Time sequence prediction method based on fusion sequence decomposition and space-time convolution

Abstract

The invention provides a time sequence prediction method based on fusion sequence decomposition and space-time convolution; prediction is carried out through a neural network model (SDBRNN, Series-Decompose-Block RNN) fusing a sequence decomposition strategy and three-dimensional time sequence convolution; firstly based on an STL decomposition strategy, the model performs time sequence decomposition on one-dimensional time sequence data, and splices into a three-dimensional data block with time and space; high-dimensional spatial-temporal characteristics of the data block is extracted by using a three-dimensional convolution model, and thus finally performing back propagation learning on the high-dimensional characteristics by using an LSTM sequence module. An improved algorithm based on trend seasonal decomposition is provided for the problem that the period is uncertain in original STL decomposition; when a time sequence block is constructed, component combination is carried out according to correlation, so that a three-dimensional convolution module can better extract spatial-temporal characteristics; meanwhile, the three-dimensional convolution block is a module embedded in the LSTM cell and replaces the gate updating operation of the LSTM, so that the LSTM can learn with the spatio-temporal characteristics of the data. Experiments prove the reasonability of the time sequence block construction and the effectiveness of the model.

Description

technical field [0001] The invention belongs to the field of data analysis, and in particular relates to a timing prediction algorithm based on fusion sequence decomposition and space-time convolution. [0002] technical background [0003] Time series data is used to describe the characteristics of things over time. Studying time series can help people understand the historical development patterns of things and use them to predict the future trends of things, such as stock price fluctuations, road traffic flow forecasts, user Behavior analysis, etc. According to the mean value, variance and covariance of the sequence, the essential characteristics of the sequence can be described, and the sequence can be divided into stationary time series and non-stationary time series. On the traditional time series forecasting problem, many machine learning methods have emerged, such as support vector machines and decision tree algorithms. At the same time, with the rapid development o...

AI Technical Summary

Problems solved by technology

Convolutional networks can also adapt to more complex time series forms. However, the existing time convolution sequence prediction methods usually predict different observation points separately. Although this method is effective, it ignores the connection between observation points, because for correlation Strong multi-observation point time series often have spatial dependencies

Images