Short-Term Forecasting of Electricity Consumption in Sparsely Populated Areas Using a Recurrent Neural Network

Abstract

The article considers the task of improving the energy efficiency of the electric power industry, namely, improving the accuracy of forecasting electricity consumption. The relevance and importance of short-term electricity consumption forecasting in sparsely populated areas is substantiated, which consists in an increase in the number of autonomous cottage settlements in many regions of Russia, and at the same time, a small number of scientific papers on this topic. A recurrent neural network LSTM was chosen to implement short-term forecasting. The mathematical model of LSTM has been adapted consisting of six equations describing: accounting for previous values of power consumption (LSTM latent state); exclusion of power consumption values that have not reached the threshold value (LSTM filter); updating the time series of power consumption values (adding new values to the LSTM memory); forecasting values of power consumption with a given step-gnosis (new latent state of LSTM). The architecture of an artificial neural network is proposed, which considers all external factors affecting the nature of electricity consumption in a sparsely populated area, namely: weather data (temperature, wind speed); day type (weekday, weekend, holiday); seasonality (winter, spring, summer, autumn). For the software implementation of recurrent LSTM in the MATLAB modeling environment, an algorithm for short-term forecasting of power consumption was developed using the Deep Learning Toolbox library functions. The developed software was tested using ANN architectures with a different number of hidden layers. The analysis of the obtained results made it possible to determine the most accurate predictive model – the recurrent LSTM model containing four hidden layers. It has been experimentally proven to increase the accuracy of short-term power consumption forecasting by considering a variety of changing environmental parameters.