Short Term Renewable Energy Forecasting with Deep Learning Neural Networks

Abstract

Renewable energy sources (RES) are sources of electricity production with a sharply uneven load coverage schedule. Unsatisfactory accuracy of RES forecasts leads to additional production/consumption imbalances and additional costs for settling these imbalances. The need to unload or load flexible electricity producers to compensate for the potential imbalance created by electricity producers at the feed-in tariff will lead to a significant increase in the cost of electricity in such organized market segments as Balancing Market and Ancillary services market. Created imbalances of RES will lead to an even greater increase in electricity prices in terms of the transmission tariff of the Transmission System Operator due to both the payment of electricity produced for feed-in and the increase of imbalances created by them. In such conditions, you can improve the quality of decision-making when planning regimes using accurate forecasts for day-ahead and intraday markets. Statistical methods for forecasting time series have become widespread in the problem of forecasting the RES energy generation. The essence of the methods of this class is to select the parameters of statistical models by minimizing the error of the forecast on historical data. The simplest are one-factor models that take into account only the previous values of generated energy volume. In this class, the most common models are ARIMA (Box-Jenkins in various modifications), exponential smoothing (Holt-Winters in various modifications), and models based on the Kalman filter. The development of machine learning theory and artificial intelligence has led to effective mathematical tools for constructing complex hierarchical models such as support vector machines (SVMs) and artificial neural networks (ANNs) including deep learning methods. The advantages of models of this class are their flexibility, high generalization ability (the ability to make accurate predictions on data that are not in the learning process), and the ability to work with data without prior selection of significant features by an expert. Usually for a neural network of deep learning requires much more data than for classical models. Their disadvantages include the complexity of developing a network architecture and calibration of the learning process. By choosing successful architecture and learning procedure parameters, it is possible to fully automate the neural network learning process for new objects and adapt to changes in existing data. In this chapter, we propose a novel architecture for short-term energy forecasting of aggregated RES generation.