Seasonal Clustering Forecasting Technique for Intelligent Hourly Solar Irradiance Systems

The main purpose of the study is to reach a forecasting accuracy level through layering and stacking clusters of weather data to reduce seasonality-related uncertainty. To enhance forecasting accuracy, high-dimensional heterogeneous weather data should be added to training datasets; yet, weather data are deemed to have seasonality-related uncertainty. Since utilizing the long short-term memory (LSTM) model does not achieve the required results, the LSTM forecasting model performance deteriorates as a result of the change from univariate to multivariate analyses. Therefore, a primary investigation was conducted to analyze seasonality-based hourly predictions for global horizontal irradiance. The investigation found that seasonality affects accuracy of predictions due to high levels of Autumn- and Winter-related weather phenomena and climate uncertainty. Accordingly, seasonal clustering forecasting technique (SCFT) based on an LSTM hybrid strategy and stacked layers of weather clusters was proposed. In order to demonstrate the validity of this model, the proposed SCFT is compared to a variety of forecasting approaches in the literature. The comparison shows the SCFT’s superiority, especially during seasons that are predominantly rainy and overcast. Finally, in accordance with previous studies, suggesting the benefit of applying model testing in other parts of the world, SCFT is employed using data from other Köppen climate classifications. Furthermore, the performance of the proposed forecasting model, SCFT, shows significant stability and reliability after measuring it against Diebold–Mariano and Granger–Newbold tests.