Deep learning based solar radiation micro forecast by fusion of infrared cloud images and radiation data

Abstract

Solar irradiance forecasting has been gaining paramount importance in recent years due to its impact on power grids. However, solar energy harvesting over shorter periods also brings new challenges due to its intermittent and uncertain attributes. Hence, accurate forecasting has become an indispensable aspect of the effective management of power system operations. The existing models focus on using only time-series data for solar radiation forecasting. But during cloudy time instances, it fails to quickly capture the nonlinear Spatio-temporal variations in the data for shorter periods. To bridge this gap, in this paper, a multi-modal fusion network is developed for studying solar irradiance micro forecasts by using both infrared images and past solar irradiance data. Here both spatial and temporal information is extracted parallelly and fused using a fully connected neural network. The solar forecasts of the proposed methods are evaluated against benchmark models in terms of Mean Absolute Percentage Error (MAPE) and other qualitative measures. The experimental results illustrate that the multi-modal fusion networks outperform the existing methods while predicting solar irradiance for cloudy days as well as mixed days (both cloudy and sunny days). Hence a transfer learning-based classifier with 99.23% accuracy is developed to categorize the cloudy days from sunny days. In the case of higher horizon forecasts, the proposed models show the optimum trade-off between performance and test time. Moreover, the Multiple Image Convolutional Long Short Term Memory Fusion Network (MICNN-L) shows a 46.42% improvement in MAPE whereas the Convolutional Long Short Term Memory Fusion Network (CNN-L) has a 42.02% increase when compared to the benchmark machine learning and deep learning models.