Time series forecasting. Forecasting payments of physical individual customers’ payments of an power supplyenergy sales company
This article solves the problem of forecasting a time series which describesing electricity payments of individualfor electricity by physical customers of a power supplyn energy marketing company. The complex and non-stationary nature of the time series, along with it is complex and non-stationary, seasonality and trend are expressed in it, which limits indicates that it is impossible to choose using simpler forecasting methods, such as extrapolation or naive forecasting method. Therefore in this regard, various machine learning methods and neural networks were chosen to solve the problem, various machine learning methods and neural networks were chosen which can find complex dependencies in data and learn to predict them. To solve the problem, several regression models of machine learning were chosen, often used in solving time-series forecasting problems, for example, XGBoost, SVM, LSSVM, LGBM, and Gradient Boosting, as well as neural networks, such as LSTM, RNN, and CNN. The best result was obtained using a regression support vector regression machine (SVR) machine. This implementation of this model has been integrated into a web application for time-series forecasting. The article also discusses an algorithm for determining the stationarity of a series using the Dickey–Fuller test and bringing a time series to a stationary series one by differentiating the former one a time series. In addition, the article addresses discusses the problem of forecasting emissions (peaks). Most forecasting methods do not solve the problem of predicting peaks because they consider them statistical outliers. In the article, this problem is solved by separating the peaks into a separate time series and predicting them separately from the main series with further connection of the two forecasts.
Authors: A. V. Eroshin, M. A. Potapova, A. E. Lonchina
Direction: Informatics, Computer Technologies And Control
Keywords: time-series, forecasting, support vector machine, seasonality, decomposition, regression, LSTM, peaks forecasting
View full article