A frequency decoupled two-branch adaptive prediction method

Abstract

The application discloses a frequency decoupling double-branch adaptive prediction method. The historical time series data is subjected to variable inversion word segmentation processing, the obtained characteristics are respectively input into a direct mapping path and a deep evolution path, and a prediction result is output after decoding by adaptively fusing the double-path characteristics through a dynamic gating mechanism. The direct mapping path is used to retain high-frequency fluctuation and local mutation information, a shared one-dimensional time series convolution and a multi-flow harmonic evolution module constrained by Sinkhorn are used to extract local dynamic characteristics and stabilize deep feature interaction, and sample-level fusion weights are generated based on global heterogeneity variance proxy, so that the technical problems in the prior art that prediction exists excessive smoothing, high-frequency information attenuation, frequency aliasing, deep feature degradation, unstable cross-variable interaction and insufficient adaptability to different operating states are solved. While retaining high-frequency fluctuation information, the deep time series feature modeling capability is enhanced, and the prediction stability, robustness and accuracy of the model in a complex dynamic scene are improved.

Description

Technical Field

[0001] This invention relates to the field of time series forecasting technology, and to a frequency-decoupled dual-branch adaptive forecasting method, particularly to a time series data forecasting method applicable to complex scenarios such as landslide forecasting, power load forecasting, traffic flow forecasting, meteorological condition forecasting, energy generation forecasting, and industrial equipment condition monitoring. Background Technology

[0002] Multivariate time series forecasting is an important task in time series analysis, widely applicable to scenarios such as landslide monitoring, traffic flow forecasting, energy load forecasting, weather forecasting, industrial sensor monitoring, and financial data analysis. This type of task typically requires predicting the future trends of multiple variables based on their historical observations within a given time window.

[0003] In practical applications, multivariate time series often exhibit significant non-stationarity, manifesting as trend changes, local abrupt changes, high-frequency fluctuations, and dynamic relationships between variables that change over time. Existing prediction models typically suffer from the following shortcomings when processing such data: First, deep feature extraction can easily weaken high-frequency fluctuations and local anomalies in the original sequence, leading to over-smoothing in the prediction results; second, a single-path structure struggles to simultaneously account for both short-term local changes and long-term evolutionary trends; and third, it lacks the ability to adapt to dynamic heterogeneity under different samples and variable states, making it difficult to adaptively adjust the feature fusion method based on the degree of change in the input sequence. Summary of the Invention

[0004] This invention provides a frequency-decoupled dual-branch adaptive prediction method, which solves the technical problems existing in the prior art, such as over-smoothing of predictions, attenuation of high-frequency information, frequency aliasing, degradation of deep features, instability of cross-variable interactions, and insufficient adaptability to different operating states. It enhances the ability to model deep temporal features while retaining high-frequency fluctuation information, and improves the prediction stability, robustness and accuracy of the model in complex dynamic scenarios.

[0005] This invention provides a frequency-decoupled dual-branch adaptive prediction method, comprising:

[0006] Obtain multivariate historical time series data;

[0007] The multivariate historical time series data is subjected to variable inversion segmentation, and the historical observation trajectory of each variable is used as an independent feature unit to obtain variable inversion features.

[0008] The variable inversion feature is input into the direct mapping path, and the direct mapping feature is obtained through linear mapping.

[0009] The variable inversion feature is input into the deep evolution path, and local dynamic features are extracted by a shared variable-wise one-dimensional temporal convolution module. The local dynamic features are then subjected to deep feature interaction by a multi-stream harmonic evolution module to obtain deep evolution features.

[0010] A global heterogeneous variance proxy is constructed based on the variance of the direct mapping features in the global variable dimension, and dynamic gating weights are generated based on the global heterogeneous variance proxy.

[0011] Based on the dynamic gating weights, the direct mapping features and deep evolutionary features are adaptively weighted and fused to obtain adaptive fused features;

[0012] The adaptive fusion features are decoded to obtain future multivariate time series prediction results.

[0013] Specifically, the variable inversion segmentation processing of the multivariate historical time series data, which treats the historical observation trajectory of each variable as an independent feature unit to obtain variable inversion features, includes:

[0014] The multivariate time series data is transformed from a representation with time steps as the main input unit to a representation with variables as the main input unit, so that the complete historical observation trajectory of each variable is treated as an independent feature unit, thus obtaining variable reversal features.

[0015] Specifically, the process involves inputting the variable inversion features into a deep evolution path, extracting local dynamic features through a shared one-dimensional temporal convolution module, and performing deep feature interaction on the local dynamic features through a multi-stream harmonic evolution module to obtain deep evolution features, including:

[0016] The variable inversion feature is input into the deep evolution path, and the variable inversion feature is flattened in the batch dimension and the variable dimension to obtain the flattened feature; then, a one-dimensional temporal convolution operation with shared weights is performed on the flattened feature to obtain the convolutional feature; then, the convolutional feature is restored to the variable dimension representation and mapped to the hidden feature dimension to obtain the local dynamic feature.

[0017] The local dynamic features are subjected to deep feature interaction through a multi-flow harmonization and evolution module, which replicates or maps the local dynamic features along the flow dimension. A series of parallel evolutionary flows are generated; then, intra-flow local aggregation is performed on multiple evolutionary flows to obtain consensus features among them; these consensus features are input into a cross-variable self-attention module to obtain cross-variable self-attention features. ;

[0018] Based on the constructed double-random routing matrix, cross-flow redistribution is performed on multi-flow features, and its expression is as follows: ;in, Indicates the first Characteristics of an evolving flow after cross-flow route reallocation. Represents the first in the double random routing matrix Line number Column routing weights Indicates the number of times before cross-flow redistribution The hidden state characteristics of the evolutionary flow;

[0019] The consensus features are injected into each evolution stream through a gated residual injection mechanism, and their expression is the deep evolution feature. ;in, Indicates the first Characteristics of an evolving flow after cross-flow route reallocation. Presentation layer normalization operation, Indicates the first The learnable gating coefficients corresponding to the evolutionary flow, Represents the Hadama product;

[0020] The obtained evolutionary flow is transformed by a feedforward neural network to obtain deep evolutionary features.

[0021] Specifically, the one-dimensional temporal convolution uses a cyclic filling method.

[0022] Specifically, the process of performing intra-flow local aggregation on multiple evolutionary flows to obtain consensus features among the multiple evolutionary flows includes:

[0023] Let the first The hidden state characteristics of parallel evolutionary flows are: , adopt the first Learnable aggregate weights corresponding to parallel evolutionary flows Weighted aggregation of multiple evolutionary flows:

[0024] ;

[0025] in, Indicates the first The normalized aggregate weights corresponding to the evolutionary flow, Indicates the use of calculating the first Learnable parameters for the aggregation contribution of strip evolution flows. Indicates the number of parallel evolutionary streams. This represents the inter-flow consensus feature obtained by weighted summation of multiple evolutionary flows based on the normalized aggregation weights.

[0026] Specifically, it also includes:

[0027] Set a learnable cross-stream interaction parameter matrix : ;

[0028] For the parameter matrix By performing an exponential mapping, we obtain a non-negative interaction matrix:

[0029] ;

[0030] in, This represents an element-wise exponential mapping;

[0031] The non-negative interaction matrix is ​​normalized through multiple rounds of Sinkhorn iterations; after a preset number of iterations, the double random routing matrix is ​​obtained.

[0032] Specifically, the multi-round Sinkhorn iterative normalization of the non-negative interaction matrix includes:

[0033] In each iteration, column normalization is performed first, followed by row normalization:

[0034] ;

[0035] in, Represents the non-negative interaction matrix in the current iteration. The Middle Line number Column elements, Indicates to The intermediate weights obtained after column normalization of the column are... Indicates to The updated weights are obtained by performing row normalization on the current row and are used as input for the next Sinkhorn iteration.

[0036] Specifically, the step of constructing a global heterogeneous variance proxy based on the variance of the directly mapped features in the global variable dimension, and generating dynamic gating weights based on the global heterogeneous variance proxy, includes:

[0037] The direct mapping feature is denoted as:

[0038]

[0039] Calculate the variance along the variable dimension to obtain a global heterogeneity variance proxy. :

[0040]

[0041] in, This indicates the variance calculation operation. This indicates that variance is calculated along the dimension of the global variable;

[0042] Utilizing the global heterogeneity variance proxy Features with direct mapping Constructing system fluctuation context :

[0043]

[0044] in, This indicates the mean pooling operation;

[0045] The system fluctuation context Input dynamic gating network, output These are the unnormalized weights corresponding to the two paths;

[0046] The dynamic gating weights are obtained by normalizing using the Softmax function:

[0047]

[0048] in, This represents the fusion weight corresponding to the direct mapping path. This represents the fusion weight corresponding to the deep evolution path.

[0049] Specifically, the adaptive weighted fusion of the direct mapping features and deep evolutionary features based on the dynamic gating weights to obtain adaptive fused features includes:

[0050] Based on the dynamic gating weights, the direct mapping features and deep evolutionary features Perform weighted fusion:

[0051]

[0052] in, This represents adaptive fusion features.

[0053] Specifically, it also includes:

[0054] The multivariate historical time series data is subjected to instance normalization to obtain normalized time series data;

[0055] The variable inversion word segmentation process for the multivariate historical time series data includes:

[0056] The normalized time series data is subjected to variable inversion word segmentation processing;

[0057] The process of decoding the adaptive fusion features to obtain future multivariate time series prediction results includes:

[0058] The adaptive fusion features are decoded, and the decoding results are denormalized to obtain the future multivariate time series prediction results.

[0059] One or more technical solutions provided in this invention have at least the following technical effects or advantages:

[0060] 1. Variable inversion segmentation is performed on historical time-series data. The resulting features are input into the direct mapping path and the deep evolutionary path respectively. After adaptive fusion of the dual-path features through a dynamic gating mechanism, the prediction results are decoded and output. The direct mapping path retains high-frequency fluctuations and local mutation information. Shared one-dimensional temporal convolution and a multi-stream harmonic evolutionary module constrained by Sinkhorn are used to extract local dynamic features and stabilize deep feature interactions. Then, sample-level fusion weights are generated based on global heterogeneity variance proxy. This invention achieves frequency decoupling modeling of high-frequency detail information and deep evolutionary information by constructing direct mapping paths and deep evolutionary paths. The direct mapping path retains high-frequency fluctuations, local mutations, and structural anomalies in the original sequence with a shorter information transmission path, which can alleviate the over-smoothing problem that easily occurs in non-stationary predictions of deep models. The deep evolutionary path is used to extract local dynamic features and deep cross-variable interaction features, improving the model's ability to express long-term trends and complex dynamic relationships. Furthermore, this invention also shares local temporal pattern extraction parameters across different variables by using a shared one-dimensional temporal convolution module, enabling the model to extract local dynamic patterns from different variables while reducing parameter redundancy. This structure is beneficial for improving the model's ability to capture short-term fluctuations, periodic segments, and local change patterns. This invention also introduces multiple parallel evolutionary flows through a multi-stream harmonic evolution module, allowing features to be modeled in multiple evolutionary subspaces. A double-random routing matrix is ​​generated through Sinkhorn constraints to constrain information exchange between different evolutionary flows, thereby improving the stability of deep feature interactions and mitigating feature degradation and information loss. This invention also constructs a dynamic gating mechanism through global heterogeneity variance proxy, enabling the model to adaptively adjust the fusion weights of the direct mapping path and the deep evolutionary path based on the variable heterogeneity and fluctuation state of the current input sample. When the input sequence has significant local mutations or high-frequency perturbations, the contribution of the direct mapping path can be increased; when the input sequence mainly reflects stable trends or complex intervariate relationships, the contribution of the deep evolutionary path can be increased, thereby enhancing the model's adaptability to different non-stationary states.

[0061] 2. This invention can combine selective recomputation strategy and operator fusion mechanism to reduce memory usage and computational scheduling overhead during training or inference while keeping the model prediction structure and output method unchanged, thereby improving the deployability of the model in long sequence, multivariate and large-scale data scenarios. Attached Figure Description

[0062] Figure 1 This is an overall flowchart of the frequency decoupling dual-branch adaptive prediction method provided in an embodiment of the present invention;

[0063] Figure 2 A flowchart illustrating the process of obtaining deep evolutionary features in the frequency decoupling dual-branch adaptive prediction method provided in this embodiment of the invention;

[0064] Figure 3 A flowchart illustrating the process of obtaining local dynamic features in the frequency decoupling dual-branch adaptive prediction method provided in this embodiment of the invention;

[0065] Figure 4 This is a schematic diagram of the dynamic gating mechanism in the frequency decoupling dual-branch adaptive prediction method provided in this embodiment of the invention. Detailed Implementation

[0066] The technical solutions in the embodiments of the present invention are to solve the above-mentioned technical problems existing in the prior art, such as... Figure 1 As shown, it includes the following steps:

[0067] S101, acquire multivariate historical time series data, which includes multivariate time series data collected by sensors, monitoring terminals, operational measurement points or operating equipment in landslide monitoring, power systems, transportation systems, meteorological observation systems, energy systems or industrial monitoring systems, such as time series monitoring data of power load, traffic flow, meteorological observation, energy generation, industrial equipment status, etc.

[0068] S102, perform instance normalization on the multivariate historical time series data to obtain normalized time series data; specifically, instance normalization includes: calculating the mean and standard deviation of the historical observation sequence of each variable in each sample within the historical backtracking window, and standardizing the historical observation sequence of the corresponding variable based on the mean and standard deviation.

[0069] S103, perform variable inversion segmentation on the normalized time series data, and use the historical observation trajectory corresponding to each variable as an independent feature unit to obtain variable inversion features; specifically, variable inversion segmentation includes: converting the normalized time series data from a representation with time steps as the main input unit to a representation with variables as the main input unit, so that the complete historical observation trajectory of each variable is input as an independent feature unit into the subsequent network, thereby enhancing the model's ability to model cross-variable relationships.

[0070] S104A: Input the variable inversion feature into the direct mapping path, and obtain the direct mapping feature through linear mapping. The direct mapping feature is used to preserve high-frequency fluctuations, local mutations and structural anomalies in multivariate historical time series data. Specifically, the direct mapping path includes a shared linear projection layer or a lightweight linear projection layer. The shared linear projection layer maps the variable inversion feature from the historical backtracking window dimension to the unified hidden feature dimension and reduces the destruction of high-frequency detail information by deep nonlinear transformation, thus obtaining the direct mapping feature.

[0071] S104B inputs variable inversion features into a deep evolution path, which includes a shared one-dimensional temporal convolution module and at least one multi-stream harmonic evolution module. The shared one-dimensional temporal convolution module shares kernel parameters among variables to extract local dynamic features from different variables. The multi-stream harmonic evolution module performs deep evolutionary modeling of local dynamic features through multiple parallel evolutionary flows. The multi-stream harmonic evolution module includes processes such as intra-flow local aggregation, self-attention modeling, Sinkhorn-constrained cross-flow routing, gated residual injection, feedforward neural network transformation, and cross-flow rerouting. Specifically, the Sinkhorn constraint normalizes the learnable cross-flow interaction matrix into a double-random routing matrix to constrain information exchange between multiple evolutionary flows, avoiding excessive concentration of feature flows or interaction imbalance, thus achieving constrained cross-flow feature interaction and deep evolutionary modeling of local dynamic features to obtain deep evolutionary features.

[0072] The multi-flow harmonic evolution module redistributes cross-flow features across multiple parallel evolution flows based on a dual random routing matrix, and injects attention consensus features into each of the cross-flow redistributed evolution flows through a gated residual injection mechanism to obtain updated multi-flow features.

[0073] In the feedforward neural network stage, a feedforward nonlinear transformation is performed on the updated multi-flow features, and a Sinkhorn-constrained cross-flow routing method corresponding to the cross-variable self-attention stage is adopted to perform cross-flow feature redistribution and gated residual injection on the multi-flow features after the feedforward nonlinear transformation.

[0074] When the deep evolution path includes multiple multi-stream harmonic evolution modules, the multiple multi-stream harmonic evolution modules are stacked and connected in sequence, and the multi-stream features output by the previous multi-stream harmonic evolution module are further modeled in a deep evolutionary manner; after the output of the last multi-stream harmonic evolution module, the multiple parallel evolutionary flows are aggregated to obtain the deep evolutionary features.

[0075] S105, construct a global heterogeneous variance proxy based on the variance of the direct mapping features in the global variable dimension, and generate dynamic gating weights based on the global heterogeneous variance proxy;

[0076] S106, Based on the dynamic gating weights, the direct mapping features and deep evolutionary features are adaptively weighted and fused to obtain adaptive fused features; specifically, the dynamic gating mechanism constructs a global heterogeneity variance proxy based on the direct mapping features to characterize the fluctuation differences between different variables in the current sample and the dynamic state of the system; after inputting the global heterogeneity variance proxy into the dynamic gating network, sample-level fusion weights of the direct mapping path and the deep evolutionary path are generated, enabling the model to dynamically adjust the dual-path contributions according to the non-stationarity of different input samples.

[0077] S107, the adaptive fusion features are decoded, and the decoding result is denormalized to obtain the future multivariate time series prediction result. Specifically, the decoding process includes: inputting the adaptive fusion features into the linear decoding layer, mapping the unified hidden feature dimension to the prediction window length to obtain the prediction sequence at the normalized scale; then, according to the instance normalization parameters recorded in step S102, the prediction sequence at the normalized scale is denormalized to obtain the future multivariate time series prediction result with actual physical meaning.

[0078] In addition, this method also includes a model training step:

[0079] In the model training step, the predicted results of future multivariate time series are compared with the actual future multivariate time series, and the learnable parameters in the direct mapping path, deep evolutionary path, multi-stream harmonic evolutionary module, dynamic gating network, and decoding layer are updated based on the prediction error. The prediction error used in the model training step includes mean squared error, mean absolute error, or a weighted combination of the two.

[0080] This method also includes a selective recalculation strategy. The selective recalculation strategy includes: during the backpropagation process of model training, at least one intermediate activation result among self-attention calculation, Sinkhorn iterative normalization, feedforward neural network calculation and cross-stream feature redistribution is not fully stored. Instead, the corresponding intermediate activation result is recalculated in the backpropagation stage based on the saved input features, learnable parameters and necessary states to reduce memory usage.

[0081] Furthermore, this method also includes an operator fusion mechanism; the operator fusion mechanism includes at least one of the following: fusing adjacent computational operations in one-dimensional temporal convolution, linear projection, and layer normalization; fusing linear transformation and activation function computation in multilayer perceptron; fusing column-wise normalization and row-wise normalization operations in the Sinkhorn iterative normalization process; and fusing weight broadcasting, feature multiplication, and feature summation operations in the dynamic gating fusion process, so as to reduce intermediate tensor read / write and computational scheduling overhead and improve training or inference efficiency.

[0082] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0083] The frequency decoupling dual-branch adaptive prediction method provided in this embodiment of the invention includes:

[0084] Step S110: Obtain multivariate historical time series data:

[0085]

[0086] in, Indicates batch size, Indicates the length of the history lookup window. Indicates the number of variables. For the first... The sample, the first The historical observation sequence of a variable can be represented as:

[0087]

[0088] In this embodiment, the multivariate historical time series data includes at least one of landslide monitoring data, power load, substation measuring points, traffic flow, road speed, meteorological observation indicators, energy generation power, industrial equipment vibration, temperature, pressure or operating status parameters, and includes continuous observation values ​​of multiple variables within a preset historical backtracking window;

[0089] Step S120: Perform variable inversion segmentation on the multivariate historical time series data, taking the historical observation trajectory of each variable as an independent feature unit to obtain variable inversion features;

[0090] This step is explained in detail. Variable inversion segmentation is performed on multivariate historical time series data, treating the historical observation trajectory of each variable as an independent feature unit to obtain variable inversion features, including:

[0091] This method transforms multivariate time series data from a representation primarily based on time steps to a representation primarily based on variables, treating the complete historical observation trajectory of each variable as an independent feature unit, thus obtaining variable inversion features. Specifically, it transforms multivariate time series data from a representation primarily based on time steps to a representation primarily based on variables, treating the complete historical observation trajectory of each variable as an independent feature unit, thus obtaining variable inversion features. Convert to ;in, This represents the number of variable feature units obtained after variable inversion segmentation. The complete historical observation trajectory of each variable is considered as an independent feature unit. That is, for the ... There are 1 variable, and its historical trajectory is as follows:

[0092]

[0093] It is treated as a variable-level token input to the subsequent network.

[0094] By using variable inversion word segmentation, the model can learn the correlation and dynamic heterogeneity relationships between different variables more directly, using variables as the basic modeling unit.

[0095] Step S130: Input the variable inversion feature into the direct mapping path, and obtain the direct mapping feature through linear mapping;

[0096] In this embodiment, a direct mapping path is used to preserve high-frequency information in the input sequence. Instead of undergoing a deep nonlinear evolution process, the direct mapping path uses a lightweight linear mapping to project the variable inversion features onto a unified hidden feature space.

[0097] A direct mapping path can be represented as:

[0098]

[0099] in, This represents the linear mapping layer in the direct mapping path. This indicates a directly mapped feature.

[0100] The role of the direct mapping path is to enable high-frequency fluctuations, local mutations, and structural anomalies in the original sequence to be transmitted to the subsequent fusion stage with less loss through a shorter information transmission path. For sudden peaks, anomalous perturbations, or local changes in non-stationary sequences, the direct mapping path can provide more sensitive feature representation.

[0101] In one specific implementation, the directly mapped feature can be represented as:

[0102]

[0103] in, and These are the learnable weights and bias parameters for the direct mapping path, respectively.

[0104] Step S140: Input the variable inversion features into the deep evolutionary path, extract local dynamic features through a shared one-dimensional temporal convolution module, and perform deep feature interaction on the local dynamic features through a multi-stream harmonic evolutionary module to obtain deep evolutionary features; the deep evolutionary path is used to extract deep evolutionary features such as trend changes, long-term dependencies, and cross-variable dynamic associations. Figure 2 As shown, the deep evolution path includes a shared one-dimensional temporal convolution module and a multi-stream harmonic evolution module.

[0105] This step is explained in detail as follows: The variable inversion features are input into the deep evolution path. Local dynamic features are extracted through a shared, one-dimensional temporal convolution module, and then deep feature interaction is performed on these local dynamic features through a multi-stream harmonic evolution module to obtain deep evolutionary features, including:

[0106] like Figure 3 As shown, the variable inversion features are input into the deep evolution path, and the variable inversion features are flattened in the batch dimension and variable dimension to obtain the flattened features. ;Right now Then, a one-dimensional temporal convolution operation with shared weights is performed on the flattened features to obtain the convolutional features, i.e. ;in, This represents the feature obtained after expanding the inverted features of the variable. This represents a one-dimensional temporal convolution operation. This represents the convolution kernel weights shared by different variables. Indicates a shared bias term. The convolutional features are then represented; these features are then restored to their variable-dimensional representation and mapped to the hidden feature dimensions. This yields local dynamic features, namely The advantage of using shared convolutional kernels is that although different variables have different numerical distributions and physical meanings, their local temporal change patterns may have similar structures, such as short-term rises, short-term falls, periodic segments, local peaks, etc. Shared convolutional kernels can extract cross-variable common local dynamic patterns while reducing the number of parameters.

[0107] To maintain periodic continuity at time boundaries and thus be more suitable for time series data with periodic fluctuations, this embodiment employs a cyclic filling method for one-dimensional temporal convolution.

[0108] The multi-stream harmonic evolution module is used for deep evolutionary modeling of local dynamic features. This module includes multiple parallel evolutionary flows, enabling the same input feature to be expressed and updated in multiple evolutionary subspaces. Through deep feature interaction of local dynamic features using the multi-stream harmonic evolution module, local dynamic features are copied or mapped along the flow dimension. Parallel evolutionary flows yield multi-stream input features, i.e. ;in, Indicates the number of parallel evolution streams. This represents the hidden feature dimension; each evolutionary flow can be viewed as a feature evolution branch, used to capture the dynamic relationships between variables from different perspectives. Through the multi-flow structure, the model can avoid the problem of insufficient expressive power of a single feature channel. In the multi-flow harmonic evolution module, intra-flow local aggregation is performed on multiple evolutionary flows to obtain consensus features among them; these consensus features are then input into the intervariate self-attention module to obtain intervariate self-attention features. ;

[0109] Specifically, intra-flow local aggregation is performed on multiple evolutionary flows to obtain consensus characteristics among the multiple evolutionary flows, including:

[0110] Let the first The hidden state characteristics of the evolutionary flow are: , adopt the first Learnable aggregate weights corresponding to parallel evolutionary flows Weighted aggregation of multiple evolutionary flows:

[0111] ;

[0112] in, Indicates the first The normalized aggregation weights corresponding to the parallel evolutionary flows, and each normalized aggregation weight satisfies , Indicates the use of calculating the first Learnable parameters for the aggregation contribution of strip evolution flows. Indicates the number of parallel evolution streams. This represents the inter-flow consensus feature obtained by weighted summation of multiple evolutionary flows based on normalized aggregation weights.

[0113] By using intra-stream local aggregation, shared information among multiple streams can be extracted while maintaining their differences, providing a stable input for subsequent cross-variable attention modeling.

[0114] The consensus features are input into the intervariate self-attention module to obtain the intervariate self-attention features. ,include:

[0115] Through expression To obtain cross-variable self-attention features ;in, This indicates a cross-variable self-attention operation.

[0116] To improve training stability, this embodiment can further employ residual connections and layer normalization:

[0117]

[0118] in, Presentation layer normalization operation.

[0119] This step can capture long-range dependencies between different variables. For example, in traffic forecasting scenarios, there may be upstream and downstream propagation relationships between different road segments; in weather forecasting scenarios, there may be dynamic coupling relationships between variables such as temperature, humidity, and wind speed. The cross-variable self-attention module can provide expressive capabilities for modeling such relationships.

[0120] Based on the constructed double-random routing matrix, cross-flow redistribution is performed on multi-flow features, and its expression is as follows: ;in, Indicates the first Characteristics of an evolving flow after cross-flow route reallocation. Represents a double random routing matrix The Middle Line number The routing weight of the column is used to characterize the routing weight of the first column. Evolutionary flow direction Information transmission intensity of the evolutionary flow Indicates the number of times before cross-flow redistribution The hidden state characteristics of the evolutionary flow.

[0121] Specifically, a learnable cross-stream interaction parameter matrix is ​​set. :

[0122]

[0123] For parameter matrix By performing an exponential mapping, we obtain a non-negative interaction matrix:

[0124]

[0125] in, This represents an element-wise exponential mapping;

[0126] The non-negative interaction matrix is ​​normalized through multiple rounds of Sinkhorn iterations; after a preset number of iterations, a double random routing matrix is ​​obtained: ;

[0127] in, This indicates the column-by-column summation operation. This indicates a row-by-row summation operation. This represents the number of iteration steps. The double random routing matrix satisfies that the sum of the elements in each row and each column is approximately 1. This matrix is ​​used to constrain information exchange between different evolutionary flows, ensuring that each evolutionary flow can receive information from other flows without causing any flow to over-absorb or over-dominate other flows.

[0128] In this embodiment, the non-negative interaction matrix undergoes multi-round Sinkhorn iterative normalization, including:

[0129] In each iteration, column normalization is performed first, followed by row normalization:

[0130] ;

[0131] in, This represents the non-negative interaction matrix in the current Sinkhorn iteration. The Middle Line number The element of the column represents the first element. Evolutionary flow and the first The intensity of unnormalized interactions between evolutionary flows; Indicates the first The sum of all elements in the list. Indicates to The intermediate weights obtained after column normalization of the column; Indicates the first The sum of all normalized elements in all columns of the row. Indicates to The updated weights are obtained by performing row normalization on the current row and are used as input for the next Sinkhorn iteration. This represents the number of parallel evolutionary flows. After multiple rounds of alternating column and row normalization, the non-negative interaction matrix is ​​normalized to an approximate double random routing matrix, where the sum of the elements in each row and column is approximately 1.

[0132] Consensus features are injected into each evolved stream after cross-stream redistribution using a gated residual injection mechanism, resulting in updated multi-stream features, the expression of which is the deep evolution feature. ;in, Indicates the first The characteristics of the evolved flow after cross-flow route reallocation represent the features of the first flow. The updated features of the evolutionary flow are obtained after cross-flow routing, gated residual injection, and layer normalization; Presentation layer normalization operation, Indicates the first The learnable gating coefficients corresponding to the evolutionary flow, Represents the Hadama product;

[0133] In this embodiment, the gating coefficient It can be obtained through the Sigmoid function:

[0134]

[0135] in, Indicates the first Learnable gating parameters corresponding to the evolutionary flow. This represents the Sigmoid activation function.

[0136] By using gated residual injection, consensus features can be prevented from indiscriminately covering each evolutionary flow, allowing the model to adjust the injection intensity of consensus information according to the state of different evolutionary flows.

[0137] The obtained evolutionary flow is transformed using a feedforward neural network to obtain deep evolutionary features, i.e. .in, A feedforward neural network can include linear transformations, nonlinear activation functions, and linear output layers.

[0138] To further enhance the coordinated evolution capability among multiple flows, this embodiment can employ the Sinkhorn constraint cross-flow routing mechanism again after the feedforward neural network stage to redistribute and update the multi-flow features after the feedforward transformation.

[0139] Through dual cross-flow harmonic evolution in the self-attention and feedforward stages, the model can stably exchange information from different evolutionary flows in the deep feature space, enhancing its ability to model complex dynamic relationships.

[0140] In a preferred embodiment, the deep evolution path includes multiple multi-stream regulation and evolution modules, which are stacked sequentially. The output of the multiple epidemiological investigation and evolution modules serves as the first... Input for each module:

[0141]

[0142] in, This indicates multiple epidemiological investigation and evolution modules.

[0143] After the output of the last multi-stream harmonic evolution module, the multiple evolutionary streams are aggregated to obtain deep evolutionary features:

[0144]

[0145] in, This indicates the number of stacked layers of the multi-flow investigation and evolution modules. It can be a mean aggregation, a weighted aggregation, or a learnable aggregation.

[0146] Deep evolutionary features It is mainly used to express local dynamic patterns, low-frequency trend information, and deep cross-variable interaction relationships in the input sequence.

[0147] Step S150: Construct a global heterogeneous variance proxy based on the variance of the direct mapping features in the global variable dimension, and generate dynamic gating weights based on the global heterogeneous variance proxy;

[0148] This step is explained in detail as follows: a global heterogeneous variance proxy is constructed based on the variance of the directly mapped features along the global variable dimension, and dynamic gating weights are generated based on the global heterogeneous variance proxy, including:

[0149] Directly mapped features are denoted as:

[0150]

[0151] Calculate the variance along the variable dimension to obtain a global heterogeneity variance proxy. :

[0152]

[0153] in, This indicates the variance calculation operation. This indicates that variance is calculated along the dimension of the global variable;

[0154] Global heterogeneity variance surrogate is used to characterize the fluctuation differences between different variables in the current sample. When the differences between variables are large or the local fluctuations are strong, variance surrogate can reflect that the current sample has strong heterogeneity; when the variables change relatively steadily, variance surrogate corresponds to a weak heterogeneity state.

[0155] Using global heterogeneity variance proxy Features with direct mapping Constructing system fluctuation context :

[0156]

[0157] in, This indicates the mean pooling operation;

[0158] like Figure 4 As shown, the system fluctuation context Input a dynamic gating network to generate fusion weights corresponding to the direct mapping path and the deep evolution path.

[0159] Dynamic gating networks can be implemented using lightweight multilayer perceptrons:

[0160]

[0161] in, This represents a multilayer perceptron, with the output... These are the unnormalized weights corresponding to the two paths;

[0162] The dynamic gating weights are obtained by normalizing using the Softmax function:

[0163]

[0164] in, This represents the fusion weight corresponding to the direct mapping path. Let represent the fusion weights corresponding to the deep evolution path, and satisfy:

[0165]

[0166] Through dynamic gating mechanisms, the model can adaptively adjust the dual-path contributions for different input samples. For example, when the input sequence contains strong local mutations or high-frequency perturbations, the dynamic gating network can increase the weights of the direct mapping path; when the input sequence has obvious trends or complex intervariate dependencies, the dynamic gating network can increase the weights of the deep evolutionary path.

[0167] Step S160: Based on the dynamic gating weights, adaptively weightedly fuse the direct mapping features and deep evolutionary features to obtain adaptive fused features;

[0168] This step is explained in detail. Based on the dynamic gating weights, the direct mapping features and deep evolutionary features are adaptively weighted and fused to obtain adaptive fused features, including:

[0169] Based on dynamic gating weights, directly mapped features and deep evolutionary features Perform weighted fusion:

[0170]

[0171] in, This represents adaptive fusion features.

[0172] In the specific implementation, and Broadcasting can be performed on a batch basis, giving each sample an independent path fusion weight. This enables sample-level adaptive fusion, rather than applying a fixed fusion ratio to all samples.

[0173] This fusion approach enables the model to dynamically select more suitable information sources under different non-stationary states, thereby improving the stability and accuracy of predictions.

[0174] Step S170: Decode the adaptive fusion features to obtain the prediction results of future multivariate time series.

[0175] The method of the present invention will be specifically described, which, after obtaining multivariate historical time series data, further includes:

[0176] Instance normalization is performed on multivariate historical time series data to obtain normalized time series data;

[0177] Specifically, for each sample and each variable Calculate its mean within the historical backtracking window:

[0178]

[0179] in, Indicates the sample index. Indicates the time step index. Indicates variable index, Indicates the length of the history lookup window. Indicates the first In the nth sample The variable in the historical backtracking window Observations at each time step Indicates the first In the nth sample The mean of each variable within the historical backtracking window.

[0180] Calculate its standard deviation:

[0181]

[0182] in, To prevent stable terms with a denominator of zero.

[0183] The normalized time series data are as follows:

[0184]

[0185] This yields the normalized multivariate time series data:

[0186]

[0187] Instance normalization reduces scale differences between different samples and variables, mitigating the impact of non-stationary distribution drift on the prediction model. Furthermore, in subsequent prediction output stages, the model can utilize recorded data... and The prediction results are denormalized to restore them to the original numerical scale.

[0188] In this context, variable inversion segmentation processing is performed on multivariate historical time series data, including:

[0189] Perform variable inversion word segmentation on the normalized time series data;

[0190] In this case, the adaptive fusion features are decoded to obtain future multivariate time series prediction results, including:

[0191] The adaptive fusion features are decoded, and the decoding results are denormalized to obtain the prediction results of future multivariate time series.

[0192] Specifically, the adaptive fused features are input into the linear decoding module, mapping the hidden feature dimensions to the prediction window length:

[0193]

[0194] in, This represents the prediction result under the normalized scale.

[0195] Subsequently, inverse normalization is performed based on the mean and standard deviation recorded in the instance normalization step:

[0196]

[0197] in, .

[0198] The goal of this embodiment is to predict the future based on observation data within a historical backtracking window. Multivariate time series at multiple time steps:

[0199]

[0200] in, Indicates the prediction window length.

[0201] In this embodiment, the multivariate historical time series data can be at least one of the following: power system operation data, traffic operation monitoring data, meteorological observation data, landslide monitoring data, energy generation data, financial data, or industrial equipment condition monitoring data. Specifically, power system operation data includes power load, substation measuring points, voltage, current, or power data; traffic operation monitoring data includes road flow, vehicle speed, congestion index, or road occupancy data; meteorological observation data includes temperature, humidity, wind speed, air pressure, or precipitation data; landslide monitoring data includes surface displacement, deep displacement, cumulative displacement, displacement rate, rainfall, groundwater level, seepage water pressure, pore water pressure, soil moisture content, crack width, tilt angle, or acceleration data; and industrial equipment condition monitoring data includes equipment vibration, temperature, pressure, motor current, or operating status parameters. Each variable can correspond to a sensor, a measuring point, a meteorological indicator, a traffic segment, or a business indicator.

[0202] This yields the following multivariate time series prediction results:

[0203]

[0204] The prediction results are on the same numerical scale as the original input data, making them easy to use directly for actual business analysis, scheduling decisions, anomaly warnings, or status assessments.

[0205] The training process of the prediction model will be explained in detail below:

[0206] During the training phase, given the historical input sequence of training samples... and the real future sequence The prediction results are obtained through the method of this invention. Then, a loss function is constructed based on the error between the predicted result and the actual result.

[0207] In one implementation, the loss function uses the mean squared error:

[0208]

[0209] in, This represents the total number of prediction elements involved in the loss calculation, which is the product of the batch size, the prediction window length, and the number of variables. Indicates batch size, Indicates the prediction window length. Indicates the number of variables. Indicates the first In the nth sample The variable in the future The predicted value for each prediction time step. Indicates the first In the nth sample The variable in the future The true value of each predicted time step.

[0210] In another implementation, the loss function uses the mean absolute error:

[0211]

[0212] In another implementation, the loss function can be a weighted combination of mean squared error and mean absolute error:

[0213]

[0214] in, These are the weighting coefficients.

[0215] Based on the loss function, the learnable parameters in the direct mapping path, the shared one-dimensional temporal convolution module, the multi-stream harmonic evolution module, the dynamic gating network, and the decoding module are updated through the backpropagation algorithm.

[0216] After training, the trained model is used to infer new multivariate historical time series data to obtain future time series prediction results.

[0217] The parameter settings in this embodiment are explained in detail below:

[0218] Historical backtracking window length The prediction window length can be set to 96, 192, 336, or 720. It can be set to 24, 48, 96, 192, 336, or 720 to hide the feature dimension. The number of evolutionary flows in the multi-flow harmonization and evolution module can be set to 64, 128, 256, or 512. The number of stacked layers for the multi-flow harmonization and evolution modules can be set to 2, 4, 6, or 8. The number of Sinkhorn iterations can be set to 1 to 6. The dynamic gating network can use a two-layer multilayer perceptron, and its hidden layer dimension can be set to... , or .

[0219] It should be noted that this embodiment further introduces a selective recalculation strategy to reduce GPU memory usage during model training. Specifically, during the forward propagation of model training, this embodiment only saves necessary input features, key normalized states, and learnable parameters; some intermediate activation results are not fully stored. During the backpropagation, when gradients need to be calculated, the corresponding intermediate activation results are recalculated based on the saved input features and model parameters.

[0220] The objects that can be recalculated include at least one of the following:

[0221] 1. Attention score matrix and attention output in the intervariable self-attention module;

[0222] 2. The intermediate normalized matrix during the Sinkhorn iteration process;

[0223] 3. Intermediate activation results in feedforward neural networks;

[0224] 4. Intermediate routing results during the cross-flow feature redistribution process;

[0225] 5. Partial intermediate activation in dynamic gating networks.

[0226] By selectively recompiling, memory usage can be significantly reduced with a small increase in computation, making the model more suitable for training scenarios involving long sequences, multiple variables, and large batches.

[0227] It should also be noted that this embodiment further introduces an operator fusion mechanism to improve model training or inference efficiency. Specifically, this embodiment uses an operator fusion mechanism to fuse multiple adjacent or dependent computational operations. The operator fusion mechanism includes at least one of the following:

[0228] First, the linear mapping and layer normalization operation after shared one-dimensional temporal convolution are fused to reduce the storage of intermediate tensors after convolution output.

[0229] Second, at least some operations in the linear transformation, nonlinear activation function, and output linear transformation of the multilayer perceptron are fused to improve the computational efficiency of the feedforward neural network.

[0230] Third, the column normalization and row normalization operations in the Sinkhorn iteration normalization process are merged to reduce the repeated reading and writing of the routing matrix during the iteration process.

[0231] Fourth, the weight broadcasting, direct mapping feature multiplication, deep evolution feature multiplication, and feature summation operations in dynamic gating fusion are fused to improve the fusion efficiency of dual-path features.

[0232] Fifth, the scaling recovery and output rearrangement operations in the denormalization process are integrated to reduce the additional overhead in the prediction output stage.

[0233] This embodiment, through the aforementioned operator fusion mechanism, can reduce intermediate tensor reads and writes without changing the model prediction logic, thereby improving computational efficiency and reducing deployment costs.

[0234] In this embodiment, the selective recomputation strategy and operator fusion mechanism can be applied to at least one network layer of the multi-stream harmonization and evolution module to reduce the computational storage overhead during model training or inference while keeping the prediction output method unchanged.

[0235] Application scenario examples

[0236] The method of this invention can be used for various non-stationary multivariate time series prediction tasks, such as:

[0237] In landslide prediction scenarios, input variables can include monitoring data such as rainfall, cumulative rainfall, soil moisture content, pore water pressure, groundwater level, slope displacement, tilt angle, crack width, and surface temperature. The direct mapping path can retain information on short-term abrupt changes and abnormal disturbances such as heavy rainfall, sudden displacement increases, and rapid crack propagation. The deep evolution path can model the complex temporal dependencies between rainfall infiltration, soil moisture state, pore water pressure changes, and slope deformation. Multiple flow survey and evolution modules can capture the dynamic coupling and evolutionary propagation characteristics between various monitoring variables. The dynamic gating mechanism can adaptively adjust the contributions of the direct mapping path and the deep evolution path based on the current stable state, slow deformation state, or accelerated deformation state of the slope, thereby improving the accuracy and robustness of landslide displacement prediction, deformation trend prediction, and disaster early warning.

[0238] In power load forecasting scenarios, input variables can include historical load, temperature, humidity, wind speed, date type, and holiday identifiers. Direct mapping paths can preserve information on sudden load changes and short-term disturbances, deep evolution paths can model the complex dependencies between weather factors and load, and dynamic gating mechanisms can adjust the contributions of dual paths based on the current load fluctuation state.

[0239] In traffic flow prediction scenarios, input variables can include traffic flow, speed, and occupancy rates collected from different road segments or sensors. A shared, one-dimensional temporal convolutional module can extract local traffic change patterns across different road segments, while multiple flow tracing and evolution modules can model the dynamic propagation relationships between road segments. This approach is suitable for non-stationary scenarios such as morning and evening rush hours, accident disturbances, and traffic control measures.

[0240] In meteorological forecasting scenarios, input variables can include temperature, humidity, air pressure, wind speed, precipitation, etc. Multi-flow analysis and evolution modules can capture the nonlinear coupling relationships between different meteorological variables, and the dynamic gating mechanism can adaptively adjust the sources of forecast information according to changes in weather conditions.

[0241] In industrial condition monitoring scenarios, input variables can include sensor data such as equipment temperature, vibration, pressure, current, and voltage. Direct mapping paths can preserve abnormal peaks and short-term abrupt changes, while deep evolution paths can capture the long-term evolution trend of equipment operating conditions, thereby improving condition prediction and anomaly early warning capabilities.

[0242] In summary, this invention provides a frequency-decoupled dual-branch adaptive prediction method for non-stationary multivariate time series prediction. This method takes multivariate historical time series data as input and obtains future multivariate time series prediction results through normalization, variable inversion word segmentation, dual-branch feature extraction, dynamic gating fusion, and decoding denormalization processing. The "frequency decoupling" refers to functionally separating and modeling sequence features at different scales of change at the network structure and feature representation levels by using direct mapping paths and deep evolution paths. The direct mapping path focuses on preserving rapidly changing features such as rapid changes, local mutations, short-period fluctuations, and anomalous perturbations, while the deep evolutionary path focuses on extracting local dynamic patterns, cross-variable dependencies, and relatively smooth deep evolutionary features. Local dynamic patterns are extracted through shared one-dimensional temporal convolutions for each variable. Stable cross-flow information exchange between multiple evolutionary flows is achieved through a multi-flow harmonic evolutionary module constrained by a Sinkhorn. Global heterogeneity variance proxy and dynamic gating mechanisms enable adaptive fusion of dual-path features based on the dynamic state of the input samples. Furthermore, selective recomputation strategies and operator fusion mechanisms can be combined to reduce computational and storage overhead during model training or inference. This invention is applicable to various non-stationary multivariate time series prediction scenarios, improving the accuracy, robustness, and stability of prediction results, and is particularly suitable for complex real-world scenarios with trend changes, periodic drift, local mutations, anomalous perturbations, and dynamic changes in variable relationships.

[0243] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0244] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0245] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0246] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0247] Any aspects of this invention not described in detail in the embodiments are well-known techniques to those skilled in the art. Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this invention and not to limit it. Although this invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of this invention without departing from the spirit and scope of this invention, and all such modifications and substitutions should be covered within the scope of the claims of this invention.

Claims

1. A frequency-decoupled dual-branch adaptive prediction method, characterized in that, include: Obtain multivariate historical time series data; The multivariate historical time series data is subjected to variable inversion segmentation, and the historical observation trajectory of each variable is used as an independent feature unit to obtain variable inversion features. The variable inversion feature is input into the direct mapping path, and the direct mapping feature is obtained through linear mapping. The variable inversion feature is input into the deep evolution path, and local dynamic features are extracted by a shared variable-wise one-dimensional temporal convolution module. The local dynamic features are then subjected to deep feature interaction by a multi-stream harmonic evolution module to obtain deep evolution features. A global heterogeneous variance proxy is constructed based on the variance of the direct mapping features in the global variable dimension, and dynamic gating weights are generated based on the global heterogeneous variance proxy. Based on the dynamic gating weights, the direct mapping features and deep evolutionary features are adaptively weighted and fused to obtain adaptive fused features; The adaptive fusion features are decoded to obtain future multivariate time series prediction results.

2. The frequency decoupling dual-branch adaptive prediction method as described in claim 1, characterized in that, The variable inversion segmentation process performed on the multivariate historical time series data uses the historical observation trajectory of each variable as an independent feature unit to obtain variable inversion features, including: The multivariate time series data is transformed from a representation with time steps as the main input unit to a representation with variables as the main input unit, so that the complete historical observation trajectory of each variable is treated as an independent feature unit, thus obtaining variable reversal features.

3. The frequency decoupling dual-branch adaptive prediction method as described in claim 1, characterized in that, The variable inversion feature is input into the deep evolution path, local dynamic features are extracted through a shared one-dimensional temporal convolution module, and deep feature interaction is performed on the local dynamic features through a multi-stream harmonic evolution module to obtain deep evolution features, including: The variable inversion feature is input into the deep evolution path, and the variable inversion feature is flattened in the batch dimension and the variable dimension to obtain the flattened feature; then, a one-dimensional temporal convolution operation with shared weights is performed on the flattened feature to obtain the convolutional feature; then, the convolutional feature is restored to the variable dimension representation and mapped to the hidden feature dimension to obtain the local dynamic feature. The local dynamic features are subjected to deep feature interaction through a multi-flow harmonization and evolution module, which replicates or maps the local dynamic features along the flow dimension. A series of parallel evolutionary flows are generated; then, intra-flow local aggregation is performed on multiple evolutionary flows to obtain consensus features among them; these consensus features are input into a cross-variable self-attention module to obtain cross-variable self-attention features. ; Based on the constructed double-random routing matrix, cross-flow redistribution is performed on multi-flow features, and its expression is as follows: ;in, Indicates the first Characteristics of an evolving flow after cross-flow route reallocation. Represents the first in the double random routing matrix Line number Column routing weights Indicates the first before cross-flow redistribution The hidden state characteristics of the evolutionary flow; The consensus features are injected into each evolution stream through a gated residual injection mechanism, and their expression is the deep evolution feature. ;in, Indicates the first Characteristics of an evolving flow after cross-flow route reallocation. Presentation layer normalization operation, Indicates the first The learnable gating coefficients corresponding to the evolutionary flow, Represents the Hadama product; The obtained evolutionary flow is transformed by a feedforward neural network to obtain deep evolutionary features.

4. The frequency decoupling dual-branch adaptive prediction method as described in claim 3, characterized in that, The one-dimensional temporal convolution uses a cyclic filling method.

5. The frequency decoupling dual-branch adaptive prediction method as described in claim 3, characterized in that, The method of performing intra-flow local aggregation on multiple evolutionary flows to obtain consensus characteristics among the multiple evolutionary flows includes: Let the first The hidden state characteristics of parallel evolutionary flows are: , adopt the first Learnable aggregate weights corresponding to parallel evolutionary flows Weighted aggregation of multiple evolutionary flows: ； ； in, Indicates the first The normalized aggregate weights corresponding to the evolutionary flow, Indicates the use of calculating the first Learnable parameters for the aggregation contribution of strip evolution flows. Indicates the number of parallel evolutionary streams. This represents the inter-flow consensus feature obtained by weighted summation of multiple evolutionary flows based on the normalized aggregation weights.

6. The frequency decoupling dual-branch adaptive prediction method as described in claim 3, characterized in that, Also includes: Set a learnable cross-stream interaction parameter matrix : ； For the parameter matrix By performing an exponential mapping, we obtain a non-negative interaction matrix: ； in, This represents an element-wise exponential mapping; Perform multiple rounds of Sinkhorn iterative normalization on the non-negative interaction matrix; After a preset number of iterations, the double random routing matrix is ​​obtained.

7. The frequency decoupling dual-branch adaptive prediction method as described in claim 6, characterized in that, The step of performing multi-round Sinkhorn iterative normalization on the non-negative interaction matrix includes: In each iteration, column normalization is performed first, followed by row normalization: ； in, Represents the non-negative interaction matrix in the current iteration. Middle Line number Column elements, Indicates to The intermediate weights obtained after column normalization of the column are... Indicates to The updated weights are obtained by performing row normalization on the current row and are used as input for the next Sinkhorn iteration.

8. The frequency decoupling dual-branch adaptive prediction method as described in claim 1, characterized in that, The step of constructing a global heterogeneous variance proxy based on the variance of the direct mapping features in the global variable dimension, and generating dynamic gating weights based on the global heterogeneous variance proxy, includes: The direct mapping feature is denoted as: ； Calculate the variance along the variable dimension to obtain a global heterogeneity variance proxy. : ； in, This indicates the variance calculation operation. This indicates that variance is calculated along the dimension of the global variable; Utilizing the global heterogeneity variance proxy Features with direct mapping Constructing system fluctuation context : ； in, This indicates the mean pooling operation; The system fluctuation context Input dynamic gating network, output These are the unnormalized weights corresponding to the two paths; The dynamic gating weights are obtained by normalizing using the Softmax function: ； in, This represents the fusion weight corresponding to the direct mapping path. This represents the fusion weight corresponding to the deep evolution path.

9. The frequency decoupling dual-branch adaptive prediction method as described in claim 8, characterized in that, The step of adaptively weighting and fusing the direct mapping features and deep evolutionary features according to the dynamic gating weights to obtain adaptive fused features includes: Based on the dynamic gating weights, the direct mapping features and deep evolutionary features Perform weighted fusion: ； in, This represents adaptive fusion features.

10. The frequency decoupling dual-branch adaptive prediction method as described in claim 1, characterized in that, After acquiring the multivariate historical time series data, the following is also included: The multivariate historical time series data is subjected to instance normalization to obtain normalized time series data; The variable inversion word segmentation process for the multivariate historical time series data includes: The normalized time series data is subjected to variable inversion word segmentation processing; The process of decoding the adaptive fusion features to obtain future multivariate time series prediction results includes: The adaptive fusion features are decoded, and the decoding results are denormalized to obtain the future multivariate time series prediction results.

Images