Deep learning approaches for PV forecasting

Deep learning (DL) is a part of the more general artificial intelligence (AI) field that focuses on building and optimizing complex "black-box" mathematical models. These models can learn and later recognize the descriptive features of complex phenomena happening in the real world, at the cost of a large amount of required data and a very high computational burden. However, due to their inherent flexibility and capabilities, DL models have found an increased interest among researchers. This chapter provides a clear outlook on how DL models are used in the field of PV forecasting, starting with general definitions of the used models and followed by case studies that demonstrate an example of how they can be utilized successfully. However, beforehand, it is important to identify the forecasting task and its variations, its role in the entire power system, and the typologies of used and available data. The definition will be followed by a short introduction to the basics of DL, with an overview of the training procedure and common challenges in the field such as overfitting. After the introduction, the chapter continues with three sections that share a similar general structure. In each section, a subset of DL models is introduced, including the necessary but concise theoretical explanations of the elementary building blocks, an analysis of the state-of-the-art applications recently proposed by researchers from around the globe, and a case study showcasing a detailed look at intricacies of the model application in a real-life scenario. The first section focuses on artificial neural networks (ANNs), also called fully connected neural networks (FCNNs), a typology that is by far the simplest of the bunch yet the most useful due to its flexibility. The next section addresses image-based forecasting, which in the PV field usually revolves around sky images and satellite images. Commonly, the convolution operation is used to extract the descriptive features from the images, hence the name of the model: convolutional neural network (CNN). The final section deals with temporal data and sequence analysis, focusing on recurrent models (RNNs) and their close relatives, such as long short-term memory (LSTM) and gated recurrent units (GRUs). Their structure allows for the recognition of both short- and long-term patterns that are often embedded in real-life temporal data.