Neural Network Approach to the Electricity Market Forecasting with Different-frequency Data among the Explanatory Variables

The effectiveness of planning in the electricity sector depends on the accuracy of electricity demand forecasting. Electricity generator companies need to have a short-term forecast one hour ahead to predict capacity utilization. However, most of the causal factors have a lower frequency of measurement, for example, macroeconomic indicators. As a result, there are problems of combining data of different frequencies. We propose a prediction model that combines convolutional (CNN) and recurrent neural networks (RNN). The input block of the network takes data of different frequencies as one pattern. The reduction of data to one frequency occurs endogenously in the model using convolutional blocks. We assume that this makes it possible to remove noise and reveal the spatial correlation between factors better than in the case of linear interpolation. The recurrent block is used to extract useful prediction information from the temporal data structure. The results of testing the model demonstrate its effectiveness in comparison with benchmarks.