A short-term load forecasting method for power systems based on temporal convolutional networks

Abstract

The present invention provides a short-term load forecasting method for power systems based on temporal convolution networks, including: collecting historical load data, and preprocessing the data; constructing a causal dilated convolution model, and inputting the preprocessed data into two different The convolution process is performed in the causal dilated convolution model; the two results processed by the causal dilated convolution model are connected to form a residual block; the residual blocks are stacked to obtain a temporal convolutional network; the temporal convolutional network is used for full The convolutional layer is calculated to predict the future power load demand. The short-term load forecasting method of the power system provided by the present invention, through the causal dilated convolution model, performs causal convolution processing and dilated convolution processing on the data, and then submits it to the residual convolution processing to simplify the learning objectives and difficulties, and finally uses the time convolution The product network is used to perform full convolutional layer calculations, which reduces the time and hardware requirements required in the prediction process, and at the same time has comparable accuracy to the mainstream cyclic neural network.

Description

technical field [0001] The present invention relates to the technical field of electric power system and automation, and more specifically, relates to a short-term load forecasting method of electric power system based on temporal convolution network. Background technique [0002] Short-term power load is a periodic non-stationary stochastic process, which has seasonal changes, periodic changes by week and hour, and differences between holidays and normal working days. In short, short-term power load forecasting is a nonlinear mapping sequence modeling problem related to various factors, and the output sequence is predicted from the input sequence. [0003] In general, recurrent neural networks (RNN) are considered the default configuration for sequence modeling, and even Ian Goodfellow uses "Sequence Modeling: Recurrent and Recurrent Networks" as the chapter title in the book "Deep Learning", which indicates that sequence modeling There is a very close connection with loop...

AI Technical Summary

Problems solved by technology

[0005] The present invention provides a short-term load forecasting method for power systems based on temporal convolutional networks to overcome the technical defects of time-consuming and large memory consumption in the existing cyclic neural network because the network only reads and parses the input load data once.

Images