Deep learning algorithms for very short term solar irradiance forecasting: A survey

Integrating solar energy with existing grid systems is difficult due to its variability, which is impacted by factors such as the predicted horizon, meteorological conditions, and geographic position. Accurate global horizontal irradiance (GHI) estimates can help to address this issue and allow for early and effective participation in the energy planning and management market. The existing models either use time series data or sky images in various network topologies to perform solar radiance forecasts. This study compares three categories of solar irradiance forecasting models such as time series-based, image-based and hybrid models. Here several state-of-the-art methods are compared against the proposed models, namely Convolutional Long Short-Term Memory Fusion Network (CNN-L) and Multiple Image Convolutional Long Short-Term Memory Fusion Network (MICNN-L). Both models use both infrared sky images as well as past values of GHI for prediction. These methods extract spatial features using convolutional neural networks and temporal features using long short-term memory networks. The extracted features are finally concatenated and passed through a fully connected layer to obtain a prediction. Further analysis also included using a feature extraction method such as optical flow (OF) on the image data before passing it to the hybrid model MICNN-L (OF). The results observed in this comparative analysis denote that MICNN-L improves the efficacy of the forecasts in cloudy conditions compared to the rest of the state-of-the-art approaches.