Electrical load prediction method based on time series data periodicity

Abstract

The invention discloses an electrical load prediction method based on time series data periodicity. The model involved in the invention comprises two modules: a trend modeling module and a period attention module, which are used for respectively capturing time sequence trend and periodicity characteristics. Wherein the trend modeling module adopts an improved TCN and is responsible for capturing long-term and short-term dependence of a time sequence and modeling the trend of the time sequence; the period attention module is composed of convolution operation and BiLSTM, extracts period characteristics of a time sequence through cross-period attention of close positions of prediction points in a historical period, enables time sequence information to be transmitted in a cross-period mode by using a bottom layer principle of a gated recurrent neural network, and simultaneously enables the time sequence information to be transmitted in a cross-period mode. And the problem that the gradient disappears when the recurrent neural network processes the long sequence input is avoided. Experiments prove that the periodic characteristics of the time sequence are explicitly concerned, so that the method has better interpretability, and meanwhile, the accuracy of an electrical load data prediction task is improved.

Description

technical field [0001] The invention belongs to the field of time series forecasting. Aiming at the forecasting task of fixed period electric load data, a periodic electric load forecasting method based on time series data is designed. Background technique [0002] Time series refers to a set of statistical data arranged in chronological order that is observed or recorded for the same phenomenon. Time series data are ubiquitous in daily life, from electricity load, traffic flow, temperature changes to oil prices, etc. In practical applications, users usually predict new development trends or potential dangerous events based on the observation and analysis of historical time series signals. For example, plan better driving routes by predicting traffic flow, or make more appropriate power supply plans by predicting residential electricity load. [0003] Traditional time series forecasting methods are mainly divided into two types: statistical methods and traditional machine l...

AI Technical Summary

Problems solved by technology

[0006] Although the current neural network has made good achievements in time series forecasting tasks, few existing deep learning models explicitly take the periodicity of time series into account. Recurrent neural networks and convolutional neural networks are not There is no focus on the periodicity of time series, the attention mechanism focuses on important time points in historical data, and there is no research on the periodicity of the series

Images