Multivariable time sequence signal prediction processing method and device and electronic equipment
By employing signal weaving and reversible proxy signal processing, the prediction accuracy of multivariable time series signals is improved, solving the problem of decreased prediction accuracy in multivariable scenarios of existing models and achieving efficient multivariable signal prediction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-01-08
- Publication Date
- 2026-04-07
AI Technical Summary
Existing high-performance time series prediction models lack the ability to directly model the complex coupling relationships between multiple signals in multivariate scenarios, leading to decreased prediction accuracy and problems such as overfitting risk and weak generalization ability.
By extracting correlation features between multivariate signals through a signal weaving module, constructing reversible proxy signals, and processing these proxy signals using a high-performance univariate prediction model, the prediction results of the original signals are reconstructed, thereby improving the prediction accuracy of multivariate signals without changing the mature univariate model.
Without altering the univariate time series prediction model, this study improves the accuracy and robustness of multivariate signal prediction, reduces training costs, and enhances the model's generalization ability.
Smart Images
Figure CN121809370A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method, apparatus, and electronic device for predictive processing of multivariable time-series signals. Background Technology
[0002] In fields involving multivariate time series analysis, such as industrial monitoring, meteorological observation, and energy management, the system state is typically characterized by multiple interrelated sensor signals. For example, electricity price forecasting requires integrating historical price data and parameters affecting prices, such as weather, supply and demand data, electricity demand, and photovoltaic changes; load fluctuations in the energy system are dynamically coupled with variables such as electricity consumption, power generation, and grid frequency. The cross-channel correlation patterns inherent in these multivariate time series are the core basis for characterizing system behavior and predicting future states.
[0003] However, while current high-performance time series prediction models excel in univariate time series tasks, their design logic focuses on capturing local patterns in a single signal channel, inherently lacking the ability to directly model complex coupling relationships between multiple signals. When directly applied to multivariate scenarios, they can only process multi-channel signals through independent prediction and simple aggregation, completely ignoring the dynamic correlations between variables, leading to a significant decrease in prediction accuracy. Alternatively, they may attempt to capture correlations by building end-to-end multivariate models, but these models often face problems such as overfitting due to data sparsity, high training costs due to complex network structures, and weak generalization ability due to domain-specific design, because they need to learn both inter-variable correlations and univariate patterns simultaneously.
[0004] Therefore, finding a method to improve the accuracy of result prediction based on multivariate time series signals without changing the mature univariate time series prediction model has become a current research hotspot. Summary of the Invention
[0005] This invention provides a method, apparatus, and electronic device for predicting and processing multivariate time series signals, which improves the accuracy of result prediction based on multivariate time series signals without changing the mature univariate time series prediction model.
[0006] This invention provides a method for predicting multivariate time-series signals. The method includes: acquiring a time-series signal of an event to be predicted, wherein the time-series signal includes multiple variables; calling a pre-trained prediction model, wherein the prediction model includes at least a signal weaving module, a surrogate signal construction module, and a univariate prediction model; wherein the signal weaving module is used to obtain a multivariate woven signal corresponding to the time-series signal to be predicted based on the time-series signal to be predicted; the surrogate signal construction module is used to construct multiple reversible surrogate signals based on the multivariate woven signal and the time-series signal to be predicted, wherein the multiple reversible surrogate signals can reconstruct the time-series signal to be predicted through linear combination; the univariate prediction model is used to obtain a surrogate signal prediction result corresponding to the reversible surrogate signals based on the reversible surrogate signals, so as to obtain a prediction result of the event to be predicted based on the surrogate signal prediction result; and inputting the time-series signal to be predicted into the prediction model to obtain a prediction result of the event to be predicted corresponding to the time-series signal output by the prediction model.
[0007] According to a prediction processing method for multivariate time series signals provided by the present invention, the step of inputting the time series signal to be predicted into the prediction model to obtain the prediction result of the event to be predicted corresponding to the time series signal to be predicted output by the prediction model includes: inputting the time series signal to be predicted into the signal weaving module for multivariate feature weaving and fusion processing to obtain a multivariate woven signal output by the signal weaving module corresponding to the time series signal to be predicted, wherein the multivariate woven signal is a signal that fuses the features of each variable in the time series signal to be predicted; inputting the multivariate woven signal and the time series signal to be predicted into the proxy signal construction module for linear combination processing to obtain multiple reversible proxy signals output by the proxy signal construction module; inputting each of the reversible proxy signals into the univariate prediction model to obtain prediction results of each proxy signal output by the univariate prediction model corresponding to each of the reversible proxy signals; and obtaining the prediction result of the event to be predicted corresponding to the time series signal to be predicted output by the prediction model based on the prediction results of each proxy signal.
[0008] According to a prediction processing method for multivariate time series signals provided by the present invention, the step of inputting the multivariate woven signal and the time series signal to be predicted into a proxy signal construction module for linear combination processing to obtain multiple reversible proxy signals output by the proxy signal construction module includes: inputting the multivariate woven signal and the time series signal to be predicted into the proxy signal construction module, and performing linear combination processing based on a weight vector to obtain multiple reversible proxy signals output by the proxy signal construction module, wherein the weight vector is obtained through pre-training of the prediction model; the step of obtaining the prediction result of the event to be predicted corresponding to the time series signal to be predicted, output by the prediction model, based on the prediction results of each proxy signal, includes: performing reconstruction processing based on the prediction results of each proxy signal and the weight vector to obtain the prediction result of the event to be predicted corresponding to the time series signal to be predicted, output by the prediction model.
[0009] According to a prediction processing method for multivariate time series signals provided by the present invention, the step of inputting the time series signal to be predicted into the signal weaving module for multivariate feature weaving and fusion processing to obtain a multivariate woven signal output by the signal weaving module corresponding to the time series signal to be predicted includes: inputting the time series signal to be predicted into the signal weaving module, and performing nonlinear transformation through a fully connected layer in the signal weaving module to extract the static correlation between variable features; performing weaving and fusion processing based on the static correlation to obtain a multivariate woven signal output by the signal weaving module corresponding to the time series signal to be predicted.
[0010] According to a method for predicting multivariate time series signals provided by the present invention, the step of inputting the time series signal to be predicted into the signal weaving module for multivariate feature weaving and fusion processing to obtain a multivariate woven signal output by the signal weaving module corresponding to the time series signal to be predicted includes: inputting the time series signal to be predicted into the signal weaving module, and extracting the dynamic correlation between variable features within a preset time window through the convolutional layer in the signal weaving module; performing weaving and fusion processing based on the dynamic correlation to obtain a multivariate woven signal output by the signal weaving module corresponding to the time series signal to be predicted.
[0011] According to the prediction processing method for multivariate time series signals provided by the present invention, the prediction model is pre-trained in the following manner: freezing the model parameters of the univariate prediction model in the prediction model, wherein the model parameters of the univariate prediction model are pre-trained; based on the loss function, iteratively training the prediction model with the frozen model parameters of the univariate prediction model is performed until the function value of the loss function meets the preset requirements, thereby obtaining the trained prediction model.
[0012] According to a method for predicting multivariate time-series signals provided by the present invention, the loss function includes a regularization term, which is used to constrain the accuracy of the prediction results output by the prediction model during iterative training. The method involves iteratively training the prediction model with its parameters frozen based on the loss function until the value of the loss function meets a preset requirement, thereby obtaining a trained prediction model. This includes: iteratively training the prediction model with its parameters frozen based on the loss function until the value of the loss function meets a preset requirement and the accuracy of the prediction results output by the prediction model is greater than or equal to the accuracy of the prediction results output by the univariate prediction model based on the time-series signal to be predicted, thereby obtaining a trained prediction model.
[0013] This invention also provides a predictive processing apparatus for multivariate time series signals. The apparatus includes: an acquisition module for acquiring a time series signal of an event to be predicted, wherein the time series signal includes multiple variables; an invocation module for invoking a pre-trained prediction model, wherein the prediction model includes at least a signal weaving module, a surrogate signal construction module, and a univariate prediction model, wherein the signal weaving module is used to obtain a multivariate woven signal corresponding to the time series signal to be predicted based on the time series signal to be predicted; the surrogate signal construction module is used to construct multiple reversible surrogate signals based on the multivariate woven signal and the time series signal to be predicted, wherein the multiple reversible surrogate signals can be linearly combined to reconstruct the time series signal to be predicted; the univariate prediction model is used to obtain a surrogate signal prediction result corresponding to the reversible surrogate signals based on the reversible surrogate signals, so as to obtain a prediction result of the event to be predicted based on the surrogate signal prediction result; and a prediction module for inputting the time series signal to be predicted into the prediction model to obtain a prediction result of the event to be predicted corresponding to the time series signal to be predicted output by the prediction model.
[0014] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the prediction processing method for multivariable time-series signals as described above.
[0015] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the prediction processing method for multivariable time-series signals as described above.
[0016] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the prediction processing method for multivariable time-series signals as described above.
[0017] This invention provides a method, apparatus, and electronic device for predicting multivariate time-series signals. The method includes: acquiring a time-series signal of an event to be predicted, wherein the time-series signal includes multiple variables; calling a pre-trained prediction model, wherein the prediction model includes at least a signal weaving module, a surrogate signal construction module, and a univariate prediction model; wherein the signal weaving module is used to obtain a multivariate woven signal corresponding to the time-series signal to be predicted based on the time-series signal to be predicted; the surrogate signal construction module is used to construct multiple reversible surrogate signals based on the multivariate woven signal and the time-series signal to be predicted; the univariate prediction model is used to obtain a surrogate signal prediction result corresponding to the reversible surrogate signals based on the reversible surrogate signals, so as to obtain a prediction result of the predicted event based on the surrogate signal prediction result; and inputting the time-series signal to be predicted into the prediction model to obtain the prediction result of the event to be predicted corresponding to the time-series signal output by the prediction model. In this invention, through signal transformation and reconstruction processes, the complexity of multivariate signal prediction is transferred to the signal representation side, thereby improving the accuracy of result prediction based on multivariate time-series signals without changing the mature univariate time-series prediction model. Attached Figure Description
[0018] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.
[0019] Figure 1 This is a flowchart illustrating the prediction and processing method for multivariable time-series signals provided by the present invention.
[0020] Figure 2 This is a schematic diagram of the process provided by the present invention, in which the time series signal to be predicted is input into the prediction model, and the prediction result of the prediction model corresponding to the time series signal to be predicted is obtained.
[0021] Figure 3 This is a schematic diagram of the training and prediction model provided by the present invention.
[0022] Figure 4 This is a schematic diagram of the structure of the predictive processing device for multivariable time-series signals provided by the present invention.
[0023] Figure 5 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation
[0024] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.
[0025] This invention provides a method for predicting and processing multivariate time-series signals. By performing feature weaving and surrogate transformation on the multivariate signals, it adapts them to a high-performance univariate prediction model. Through a learnable signal weaving module, the correlation features between multivariate signals are explicitly extracted. Then, a pair of surrogate signals with an invertible transformation relationship are constructed, transforming the complex multivariate joint prediction problem into the prediction problem of two more regular and predictable single surrogate signals. The high-performance univariate prediction model is then used to process the surrogate signals, and the predicted values of the original multivariate signals are reconstructed through invertible transformation. Essentially, this process seeks a better "representation space" for the multivariate signals while ensuring the invertibility of the transformation, making predictions more efficient and accurate within this space. This invention can be widely applied to scenarios requiring accurate prediction of interrelated multivariate time series signals, such as electricity price forecasting, mineral product price forecasting, power system load forecasting, industrial equipment fault early warning, financial market analysis, weather forecasting, and IoT sensor data analysis.
[0026] Figure 1 This is a flowchart illustrating the prediction and processing method for multivariable time-series signals provided by the present invention.
[0027] The following will combine Figure 1 The process of the prediction and processing method for multivariable time-series signals provided by this invention will be described.
[0028] In an exemplary embodiment of the present invention, combined with Figure 1 As can be seen, the prediction processing method for multivariable time series signals may include steps 110 to 130, and each step will be described below.
[0029] In step 110, the time series signal to be predicted for the event to be predicted is obtained, wherein the time series signal to be predicted includes multiple variables.
[0030] In one embodiment, a time-series signal of the event to be predicted can be obtained. This time-series signal can be historical observation data of multiple related variables (channels) related to the event. The present invention focuses on the technical task of jointly predicting the values of these variables over a future period based on the time-series signal, thus obtaining the prediction result of the event. For ease of explanation, the prediction of electricity prices is used as an example. Correspondingly, the prediction result of the event corresponding to the time-series signal can be the predicted electricity price within a preset future time period. Electricity prices are subject to complex influences from their own historical patterns and numerous external factors, making it a typical multivariate, nonlinear time-series prediction problem.
[0031] In another embodiment, the time-series signal of the event to be predicted can be obtained. These signals constitute a multivariate time-series dataset X, with a length (historical window) of L (e.g., hourly data from the past 7 days) and the number of variables C. Specifically, the time-series signal to be predicted may include historical electricity price sequences from the past L time points. In addition, the time-series signal to be predicted may also include weather data, electricity demand data, renewable energy output data, etc.
[0032] In step 120, a pre-trained prediction model is invoked. The prediction model includes at least a signal weaving module, a surrogate signal construction module, and a univariate prediction model. The signal weaving module is used to obtain a multivariate woven signal corresponding to the time series signal to be predicted based on the time series signal to be predicted. The surrogate signal construction module is used to construct multiple reversible surrogate signals based on the multivariate woven signal and the time series signal to be predicted. The multiple reversible surrogate signals can be linearly combined to reconstruct the time series signal to be predicted. The univariate prediction model is used to obtain the surrogate signal prediction result corresponding to the reversible surrogate signal based on the reversible surrogate signal, so that the prediction result of the predicted event can be obtained based on the surrogate signal prediction result.
[0033] In one embodiment, a pre-trained prediction model can be invoked, wherein the prediction model may include a signal weaving module, a proxy signal construction module, and a univariate prediction model.
[0034] The signal weaving module can be implemented using a lightweight one-dimensional convolutional neural network (CNN) or a multilayer perceptron (MLP). This module is trained to capture and fuse dynamic dependencies across variables from historical data of the aforementioned multiple variables. For example, it can learn complex patterns such as "how a sudden rise in temperature affects peak electricity prices by increasing load demand and thus delaying the peak." Given an input X, the module outputs a multivariate woven signal F(X) of the same dimension, which encodes these interactive features.
[0035] The surrogate signal construction module can contain learnable channel weight vectors w_α and w_β. It receives the original signal X and the woven signal F(X) and constructs multiple, for example, two univariate, invertible surrogate signals, thus obtaining invertible surrogate signals. It should be noted that multiple invertible surrogate signals can be linearly combined to reconstruct the time series signal to be predicted.
[0036] A univariate prediction model can be a high-performance, parameter-frozen pre-trained model, such as a univariate time series base model based on the Transformer architecture (e.g., Informer or Autoformer). This model excels only at handling single time series. For example, it can be used to independently predict two more regular reversible surrogate signals, obtaining surrogate signal prediction results corresponding to the reversible surrogate signals, so that the prediction results of the predicted event can be obtained based on the surrogate signal prediction results.
[0037] In step 130, the time series signal to be predicted is input into the prediction model to obtain the prediction result of the event to be predicted corresponding to the time series signal to be predicted, which is output by the prediction model.
[0038] In one embodiment, the time series signal X to be predicted can be input into the prediction model. During application, the time series signal X is processed by a signal weaving module to obtain a multivariate woven signal F(X) that integrates the complex relationships between electricity price, temperature, load, and photovoltaic power. Further, F(X) and X can be processed by a proxy signal construction module to generate two univariate sequences containing the original information but with simpler forms and clearer trends, which are also known as reversible proxy signals. The reversible proxy signals are then input into the univariate prediction model, and its powerful time series pattern recognition capability is used to output the predicted value of the proxy signal at future times, which is the proxy signal prediction result. Further, the proxy signal prediction result is reconstructed into a prediction result of an event to be predicted.
[0039] In this embodiment, by constructing multiple reversible proxy signals, several comprehensive state indicators are essentially found for complex multivariate systems. These indicators inherit all the information from the original signals, but through weaving and weighting, their sequences often exhibit stronger trends and regularities. This is equivalent to transforming the prediction task from a noisy original data space to a smooth feature space, allowing subsequent powerful univariate models to achieve optimal performance. This improves the accuracy of result prediction based on multivariate time series signals without altering the mature univariate time series prediction model.
[0040] This invention provides a method for predicting multivariate time-series signals, comprising: acquiring a time-series signal to be predicted for an event to be predicted, wherein the time-series signal to be predicted includes multiple variables; calling a pre-trained prediction model, wherein the prediction model includes at least a signal weaving module, a surrogate signal construction module, and a univariate prediction model, wherein the signal weaving module is used to obtain a multivariate woven signal corresponding to the time-series signal to be predicted based on the time-series signal to be predicted; the surrogate signal construction module is used to construct multiple reversible surrogate signals based on the multivariate woven signal and the time-series signal to be predicted; the univariate prediction model is used to obtain a surrogate signal prediction result corresponding to the reversible surrogate signals based on the reversible surrogate signals, so as to obtain a prediction result of the predicted event based on the surrogate signal prediction result; and inputting the time-series signal to be predicted into the prediction model to obtain the prediction result of the predicted event corresponding to the time-series signal to be predicted output by the prediction model. In this invention, through signal transformation and reconstruction processes, the complexity of multivariate signal prediction is transferred to the signal representation side, thereby improving the accuracy of result prediction based on multivariate time-series signals without changing the mature univariate time-series prediction model.
[0041] Figure 2 This is a schematic diagram of the process provided by the present invention, in which the time series signal to be predicted is input into the prediction model, and the prediction result of the prediction model corresponding to the time series signal to be predicted is obtained.
[0042] The following will combine Figure 2 The process provided by the present invention, in which the time series signal to be predicted is input into the prediction model, and the prediction result of the prediction model corresponding to the time series signal to be predicted is obtained, is described.
[0043] In an exemplary embodiment of the present invention, combined with Figure 2 As can be seen, inputting the time series signal to be predicted into the prediction model and obtaining the prediction result of the prediction model outputting the prediction result of the event to be predicted corresponding to the time series signal to be predicted can include steps 210 to 240, and each step will be described below.
[0044] In step 210, the time series signal to be predicted is input to the signal weaving module for multivariate feature weaving and fusion processing to obtain the multivariate woven signal output by the signal weaving module corresponding to the time series signal to be predicted. The multivariate woven signal is a signal that fuses the features of each variable in the time series signal to be predicted.
[0045] In one embodiment, continuing with the previously described example of predicting electricity prices, the time-series signal X to be predicted is input into the signal weaving module. The resulting multivariate woven signal F(X) is not a simple weighted sum of the original signals, but rather a feature representation that incorporates the complex dependencies between electricity prices, temperature, load, and wind power. In another example, the signal weaving module can be implemented using a two-layer one-dimensional convolutional neural network. The convolutional kernel slides along the time dimension, simultaneously acting on all variable channels, thereby automatically learning and fusing collaborative change patterns between variables within a local time window. For example, learning the local dynamic correlation of "evening temperature decrease, load increase, and wind power reduction jointly pushing up electricity prices" yields the multivariate woven signal.
[0046] In step 220, the multivariate woven signal and the time series signal to be predicted are input to the proxy signal construction module for linear combination processing to obtain multiple reversible proxy signals output by the proxy signal construction module.
[0047] In one embodiment, the multivariate woven signal F(X) and the time series signal X to be predicted can be input to the surrogate signal construction module for linear combination processing, thereby obtaining multiple reversible surrogate signals output by the surrogate signal construction module. In one example, the multiple reversible surrogate signals can be two. The following description uses two reversible surrogate signals as an example, wherein the two reversible surrogate signals S_α and S_β can be obtained using the following formulas (1) and (2): S_α = F(X) + w_α ⊙ X (1) S_β = F(X) - w_β ⊙ X (2) Here, ⊙ denotes element-wise multiplication of channels. The key to this transformation lies in its strict invertibility. As long as the weights are non-zero, the original signal X can be accurately reconstructed using (S_α - S_β) / (w_α + w_β). This ensures that the processing result of the surrogate signal can be losslessly mapped back to the original signal space. For ease of explanation, S_α is called the in-phase surrogate signal; S_β is called the out-of-phase surrogate signal.
[0048] As a variation, more reversible surrogate signals can be constructed to capture correlation information through linear relationships with different weight combinations. This scheme can further improve the accuracy of correlation capture, but it increases training complexity and inference time, making it suitable for scenarios with extremely high prediction accuracy requirements and ample resources.
[0049] In this embodiment, a specific linear combination and decomposition process is designed. The woven signal and the original signal are weighted by channel to generate a pair of new single-dimensional surrogate signals (in-phase and out-of-phase), i.e., multiple reversible surrogate signals. The key to this transformation lies in its strict reversibility design, ensuring that information can be converted between the original space and the surrogate space without loss and with determinism. In application, the original multivariate signal, which is difficult to predict directly, can be mapped to a "feature subspace" (corresponding to the reversible surrogate signal) that has stronger regularity and is easier for existing mature models (univariate prediction models) to understand.
[0050] In step 230, each reversible proxy signal is input into the univariate prediction model to obtain the prediction results of each proxy signal corresponding to each reversible proxy signal output by the univariate prediction model.
[0051] In one embodiment, the generated reversible surrogate signals S_α and S_β can be input into a pre-trained, parameter-frozen univariate time series prediction model (such as TimerXL, Sundial), i.e., a univariate prediction model, to obtain prediction results for each surrogate signal corresponding to each reversible surrogate signal. The univariate prediction model utilizes its powerful single-signal time series pattern recognition capability to independently and accurately predict the future trend of each surrogate signal, thus obtaining the prediction results for each surrogate signal. and .
[0052] In step 240, based on the prediction results of each proxy signal, the prediction results of the event to be predicted, which are output by the prediction model and correspond to the time series signal to be predicted, are obtained.
[0053] In another embodiment, the reversible transformation relationship can also be used to transform the prediction results of each proxy signal. and By combining the learned weight vectors w_α and w_β, the predicted values of the original multivariate signal are reconstructed. That is, to obtain the prediction result of the event to be predicted.
[0054] In another exemplary embodiment of the present invention, the following description continues using the previously described embodiments as an example. The multivariate woven signal and the time series signal to be predicted are input to the proxy signal construction module for linear combination processing to obtain multiple reversible proxy signals output by the proxy signal construction module. This can be achieved in the following way: The multivariate woven signal and the time series signal to be predicted are input into the proxy signal construction module, and linear combination processing is performed based on the weight vector to obtain multiple invertible proxy signals output by the proxy signal construction module. The weight vector is obtained by pre-training the prediction model. Based on the prediction results of each proxy signal, the prediction result of the event to be predicted, corresponding to the time series signal to be predicted, output by the prediction model can be obtained in the following way: Based on the prediction results of each proxy signal and the weight vector, reconstruction processing is performed to obtain the prediction result of the event to be predicted corresponding to the time series signal to be predicted, which is output by the prediction model.
[0055] In one embodiment, the prediction model undergoes a pre-training phase before deployment. During application, a large amount of historical multivariate time-series data can be used to train the model. Besides the parameters of the signal weaving module, the weight vectors w_α and w_β in the surrogate signal construction module are also trained. These two weight vectors, along with the parameters of the signal weaving module, can be optimized end-to-end. During application, optimization algorithms can be used to adjust these parameters to minimize the error between the final reconstructed prediction and the true value. After training, w_α and w_β can be fixed to specific values and considered as part of the model's knowledge.
[0056] In another embodiment, the multivariate woven signal and the time series signal to be predicted can be input into the proxy signal construction module and linearly combined based on the weight vectors w_α and w_β to obtain multiple reversible proxy signals output by the proxy signal construction module. The reversible proxy signals can be obtained by referring to the formulas (1) and (2) above.
[0057] In another embodiment, prediction results can also be based on each proxy signal. , The weight vectors w_α and w_β are reconstructed to obtain the prediction result of the event to be predicted, which corresponds to the time series signal to be predicted, as output by the prediction model. In one example, the prediction result of the event to be predicted can be expressed by formula (3): (3) in, This indicates the predicted result for the event to be predicted. This indicates that the channel is divided element by element.
[0058] This embodiment ensures that the transformation from X to (S_α, S_β) is consistent with the transformation from ( , ) arrive The inverse transform of the is a strict, closed mathematical duality. As long as the weight vector is determined, the transformation is lossless and reversible. This mechanism guarantees zero loss of information during the conversion between the surrogate space and the original space, ensuring that all accuracy gains obtained in the easily predictable surrogate space can be completely and accurately mapped back to the original time-series signal that the user cares about, such as electricity prices and other physical variables, thus meeting the stringent requirements of industrial applications for accuracy and reliability.
[0059] In another exemplary embodiment of the present invention, the following description continues using the previously described embodiments. The time series signal to be predicted is input to the signal weaving module for multivariate feature weaving fusion processing to obtain the multivariate woven signal output by the signal weaving module corresponding to the time series signal to be predicted. This can be achieved in the following way: The time series signal to be predicted is input into the signal weaving module, and nonlinear transformation is performed through the fully connected layer in the signal weaving module to extract the static correlation between variable features. Based on static correlation, weaving and fusion processing is performed to obtain the multivariate woven signal output by the signal weaving module, which corresponds to the time series signal to be predicted.
[0060] In one embodiment, at each time point t, the observed values of C variables in the time series signal to be predicted are treated as a vector and nonlinearly transformed through a fully connected layer to extract the static correlation pattern between the variables at that time. Furthermore, the extracted static correlation features at each time point are then woven and fused to obtain a multivariate woven signal corresponding to the time series signal to be predicted. Here, the static correlation pattern can refer to the inherent connection and mutual explanatory relationship between the observed values of multiple variables at the same time segment (time t).
[0061] In another exemplary embodiment of the present invention, the following description continues using the previously described embodiments. The time series signal to be predicted is input to the signal weaving module for multivariate feature weaving fusion processing to obtain the multivariate woven signal output by the signal weaving module corresponding to the time series signal to be predicted. This can be achieved in the following way: The time series signal to be predicted is input into the signal weaving module, and the dynamic correlation between variable features within a preset time window is extracted through the convolutional layer in the signal weaving module. Based on dynamic correlation, weaving and fusion processing is performed to obtain a multivariate woven signal output by the signal weaving module that corresponds to the time series signal to be predicted.
[0062] In one embodiment, each variable in the time series signal to be predicted can be regarded as a channel. A one-dimensional convolutional layer slides along the time dimension, and the convolutional kernel acts on all variable channels simultaneously within a local time window. This automatically learns the collaborative change patterns between multi-variable signals in adjacent time periods and extracts the dynamic correlation between variable features within a preset time window. Furthermore, based on the dynamic correlation, a weaving and fusion process is performed to obtain the multi-variable woven signal output by the signal weaving module, which corresponds to the time series signal to be predicted. The dynamic correlation refers to the dynamic pattern of collaborative evolution between variables over time, rather than the static relationship at a single moment.
[0063] Figure 3 This is a schematic diagram of the training and prediction model provided by the present invention.
[0064] The following will combine Figure 3 The process of training the prediction model provided by this invention will be described.
[0065] In an exemplary embodiment of the present invention, combined with Figure 3 As can be seen, training the prediction model may include steps 310 and 320, which will be described in detail below.
[0066] In step 310, the model parameters of the univariate prediction model in the prediction model are frozen, wherein the model parameters of the univariate prediction model are obtained through pre-training.
[0067] In step 320, based on the loss function, the prediction model of the frozen univariate prediction model is iteratively trained until the function value of the loss function meets the preset requirements, and the trained prediction model is obtained.
[0068] In one embodiment, the univariate prediction model can be a pre-selected high-performance model pre-trained on massive amounts of general univariate time-series data, such as a time-series basic model based on the Transformer architecture. Its model parameters already contain powerful univariate time-series pattern recognition capabilities. A neural network module with a defined structure (such as an MLP or CNN) but whose parameters are yet to be trained can be initialized; its learnable weight vectors w_α and w_β can also be initialized.
[0069] Furthermore, all trainable parameters of the univariate prediction model are set to a non-updateable state. During backpropagation, gradient calculation stops at this module without updating any parameters. The model's parameters are introduced as a pre-trained, fixed knowledge base. The overall prediction model is then trained while the univariate prediction model remains frozen. Continuing with the example of an electricity price prediction model, multivariate time-series data from several years of history, including electricity prices and covariates (temperature, load, wind power), can be used as training samples. The loss function can be either mean squared error (MSE) or mean absolute error (MAE), used to measure the difference between the model's reconstructed future electricity price prediction and the actual value.
[0070] In another embodiment, a batch of training data can be input and sequentially passed through a signal weaving module, a surrogate signal construction module, and a frozen univariate prediction model to reconstruct the prediction result. The loss value is calculated based on the prediction result and the true label. The gradient is backpropagated only to update the parameters of the signal weaving module and the surrogate weight vectors w_α and w_β. The frozen univariate prediction module does not receive gradient updates. The training process is repeated iteratively until the loss function meets a preset requirement, for example, the loss value no longer decreases significantly on the validation set, or a preset number of iterations is reached, thus obtaining a trained prediction model.
[0071] In this embodiment, the general time-series representation capabilities learned by the univariate time-series basic model from massive amounts of data are fully preserved, avoiding catastrophic forgetting that might result from full parameter fine-tuning. By freezing its parameters and reusing them directly, this is equivalent to injecting the core capabilities of a top-level model into a specific prediction task at near-zero marginal cost, maximizing the efficiency of capability transfer. This invention involves zero intrusion into the core prediction model, with minimal parameter increments (requiring only training the weaving module and a small number of weights), achieving plug-and-play functionality. Simultaneously, this method is insensitive to hyperparameters, reducing the difficulty and cost of deployment and optimization in real-world industrial scenarios, and making large-scale applications possible.
[0072] In another exemplary embodiment of the present invention, the following description continues using the previously described embodiments as examples. The loss function includes a regularization term, which is used to constrain the accuracy of the prediction results output by the prediction model during iterative training. Specifically, based on the loss function, the prediction model with frozen univariate prediction model parameters is iteratively trained until the function value of the loss function meets a preset requirement, thus obtaining a trained prediction model. This can be achieved in the following manner: Based on the loss function, the prediction model with frozen univariate prediction model parameters is iteratively trained until the function value of the loss function meets the preset requirements and the accuracy of the prediction result output by the prediction model is greater than or equal to the accuracy of the prediction result output by the univariate prediction model based on the time series signal to be predicted, thus obtaining the trained prediction model.
[0073] In one embodiment, the loss function may include a main loss component and a regularization term. The main loss component measures the error between the model's final multivariate prediction (e.g., predicted electricity price) and the true value, such as mean squared error. The regularization term constrains the accuracy of the predictions output by the model during iterative training. It is understood that training not only continues until the loss function meets a preset requirement, but also must satisfy a strengthened performance guarantee condition: "The accuracy of the predictions output by the model is greater than or equal to the accuracy of the predictions output by the univariate prediction model based on the time-series signal to be predicted." A well-trained prediction model that satisfies these two conditions not only has a convergent loss function, but also rigorously guarantees that its performance is not lower than a specified baseline.
[0074] In the application, a theoretically guaranteed regularization constraint is used when training the signal weaving module and weight parameters. This constraint is derived from the surrogate prediction error. The theoretical derivation of the relationship between the error and the original prediction has the physical significance of ensuring that the surrogate signal transformation sought by the forced optimization process makes the prediction difficulty in its space lower than or equal to that in the original space. This is achieved by adding a regularization term to the loss function, ensuring the performance lower bound of the entire signal processing flow and enhancing the robustness of the method.
[0075] The regularization term can be represented by formula (4): (4) in, The regularization term is represented by C; the number of signals in the multivariate woven signal is represented by C; and i represents the i-th signal. This represents the prediction error based on the prediction result obtained from the i-th in-phase surrogate signal S_α; This represents the prediction error based on the prediction result obtained from the i-th inverted proxy signal S_β; This represents the weight vector corresponding to the i-th signal, which can correspond to w_α mentioned earlier; This represents the weight vector corresponding to the i-th signal, which can correspond to w_β mentioned earlier; This represents the prediction result of the original signal corresponding to the i-th signal; The actual prediction result corresponding to the i-th signal.
[0076] In this embodiment, a regularization term based on the upper bound of the error is introduced to guide signal weaving and proxy generation towards "reducing the difficulty of prediction" from the optimization objective, avoiding instability and overfitting during the training process, and enabling the method to have good convergence and generalization ability.
[0077] As described above, this invention provides a predictive processing method for multivariate time-series signals, pioneering a new paradigm of "signal weaving-surrogate prediction-lossless inverse transformation." Instead of simply modifying or fine-tuning the model, this invention starts from the signal processing front-end, designing a universal signal transformation and reconstruction process. It shifts the complexity of multivariate prediction from the model side to the signal representation side, providing a "preprocessing interface" for the core prediction model. Through a trainable signal weaving module (MLP / CNN), the dynamic association of multivariates, which is difficult to model directly, is transformed into specific network parameters for learning and representation. This is an active and interpretable signal feature engineering approach. The dual surrogate transformation and its reconstruction formula constitute a rigorous and reversible mathematical transformation pair, ensuring no loss of signal information throughout the processing flow. The prediction results accurately correspond to each original physical variable, meeting the industrial application's need for independent prediction results for each variable. The introduction of a regularization term based on the upper bound of the error guides signal weaving and surrogate generation towards "reducing prediction difficulty" from an optimization perspective, avoiding instability and overfitting during training, thus giving the method good convergence and generalization capabilities.
[0078] The prediction processing apparatus for multivariable time series signals provided by the present invention will be described below. The prediction processing apparatus for multivariable time series signals described below can be referred to in correspondence with the prediction processing method for multivariable time series signals described above.
[0079] Figure 4 This is a schematic diagram of the structure of the predictive processing device for multivariable time-series signals provided by the present invention.
[0080] The following will combine Figure 4 The structure of the predictive processing apparatus for multivariable time-series signals provided by the present invention will be described.
[0081] In an exemplary embodiment of the present invention, combined with Figure 4 As can be seen, the predictive processing device for multivariable time series signals may include an acquisition module 410, a calling module 420, and a prediction module 430. Each module will be described in detail below.
[0082] The acquisition module 410 can be configured to acquire the time series signal to be predicted of the event to be predicted, wherein the time series signal to be predicted includes multiple variables; The calling module 420 can be configured to call a pre-trained prediction model, wherein the prediction model includes at least a signal weaving module, a surrogate signal construction module, and a univariate prediction model. The signal weaving module is used to obtain a multivariate woven signal corresponding to the time series signal to be predicted based on the time series signal to be predicted. The surrogate signal construction module is used to construct multiple reversible surrogate signals based on the multivariate woven signal and the time series signal to be predicted, wherein the multiple reversible surrogate signals can be linearly combined to reconstruct the time series signal to be predicted. The univariate prediction model is used to obtain a surrogate signal prediction result corresponding to the reversible surrogate signals based on the reversible surrogate signals, so that a prediction result for the predicted event can be obtained based on the surrogate signal prediction result. The prediction module 430 can be configured to input the time series signal to be predicted into the prediction model to obtain the prediction result of the event to be predicted corresponding to the time series signal to be predicted, output by the prediction model.
[0083] In an exemplary embodiment of the present invention, the prediction module 430 may input the time series signal to be predicted into the prediction model in the following manner to obtain the prediction result of the event to be predicted corresponding to the time series signal to be predicted, output by the prediction model: The time series signal to be predicted is input into the signal weaving module for multivariate feature weaving and fusion processing to obtain the multivariate woven signal output by the signal weaving module corresponding to the time series signal to be predicted. The multivariate woven signal is a signal that fuses the features of each variable in the time series signal to be predicted. The multivariate woven signal and the time series signal to be predicted are input into the proxy signal construction module for linear combination processing to obtain multiple reversible proxy signals output by the proxy signal construction module. Each of the reversible proxy signals is input into the univariate prediction model to obtain the prediction results of each proxy signal corresponding to each of the reversible proxy signals output by the univariate prediction model. Based on the prediction results of each proxy signal, the prediction result of the event to be predicted, which corresponds to the time series signal to be predicted, is obtained from the prediction model output.
[0084] In an exemplary embodiment of the present invention, the prediction module 430 may input the multivariate woven signal and the time series signal to be predicted into the proxy signal construction module for linear combination processing to obtain multiple reversible proxy signals output by the proxy signal construction module: The multivariate woven signal and the time series signal to be predicted are input into the proxy signal construction module, and linear combination processing is performed based on the weight vector to obtain multiple reversible proxy signals output by the proxy signal construction module. The weight vector is obtained by pre-training the prediction model. The prediction module 430 can obtain the prediction result of the event to be predicted, which corresponds to the time series signal to be predicted, output by the prediction model based on the prediction results of each proxy signal, in the following manner: Based on the prediction results of each proxy signal and the weight vector, reconstruction processing is performed to obtain the prediction result of the event to be predicted, which corresponds to the time series signal to be predicted, output by the prediction model.
[0085] In an exemplary embodiment of the present invention, the prediction module 430 can input the time series signal to be predicted into the signal weaving module for multivariate feature weaving fusion processing in the following manner to obtain the multivariate woven signal output by the signal weaving module corresponding to the time series signal to be predicted: The time series signal to be predicted is input into the signal weaving module, and nonlinear transformation is performed through the fully connected layer in the signal weaving module to extract the static correlation between variable features. Based on the static correlation, a weaving and fusion process is performed to obtain a multivariate woven signal output by the signal weaving module that corresponds to the time series signal to be predicted.
[0086] In an exemplary embodiment of the present invention, the prediction module 430 can input the time series signal to be predicted into the signal weaving module for multivariate feature weaving fusion processing in the following manner to obtain the multivariate woven signal output by the signal weaving module corresponding to the time series signal to be predicted: The time series signal to be predicted is input into the signal weaving module, and the dynamic correlation between variable features within a preset time window is extracted through the convolutional layer in the signal weaving module. Based on dynamic correlation, a weaving and fusion process is performed to obtain a multivariate woven signal output by the signal weaving module that corresponds to the time series signal to be predicted.
[0087] In an exemplary embodiment of the present invention, the calling module 420 may pre-train the prediction model in the following manner: Freeze the model parameters of the univariate prediction model in the prediction model, wherein the model parameters of the univariate prediction model are obtained through pre-training; Based on the loss function, the prediction model with its parameters frozen is iteratively trained until the value of the loss function meets a preset requirement, thus obtaining a trained prediction model.
[0088] In an exemplary embodiment of the present invention, the loss function includes a regularization term, which is used to constrain the accuracy of the prediction results output by the prediction model during iterative training. The calling module 420 can perform iterative training on the prediction model with its model parameters frozen based on the loss function, until the function value of the loss function meets a preset requirement, thereby obtaining a trained prediction model: Based on the loss function, the prediction model with the model parameters of the frozen univariate prediction model is iteratively trained until the function value of the loss function meets the preset requirements and the accuracy of the prediction result output by the prediction model is greater than or equal to the accuracy of the prediction result output by the univariate prediction model based on the time series signal to be predicted, thus obtaining the trained prediction model.
[0089] Figure 5 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 5 As shown, the electronic device may include: a processor 510, a communications interface 520, a memory 530, and a communications bus 540, wherein the processor 510, the communications interface 520, and the memory 530 communicate with each other through the communications bus 540. Processor 510 can call logic instructions in memory 530 to execute a prediction processing method for multivariate time series signals. The method includes: acquiring a time series signal to be predicted for an event to be predicted, wherein the time series signal to be predicted includes multiple variables; calling a pre-trained prediction model, wherein the prediction model includes at least a signal weaving module, a proxy signal construction module, and a univariate prediction model; wherein the signal weaving module is used to obtain a multivariate woven signal corresponding to the time series signal to be predicted based on the time series signal to be predicted; the proxy signal construction module is used to construct multiple reversible proxy signals based on the multivariate woven signal and the time series signal to be predicted, wherein the multiple reversible proxy signals can be linearly combined to reconstruct the time series signal to be predicted; the univariate prediction model is used to obtain a proxy signal prediction result corresponding to the reversible proxy signal based on the reversible proxy signal, so as to obtain a prediction result for the event to be predicted based on the proxy signal prediction result; and inputting the time series signal to be predicted into the prediction model to obtain a prediction result for the event to be predicted output by the prediction model corresponding to the time series signal to be predicted.
[0090] Furthermore, the logical instructions in the aforementioned memory 530 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0091] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the prediction processing method for multivariate time-series signals provided by the above methods. The method includes: acquiring a prediction time-series signal of an event to be predicted, wherein the prediction time-series signal includes multiple variables; invoking a pre-trained prediction model, wherein the prediction model includes at least a signal weaving module, a proxy signal construction module, and a univariate prediction model, wherein the signal weaving module is used to obtain a prediction signal based on the prediction time-series signal. The multivariate woven signal corresponding to the time series signal to be predicted; the proxy signal construction module is used to construct multiple reversible proxy signals based on the multivariate woven signal and the time series signal to be predicted, wherein the multiple reversible proxy signals can be linearly combined to reconstruct the time series signal to be predicted; the univariate prediction model is used to obtain the proxy signal prediction result corresponding to the reversible proxy signal based on the reversible proxy signal, so as to obtain the prediction result of the prediction event based on the proxy signal prediction result; the time series signal to be predicted is input into the prediction model to obtain the prediction result of the prediction event corresponding to the time series signal to be predicted output by the prediction model.
[0092] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements a method for predicting and processing multivariate time-series signals provided by the methods described above. This method includes: acquiring a time-series signal to be predicted of an event to be predicted, wherein the time-series signal to be predicted includes multiple variables; and invoking a pre-trained prediction model, wherein the prediction model includes at least a signal weaving module, a proxy signal construction module, and a univariate prediction model, wherein the signal weaving module is used to obtain a multivariate signal corresponding to the time-series signal to be predicted based on the time-series signal to be predicted. The post-weaving signal; the proxy signal construction module is used to construct multiple reversible proxy signals based on the multivariate post-weaving signal and the time series signal to be predicted, wherein the multiple reversible proxy signals can be linearly combined to reconstruct the time series signal to be predicted; the univariate prediction model is used to obtain the proxy signal prediction result corresponding to the reversible proxy signal based on the reversible proxy signal, so as to obtain the prediction result of the prediction event based on the proxy signal prediction result; the time series signal to be predicted is input into the prediction model to obtain the prediction result of the prediction event corresponding to the time series signal to be predicted output by the prediction model.
[0093] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.
[0094] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.
[0095] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A predictive processing method for multivariate time-series signals, characterized in that, The method includes: Obtain the time series signal to be predicted for the event to be predicted, wherein the time series signal to be predicted includes multiple variables; A pre-trained prediction model is invoked, wherein the prediction model includes at least a signal weaving module, a surrogate signal construction module, and a univariate prediction model. The signal weaving module is used to obtain a multivariate woven signal corresponding to the time series signal to be predicted based on the time series signal to be predicted. The surrogate signal construction module is used to construct multiple reversible surrogate signals based on the multivariate woven signal and the time series signal to be predicted, wherein the multiple reversible surrogate signals can be linearly combined to reconstruct the time series signal to be predicted. The univariate prediction model is used to obtain a surrogate signal prediction result corresponding to the reversible surrogate signals based on the reversible surrogate signals, so that a prediction result for the predicted event is obtained based on the surrogate signal prediction result. The time series signal to be predicted is input into the prediction model to obtain the prediction result of the event to be predicted corresponding to the time series signal to be predicted, which is output by the prediction model.
2. The prediction processing method for multivariable time-series signals according to claim 1, characterized in that, The step of inputting the time series signal to be predicted into the prediction model to obtain the prediction result of the event to be predicted corresponding to the time series signal to be predicted, output by the prediction model, includes: The time series signal to be predicted is input into the signal weaving module for multivariate feature weaving and fusion processing to obtain the multivariate woven signal output by the signal weaving module corresponding to the time series signal to be predicted. The multivariate woven signal is a signal that fuses the features of each variable in the time series signal to be predicted. The multivariate woven signal and the time series signal to be predicted are input into the proxy signal construction module for linear combination processing to obtain multiple reversible proxy signals output by the proxy signal construction module. Each of the reversible proxy signals is input into the univariate prediction model to obtain the prediction results of each proxy signal corresponding to each of the reversible proxy signals output by the univariate prediction model. Based on the prediction results of each proxy signal, the prediction result of the event to be predicted, which corresponds to the time series signal to be predicted, is obtained from the prediction model output.
3. The prediction processing method for multivariate time series signals according to claim 2, characterized in that, The process involves inputting the multivariate woven signal and the time series signal to be predicted into the proxy signal construction module for linear combination processing to obtain multiple reversible proxy signals output by the proxy signal construction module, including: The multivariate woven signal and the time series signal to be predicted are input into the proxy signal construction module, and linear combination processing is performed based on the weight vector to obtain multiple reversible proxy signals output by the proxy signal construction module. The weight vector is obtained by pre-training the prediction model. The step of obtaining the prediction result of the event to be predicted, corresponding to the time series signal to be predicted, output by the prediction model based on the prediction results of each proxy signal includes: Based on the prediction results of each proxy signal and the weight vector, reconstruction processing is performed to obtain the prediction result of the event to be predicted, which corresponds to the time series signal to be predicted, output by the prediction model.
4. The prediction processing method for multivariable time-series signals according to claim 2, characterized in that, The step of inputting the time series signal to be predicted into the signal weaving module for multivariate feature weaving and fusion processing to obtain the multivariate woven signal output by the signal weaving module corresponding to the time series signal to be predicted includes: The time series signal to be predicted is input into the signal weaving module, and nonlinear transformation is performed through the fully connected layer in the signal weaving module to extract the static correlation between variable features. Based on the static correlation, a weaving and fusion process is performed to obtain a multivariate woven signal output by the signal weaving module that corresponds to the time series signal to be predicted.
5. The prediction processing method for multivariable time-series signals according to claim 2, characterized in that, The step of inputting the time series signal to be predicted into the signal weaving module for multivariate feature weaving and fusion processing to obtain the multivariate woven signal output by the signal weaving module corresponding to the time series signal to be predicted includes: The time series signal to be predicted is input into the signal weaving module, and the dynamic correlation between variable features within a preset time window is extracted through the convolutional layer in the signal weaving module. Based on dynamic correlation, a weaving and fusion process is performed to obtain a multivariate woven signal output by the signal weaving module that corresponds to the time series signal to be predicted.
6. The prediction processing method for multivariable time-series signals according to claim 1 or 3, characterized in that, The prediction model was pre-trained using the following method: Freeze the model parameters of the univariate prediction model in the prediction model, wherein the model parameters of the univariate prediction model are obtained through pre-training; Based on the loss function, the prediction model with its parameters frozen is iteratively trained until the value of the loss function meets a preset requirement, thus obtaining a trained prediction model.
7. The prediction processing method for multivariable time series signals according to claim 6, characterized in that, The loss function includes a regularization term, which is used to constrain the accuracy of the prediction results output by the prediction model during iterative training. The step of iteratively training the prediction model with its parameters frozen based on the loss function until the value of the loss function meets a preset requirement, to obtain a trained prediction model, includes: Based on the loss function, the prediction model with the model parameters of the frozen univariate prediction model is iteratively trained until the function value of the loss function meets the preset requirements and the accuracy of the prediction result output by the prediction model is greater than or equal to the accuracy of the prediction result output by the univariate prediction model based on the time series signal to be predicted, thus obtaining the trained prediction model.
8. A predictive processing apparatus for multivariable time-series signals, characterized in that, The device includes: An acquisition module is used to acquire the time series signal to be predicted of the event to be predicted, wherein the time series signal to be predicted includes multiple variables; The invocation module is used to invoke a pre-trained prediction model, wherein the prediction model includes at least a signal weaving module, a surrogate signal construction module, and a univariate prediction model. The signal weaving module is used to obtain a multivariate woven signal corresponding to the time series signal to be predicted based on the time series signal to be predicted. The surrogate signal construction module is used to construct multiple reversible surrogate signals based on the multivariate woven signal and the time series signal to be predicted, wherein the multiple reversible surrogate signals can be linearly combined to reconstruct the time series signal to be predicted. The univariate prediction model is used to obtain a surrogate signal prediction result corresponding to the reversible surrogate signals based on the reversible surrogate signals, so that a prediction result for the predicted event can be obtained based on the surrogate signal prediction result. The prediction module is used to input the time series signal to be predicted into the prediction model and obtain the prediction result of the event to be predicted corresponding to the time series signal to be predicted, output by the prediction model.
9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the prediction processing method for multivariable time-series signals as described in any one of claims 1 to 7.
10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the prediction processing method for multivariable time-series signals as described in any one of claims 1 to 7.