Application of Neural Networks to Predict Power Consumption of a Megapolis

Abstract

A model for daily schedules short-term forecasting of active power in Moscow is proposed in the article. It is formed on the basis of neural network algorithms, taking into account actual and forecast data on air temperature, cloudiness, precipitation and natural light. To predict daily active power curves, a hybrid model based on multivariate singular spectral analysis and fuzzy neural networks is proposed. Time series of power consumption and meteorological factors are decomposed into independent components in this model, with the help of which additive trend, harmonic and random components, used in the fuzzy neural network module, are formed. The initial data were archives of power consumption time series, air temperature, cloudiness, precipitation and natural light for the territory of Moscow in the period from 2017 to 2019. Predictive data on natural light were obtained using a neural network with long-term memory LSTM. Short-term forecasts of daily active power schedules for August 2019 were made. The results of short-term forecasting of daily active power schedules show that in most test cases the outcomes were obtained within acceptable errors.