Correlation-based time sequence data multi-step prediction method and system

Abstract

The invention relates to a correlation-based time sequence data multi-step prediction method and system. The method comprises the steps of setting an output length and a regression order according to the correlation between adjacent time point data in to-be-tested time sequence data, and dividing the to-be-tested time sequence data into input output data pair sets according to the output length and the regression order; and converting the input output data pair sets into training data of a multi-output Gaussian process model, performing training to generate a prediction model, and inputting the to-be-tested time sequence data to the prediction model to obtain final predicted values, and until the total number of the final predicted values is greater than or equal to a predicted step length, outputting the final predicted values. According to the method, a plurality of future values are predicted at the same time by using the multi-output Gaussian process model based on a multi-input multi-output policy according to the correlation between the adjacent time point data in the to-be-tested time sequence data, and gradual backward prediction is performed in an iterative manner, so that high prediction accuracy is achieved.

Description

technical field [0001] The invention relates to the field of multi-step forecasting of time series data, in particular to a method and system for multi-step forecasting of time series data based on correlation. Background technique [0002] Time series data covers almost any field of science and engineering. Time series data forecasting utilizes associated forecasting models that take historical data as input to predict future values ​​based on historical observations. Time series data forecasting has been very important in financial market analysis, economic forecasting, environmental monitoring and other aspects. How to find and describe the change law of time series data, and establish corresponding forecasting models to reduce the forecast error of the model is very important for time series data forecasting. [0003] Multi-step forecasting of time series data involves two aspects of forecasting strategy and model selection. In terms of prediction strategies, there ar...

AI Technical Summary

Problems solved by technology

[0005] Traditional multi-step forecasting methods for time series data are based on some models and strategies introduced above, such as deep neural network regression models with multiple input and multiple output structures, using Gaussian process iterations or direct strategy models, etc. There are also some methods in When forecasting time series data, factors related to the time series data are added to jointly predict its future value, but these models do not consider the correlation between time points before and after time series data, and the Gaussian process method is single-output, there is no multiple-input multiple-output strategy

Images