CN-121980238-A - Time sequence prediction method and system based on multi-scale feature reconstruction and multi-expert fusion

CN121980238ACN 121980238 ACN121980238 ACN 121980238ACN-121980238-A

Abstract

The invention relates to the technical field of machine learning, in particular to a time sequence prediction method and a time sequence prediction system based on multi-scale feature reconstruction and multi-expert fusion, wherein the method comprises the steps of acquiring multi-source time sequence data and preprocessing to construct a structural feature matrix; the method comprises the steps of carrying out multi-scale feature reconstruction on sequences in a matrix by utilizing three groups of one-dimensional convolution networks with different receptive fields to obtain a local, medium and global three-dimensional reconstructed feature sample set, carrying out time sequence slicing on the sample set, calculating statistical features in each slice, respectively obtaining three predicted components of basic linear extrapolation, covariant linear correction and nonlinear residual error by parallel branches for each-dimensional sample slice, dynamically fusing the three components by a gating network based on the statistical features to obtain branch predicted results of each scale, and finally fusing all the branch predicted results by a full-connection network to output final predicted values. The method and the device effectively improve accuracy, robustness and practicability of time sequence prediction.

Inventors

ZHAO JUN
DENG HAO
LU YUAN
AN BAIJING
LIN XIUHAN
Guo Kangzhuang
ZHANG YIYONG
LIU KANGXU
LI HAIBIN
LIU TAO

Assignees

山东鲁软数字科技有限公司

Dates

Publication Date: 20260505
Application Date: 20260114

Claims (10)

1. A time sequence prediction method based on multi-scale feature reconstruction and multi-expert fusion is characterized by comprising the following steps: Acquiring multi-source time sequence data comprising a target sequence and a covariate sequence, preprocessing the multi-source time sequence data and constructing a structural feature matrix comprising multi-dimensional time features; performing multi-scale feature reconstruction on sequences in the feature matrix by using three groups of one-dimensional convolution networks with different receptive fields to obtain a reconstructed feature sample set with three groups of local, medium and global scales; performing time sequence slicing on the reconstructed feature sample set, and calculating statistical features in each slice; for samples of each scale subjected to time sequence slicing, respectively acquiring a plurality of prediction components through parallel branches, wherein the prediction components comprise a basic linear extrapolation component based on a historical target sequence, a linear correction component based on a future covariate sequence and a nonlinear residual component based on history and future sequence splicing; Based on the statistical characteristics, dynamically fusing a plurality of prediction components through a gating network to obtain branch prediction results of all scales, and fusing all branch results through a fully-connected network to output a final prediction value.
2. The method of claim 1, wherein the target sequence is a power load sequence or a power usage sequence, and wherein the covariate sequence comprises meteorological data and time characterization data.
3. The method of claim 2, wherein preprocessing the multi-source time series data and constructing a structured feature matrix comprising multi-dimensional time features comprises: Dividing the multi-source time sequence data into a training set, a verification set and a test set according to a preset proportion in time sequence; Encoding the time stamp of the training set, and constructing a multi-dimensional time feature containing a date attribute and a period attribute; and splicing the target sequence, the covariate sequence and the multidimensional time feature to generate a structured feature matrix.
4. The method of claim 2, wherein preprocessing the multi-source time series data and constructing a structured feature matrix comprising multi-dimensional time features comprises: performing time sequence alignment based on sampling characteristics of each data source, and adaptively dividing a training set, a verification set and a test set according to fluctuation indexes of time sequence data; Inputting the time stamp information into a time sequence feature coding network, wherein the time sequence feature coding network automatically extracts and outputs multi-dimensional time sequence features integrating periodicity and eventuality, and the time sequence feature coding network comprises a time sequence self-encoder architecture, wherein an encoder part of the time sequence self-encoder comprises an attention mechanism for capturing global dependence and a convolution module for capturing a local periodic mode; And constructing a characteristic relation graph taking the target sequence and the covariate sequence as nodes, aggregating node information through a graph neural network, and generating a characteristic matrix which codes the association relation between variables and is used as the structural characteristic matrix, wherein the characteristic relation graph is a full-connection graph, and the edge weights of the characteristic relation graph are initialized according to the statistical correlation or mutual information between the variables.
5. The method of claim 1, wherein performing multi-scale feature reconstruction on sequences in the feature matrix using a one-dimensional convolutional network of three different receptive fields, results in a set of reconstructed feature samples of three sets of local, medium, and global scales, comprising: Adopting three-layer depth-increasing one-dimensional convolution network groups which respectively correspond to the local, medium and global scales; The first group of convolution networks are formed by single-layer convolution and are used for extracting detail features of the local scale; the second group of convolution networks is formed by two layers of cascaded convolutions and is used for extracting the periodic characteristics of the mesoscale; the third group of convolution networks is composed of three layers of cascaded convolutions and is used for extracting trend features of the global scale.
6. The method of claim 1, wherein time-series slicing the reconstructed feature sample set and calculating statistical features within each slice comprises: taking a preset fixed length as a history window, sliding on a time axis by a step length 1, and slicing the reconstructed characteristic sample set; For each slice, statistics of the target sequence and at least one key covariate within its history window are calculated, the statistics comprising at least one of mean, standard deviation and maximum.
7. The method according to claim 1, wherein the obtaining of the plurality of prediction components by parallel branches, respectively, for each scale of the time-sliced samples, comprises: Inputting the historical target sequence of each scale sample into a basic linear projection layer to obtain a basic linear extrapolation component; inputting a future covariate sequence of each scale sample into a covariate linear projection layer to obtain covariate linear correction components; And splicing the historical target sequence, the historical covariate sequence and the future covariate sequence of each scale sample into a conditional sequence, and inputting the conditional sequence into a conditional self-encoder based on a multi-layer perceptron to obtain a nonlinear residual error component.
8. The method of claim 1, wherein dynamically fusing a plurality of prediction components through a gating network based on the statistical features to obtain branch prediction results for each scale comprises: inputting the statistical features into a gating network, wherein the gating network outputs dynamic weights corresponding to the basic linear extrapolation component, the covariate linear correction component and the nonlinear residual component respectively; And carrying out normalization processing on the dynamic weights, and carrying out weighted summation on the results of the three prediction components to obtain branch prediction results of corresponding scales.
9. The method of claim 8, wherein the gating network comprises a multi-layer perceptron that maps the input statistical features to initial weights of the predicted components by nonlinear transformation and normalizes by Softmax function to output the dynamic weights summed to 1.
10. A time series prediction system based on multi-scale feature reconstruction and multi-expert fusion, comprising: The data acquisition module is used for acquiring multi-source time sequence data comprising a target sequence and a covariate sequence, preprocessing the multi-source time sequence data and constructing a structural feature matrix comprising multi-dimensional time features; The feature reconstruction module is used for carrying out multi-scale feature reconstruction on sequences in the feature matrix by utilizing three groups of one-dimensional convolution networks with different receptive fields to obtain a reconstructed feature sample set with three groups of local, medium and global scales; The slice processing module is used for carrying out time sequence slicing on the reconstructed characteristic sample set and calculating statistical characteristics in each slice; the branch prediction module is used for respectively acquiring a plurality of prediction components through parallel branches according to samples of each scale after time sequence slicing, and comprises a basic linear extrapolation component based on a historical target sequence, a linear correction component based on a future covariant sequence and a nonlinear residual component based on history and future sequence splicing; And the branch fusion module is used for dynamically fusing a plurality of prediction components through a gating network based on the statistical characteristics to obtain branch prediction results of all scales, and fusing all the branch results through a fully-connected network to output a final prediction value.

Description

Time sequence prediction method and system based on multi-scale feature reconstruction and multi-expert fusion Technical Field The invention belongs to the technical field of machine learning, and particularly relates to a time sequence prediction method and system based on multi-scale feature reconstruction and multi-expert fusion. Background The multi-input single-parameter time sequence prediction technology has important application value in the fields of power load prediction, power generation power prediction and the like. The conventional method has the obvious defects that a traditional statistical model such as vector autoregressive dependence linear assumption is difficult to describe nonlinear interaction, a deep learning model such as a cyclic neural network and a variant thereof can process sequence dependence, but have the problems of low training efficiency, gradient disappearance and the like, a self-attention-based Transformer architecture advances in long-range dependence modeling, different variables are generally mixed in characteristic dimensions when multi-variable input is processed, multi-scale features with physical significance are difficult to be explicitly extracted, and an adaptive modeling mechanism aiming at dynamic association relation between target variables and different covariates is lacked. In addition, the common defects of single feature extraction granularity, static fusion mechanism, low residual error learning efficiency and the like of the existing method generally exist, so that the model is limited in aspects of multi-scale feature self-adaptive extraction, multi-branch result dynamic fusion and efficient nonlinear residual error learning. Disclosure of Invention Aiming at the defects in the prior art, the invention provides a time sequence prediction method and a time sequence prediction system based on multi-scale feature reconstruction and multi-expert fusion, so as to solve the technical problems. In a first aspect, the present invention provides a time sequence prediction method based on multi-scale feature reconstruction and multi-expert fusion, including: Acquiring multi-source time sequence data comprising a target sequence and a covariate sequence, preprocessing the multi-source time sequence data and constructing a structural feature matrix comprising multi-dimensional time features; performing multi-scale feature reconstruction on sequences in the feature matrix by using three groups of one-dimensional convolution networks with different receptive fields to obtain a reconstructed feature sample set with three groups of local, medium and global scales; performing time sequence slicing on the reconstructed feature sample set, and calculating statistical features in each slice; for samples of each scale subjected to time sequence slicing, respectively acquiring a plurality of prediction components through parallel branches, wherein the prediction components comprise a basic linear extrapolation component based on a historical target sequence, a linear correction component based on a future covariate sequence and a nonlinear residual component based on history and future sequence splicing; Based on the statistical characteristics, dynamically fusing a plurality of prediction components through a gating network to obtain branch prediction results of all scales, and fusing all branch results through a fully-connected network to output a final prediction value. In an alternative embodiment, the target sequence is a power load sequence or a power consumption sequence, and the covariate sequence comprises meteorological data and time characteristic data. In an alternative embodiment, preprocessing the multi-source time series data and constructing a structured feature matrix containing multi-dimensional time features includes: Dividing the multi-source time sequence data into a training set, a verification set and a test set according to a preset proportion in time sequence; Encoding the time stamp of the training set, and constructing a multi-dimensional time feature containing a date attribute and a period attribute; and splicing the target sequence, the covariate sequence and the multidimensional time feature to generate a structured feature matrix. In an alternative embodiment, preprocessing the multi-source time series data and constructing a structured feature matrix containing multi-dimensional time features includes: performing time sequence alignment based on sampling characteristics of each data source, and adaptively dividing a training set, a verification set and a test set according to fluctuation indexes of time sequence data; Inputting the time stamp information into a time sequence feature coding network, wherein the time sequence feature coding network automatically extracts and outputs multi-dimensional time sequence features integrating periodicity and eventuality, and the time sequence feature coding network comprises a time sequence self-en