting Solar Irradiance

Abstract

Even though solar energy usage for power generation is currently gaining popularity in many nations of the world, its unpredictability is still posing a great challenge. This problem is mainly caused by the stochastic nature of radiative power responsible for solar energy use in electric power generation. In this paper, a Convolutional Long Short-Term Memory (ConvLSTM) network is introduced to predict solar irradiance on an hourly basis. Five-year historical weather information obtained from Johannesburg was utilized as the input dataset. Using 80% of the entire dataset, the model was trained up to 1,000 epochs until there was no significant improvement in its accuracy. The testing data were introduced to determine the effectiveness of the predictive model. To measure its accuracy, the Root Mean Square Error (RMSE) metric was used. The result obtained from this experiment was compared with the ones obtained using Support Vector Machine (SVM) and Extreme Gradient Boosting (XGB) models trained separately with the same amount of data. The performance indicators showed that the ConvLSTM model outperformed the XGB and SVM models with the nRMSE value of 1.62%. It is suggested that implementing ConvLSTM for solar power prediction in Johannesburg could ensure better control of the problems of cascading solar radiation on power grid connections.