Dual-end time series prediction method based on multi-core adaptive feature and residual connection
By using a multi-kernel adaptive feature and residual connection method, the problems of information loss and training instability in multivariate time series prediction are solved, achieving high-precision prediction results and improving the stability and prediction accuracy of the model.
Patent Information
- Application Number
- CN202611144807.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-30
- Publication Date
- 2026-11-06
AI Technical Summary
Existing deep models suffer from problems in multivariate time series prediction, such as loss of temporal evolution information of exogenous variables, difficulty in fitting complex nonlinear physical processes with linear prediction heads, and the tendency to overfit and unstable training when directly introducing nonlinear networks.
This paper employs a multi-kernel adaptive feature and residual connection approach. By dividing the time series into endogenous and exogenous variables, features are extracted and embedded separately. Chebyshev KAN kernels and one-dimensional convolutional kernels are used to capture high-order nonlinear coupling relationships and local temporal patterns. Adaptive weighted fusion is performed through learnable scaling factors and residual connections. Combined with a Transformer encoder, long-range temporal dependencies and cross-variable causal relationships are captured. Finally, prediction results are generated through residual decoding.
It effectively solves the training instability and overfitting problems introduced by deep nonlinear transformation, improves prediction accuracy and model convergence safety, and reduces mean squared error and mean absolute error.
Smart Images
Figure CN123049459A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of multivariate time series prediction technology, specifically relating to a two-terminal time series prediction method based on multi-kernel adaptive features and residual connections. Background Technology
[0002] Multivariate time series forecasting is a core task in time series data analysis, widely applied in weather forecasting, intelligent transportation, stock price prediction, and industrial control. This task infers future values by analyzing the evolution patterns of multiple variables in historical data, providing a data foundation for scientific decision-making and resource allocation. These forecasts provide valuable information for decision-makers and planners, helping them make accurate predictions and decisions, thereby improving production efficiency, optimizing resource allocation, and reducing risks. Among existing deep learning models, recurrent neural networks (RNNs), temporal convolutional networks (TCNs), and transformer networks are the most commonly used. RNNs suffer from vanishing and exploding gradients when handling long-term dependencies; while they excel at capturing dynamic changes between nearby time points, their ability to model long-term dependencies is relatively poor. TCNs possess parallel computing capabilities and can handle long-term dependencies well, but they rely on fixed-length windows for local modeling, primarily emphasizing local periodic structures, thus neglecting key global periodic patterns. While Transformer has structural advantages in capturing long-range dependencies, real-world time-series signals are often accompanied by noise, abrupt changes, and complex periodic superpositions. Relying solely on point-to-point attention computation makes it difficult to stably and effectively extract generalizable time-series dependencies. Summary of the Invention
[0003] The purpose of this invention is to propose a two-terminal time series prediction method based on multi-kernel adaptive features and residual connections, in order to solve the problems of loss of temporal evolution information of exogenous variables, difficulty of linear prediction heads to fit complex nonlinear physical processes, and overfitting and training instability caused by directly introducing nonlinear networks in existing technologies.
[0004] The method of this invention first decouples the original time series into endogenous variable end and exogenous variable end. For the endogenous variable, a block embedding strategy is adopted to preserve local temporal details, and for the exogenous variable, an inverse embedding strategy is adopted to capture the overall dynamic trend as conditional information. At the same time, a multi-kernel parallel feature extraction mechanism is designed. Chebyshev KAN kernels are used to expand multi-order polynomial mapping along the variable dimension to mine high-order nonlinear coupling relationships between variables. One-dimensional convolution kernels are used to capture local temporal patterns along the time dimension with a sliding window. Finally, the dual-branch outputs are adaptively weighted and fused through a learnable scaling factor. The multi-core fused features are then used as incremental correction terms and residually superimposed with the original endogenous variable sequence before entering the network. The learnable scaling factor is set to zero at the initial training stage to ensure complete alignment between the training starting point and the unenhanced baseline. As training progresses, the nonlinear enhancement gain is gradually released, fundamentally eliminating the training instability risk caused by deep transformations. Based on this, the enhanced endogenous block-level representation and the exogenous variable mapping representation are concatenated and input into a stacked Transformer encoder. A multi-head self-attention mechanism is used to jointly capture long-range temporal dependencies within the sequence and cross-variable causal relationships between internal and external variables. In the result decoding stage, the encoded endogenous variable features are simultaneously fed into the linear main path and the nonlinear auxiliary path to generate preliminary predicted values and incremental correction terms, respectively. These are then residually fused using a scaling factor initialized to zero. Finally, the mean and standard deviation parameters saved in the preprocessing stage are used to inversely normalize the fused prediction results, outputting the final predicted value.
[0005] Compared with existing technologies, this invention, through a complete technical chain of "dual-end separation - multi-core enhancement - secure residual - joint modeling - dual-path decoding", fully explores the complex coupling relationship of time series data while taking into account the convergence security and prediction accuracy of deep models.
[0006] To achieve the above objectives, the present invention provides the following technical solution:
[0007] A two-terminal time series prediction method based on multi-kernel adaptive feature enhancement and residual connection includes the following steps: Step 1, Data Preprocessing and Two-End Separation: After inputting the original time series data into the model, first calculate the mean and standard deviation of its sequence dimension, then normalize the original time series data, and divide the normalized data into endogenous variable sequences and exogenous variable sequences.
[0008] Step 2, Multi-kernel Adaptive Feature Enhancement and Residual Connection: After extracting features from the endogenous variable sequence obtained in Step 1 using a two-branch method, the extracted features are fused and then connected via residual connections to obtain the enhanced endogenous variable feature representation. This includes the following sub-steps: Step 2.1, Multi-core parallel representation: The endogenous variable sequence obtained in Step 1 is simultaneously input into the first feature extraction branch and the second feature extraction branch; Step 2.2: Perform weighted fusion of the output of the first feature branch and the output of the second feature branch using a learnable scaling factor initialized to zero; Step 2.3: The weighted fused features are randomly deactivated and then directly passed to the residual stacking layer along with the endogenous variable sequence obtained in Step 1 through the Identity mapping path to generate the enhanced endogenous variable feature representation. Step 3, endogenous variable partitioning and exogenous variable inversion embedding: The enhanced variable feature representation obtained in step 2 is divided into multiple local blocks along the time dimension, transformed into block-level temporal representations through linear mapping, and then overlaid with positional encoding to preserve the temporal order. Simultaneously, a learnable global representation vector is initialized as the central hub for cross-variable interactions. The exogenous variable sequences obtained in step 1 employ an inverted embedding strategy, i.e., instead of block processing, the complete time series of each exogenous variable obtained in step 1 is mapped as a single variable-level representation vector to preserve the global dynamic information of the external covariates in the time dimension.
[0009] Step 4, Dependency Modeling Based on Transformer Encoder: The block-level temporal representation, global representation vector and variable-level global representation vector obtained in Step 3 are concatenated along the sequence dimension and input into the stacked Transformer encoder to obtain the encoded endogenous variable feature representation.
[0010] This step utilizes the multi-head self-attention mechanism of stacked Transformer encoders to simultaneously capture long-range temporal dependencies within a sequence as well as cross-variable causal relationships between endogenous and exogenous variables, achieving joint representation learning across the time and variable dimensions.
[0011] Step 5, linear mapping and result restoration based on residual dual-path decoding, includes the following sub-steps: Step 5.1: Input the encoded endogenous variable feature representation obtained in Step 4 into both the first and second decoding branches simultaneously; perform residual superposition of the outputs of the first and second decoding branches using a learnable scaling factor initialized to zero to obtain the dual-path fusion prediction result; Step 5.2: Based on the mean and standard deviation parameters obtained in Step 1, perform an inverse normalization operation on the dual-path fusion prediction results obtained in Step 5.1 to obtain the prediction results.
[0012] Furthermore, in step 1, let the input original time series be... The input data is normalized as follows:
[0013] Where μ and σ are the mean and standard deviation of the sequence dimension, respectively.
[0014] Furthermore, in step 2.1, the first feature extraction branch uses a Chebyshev KAN kernel to perform a multi-order polynomial expansion mapping on the endogenous variable sequence obtained in step 1 along the variable dimension. It uses Chebyshev basis functions of different orders to approximate the high-order nonlinear coupling relationship between variables, generating cross-variable global features; the global output feature G of the first feature extraction branch... C Represented as:
[0015]
[0016] in, It is the first Chebyshev basis functions, acting on input variables ; This is the maximum order set in this study. This value strikes a balance between non-linear expressiveness and computational complexity. These are learnable weighting coefficients, whose values are obtained by the network during training through gradient descent optimization, and are used to combine basis functions of different orders and different variables; Indicates the first The value of each input variable Substitute the first The basis function values obtained after the Chebyshev basis functions.
[0017] The second feature extraction branch uses a one-dimensional convolution kernel to perform a sliding convolution operation along the time dimension with a preset window size to extract local temporal patterns in the sequence. The convolution method is as follows:
[0018] in, Given the normalized sequence of endogenous variables as input. For learnable convolution kernel parameters, Sets the default window size.
[0019] The two branches execute in parallel without interfering with each other, capturing complementary information about the data in the variable dimension and the time dimension, respectively.
[0020] In step 2.2, the learnable scaling factor is used as a trainable parameter, and its gradient is calculated based on the loss function and updated by an optimizer (such as Adam or SGD). During the model optimization process, the contribution ratio of the bi-branch features is adaptively adjusted to determine the optimal fusion weight in a data-driven manner.
[0021] In step 2.3, the physical distribution of the endogenous variable sequence obtained in step 1 is retained as the main signal, and the weighted fusion feature obtained in step 2.2 is superimposed on the main signal as an incremental correction term.
[0022] Furthermore, the combined expression for steps 2.2 and 2.3 is:
[0023] in, To enhance the representation of endogenous variable characteristics, For an endogenous variable sequence, For learnable scaling factors initialized to zero, For random inactivation, , These are the outputs of the first feature branch and the second feature branch, respectively.
[0024] Furthermore, in step 5.1, the first decoding branch is a linear main path, which directly maps the encoded endogenous variable feature representation to a preliminary prediction value of a preset prediction length through a fully connected layer, serving as the decoded main signal; The second decoding branch is a nonlinear auxiliary path. It performs a nonlinear transformation on the encoded endogenous variable feature representation through a fully connected layer containing an activation function to generate an incremental correction term. Then, the incremental correction term is multiplied by a learnable scaling factor initialized to zero and superimposed with the preliminary prediction value to generate a dual-path fusion prediction result. Furthermore, in step 5.2, the dual-path fusion prediction result obtained in step 5.1 is multiplied by the standard deviation and the mean is added to restore it to the original data units, and the final prediction value is output.
[0025] Specifically, the calculation formula is as follows:
[0026] in, The normalized predicted value output by the model. and These are the standard deviation and mean of the original sequence recorded in step 1, respectively. This indicates element-wise multiplication.
[0027] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention introduces a dual-branch parallel structure consisting of Chebyshev KAN kernels and one-dimensional convolution kernels into multi-kernel adaptive feature enhancement. It captures cross-variable high-order nonlinear coupling features and local temporal patterns along the variable dimension and time dimension, respectively, to achieve dual-dimensional decoupling and refined reconstruction of the intrinsic physical structure of the data. At the same time, it uses a learnable scaling factor initialized to zero to adaptively weight and fuse the dual-branch outputs. Combined with residual connection and random deactivation mechanism, the enhancement branch is superimposed on the original distribution in a safe incremental manner. This can maintain baseline performance and obtain accuracy gain, effectively solving the training instability and overfitting problems introduced by deep nonlinear transformation.
[0028] 2. This invention employs a block embedding strategy on the endogenous variable side, dividing the enhanced endogenous sequence into multiple local blocks along the time dimension and superimposing positional codes. Simultaneously, it initializes a learnable global representation vector as the central hub for cross-variable information interaction, achieving collaborative representation of local temporal details and global semantics. On the exogenous variable side, an inverse embedding strategy is used to preserve the complete time series, with the variable-level overall representation participating in cross-attention. While fully preserving the temporal dynamic information of external covariates, it avoids redundant noise and computational overhead introduced by fine-grained block division, effectively solving the representation misalignment problem caused by the mismatch of information granularity between endogenous and exogenous variables in traditional methods.
[0029] 3. This invention constructs a decoupled enhancement-mapping architecture by combining multi-core adaptive enhancement at the encoder front end with residual linear decoding at the back end. The front end focuses on improving the representation quality of endogenous variable features, while the back end achieves efficient linear decoding through a flattened fully connected layer. Combined with inverse normalization operations, it restores the model to the true physical dimensions, forming an end-to-end collaborative optimization link. By combining a zero-initialization strategy with a learnable scaling factor, it ensures that the model training starting point is equivalent to the original baseline, eliminating the risk of random fluctuations caused by deep transformations, and achieving a significant reduction in mean square error and mean absolute error on public data benchmarks. Attached Figure Description
[0030] The accompanying drawings are provided to further illustrate the present disclosure and form part of the specification. They are used together with the following detailed description to explain the present disclosure, but do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of the two-end time series prediction method based on multi-kernel adaptive feature enhancement and residual connection of the present invention.
[0031] Figure 2 This is a diagram of the internal structure of the multi-kernel adaptive feature enhancement process in step 2.
[0032] Figure 3 This is a diagram of the post-correction structure based on residual dual-path decoding in step 5.
[0033] The technical solution of the present invention will be further explained and described below with reference to the accompanying drawings and specific embodiments. Detailed Implementation
[0034] It should be noted that, unless otherwise specified, all models and algorithms in this invention, and all technical and scientific terms used in this invention, have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.
[0035] The overall technical concept of this invention is as follows: This invention proposes a two-terminal time series prediction method based on multi-kernel adaptive feature enhancement and residual connection. In the model construction process of this method: the original endogenous variable sequence is normalized for mean and variance, and the normalization parameters are saved for result reconstruction; the normalized endogenous sequence is input into the multi-kernel adaptive feature enhancement module, which captures high-order nonlinear coupling features across variables along the variable dimension using a Chebyshev KAN kernel, and simultaneously extracts local temporal patterns along the time dimension using a one-dimensional convolution kernel. The outputs of the two branches are adaptively weighted and fused with a learnable scaling factor initialized to zero, and then residually superimposed with the original sequence to generate an enhanced endogenous variable feature representation. The enhanced endogenous sequence is divided into multiple local blocks along the time dimension and positional encodings are superimposed, while a learnable global representation vector is initialized; the exogenous variable sequence maintains its complete time length, and a variable-level overall representation is generated using an inverse embedding strategy. The block-level temporal representation, the global representation vector, and the variable-level overall representation are concatenated along the sequence dimension and input into a Transformer encoder, which simultaneously captures long-range temporal dependencies and cross-variable interactions within the sequence through a multi-head self-attention mechanism. The endogenous variable features output by the encoder are input to the flattened fully connected layer and mapped to the preliminary predicted values. The preliminary predicted values are then restored to the original data dimensions through an inverse normalization operation, and the final prediction result is output.
[0036] Following the above technical solution, the specific implementation method of the present invention is given below. It should be noted that the present invention is not limited to the following specific implementation method. All equivalent modifications made based on the technical solution of this application fall within the protection scope of the present invention.
[0037] The present invention provides a two-terminal time series prediction method based on multi-kernel adaptive feature enhancement and residual connection, such as... Figure 1 As shown, it includes the following steps: S1, Dataset Preparation and Model Pre-training: The ETTh1 dataset was used to validate the model's accuracy. The ETTh1 dataset contains sensor data from different power transformers, including seven features such as temperature, current, and voltage, with a sampling rate of one hour. The training, validation, and test sets were divided in a 6:2:2 ratio. The model was trained using the training and validation sets. An end-to-end pre-training strategy using the Adam optimizer and cosine annealing learning rate scheduling was employed. Training was terminated early when the validation set loss did not decrease for several consecutive rounds, and the model parameters with the minimum validation set loss were saved as the optimal model.
[0038] S2, Data Preprocessing and Normalization: Obtain the original endogenous variable sequence and the original exogenous variable sequence from the dataset, perform mean and variance normalization on the endogenous variables to eliminate dimensional differences, and save the normalization parameters for subsequent result restoration.
[0039] S3, Pre-feature Refinement Based on Multi-Kernel Adaptive Network: The normalized endogenous variable sequence is input into the multi-kernel adaptive feature enhancement module, that is, the output of the first feature branch and the second feature branch are input. The Chebyshev KAN kernel captures cross-variable high-order coupling features along the variable dimension, and at the same time, the one-dimensional convolutional kernel extracts local temporal patterns along the time dimension. The outputs of the two branches are adaptively weighted and fused by a learnable scaling factor initialized to zero, and then residual superposition is performed with the original sequence to generate the enhanced endogenous variable feature representation.
[0040] S4, Endogenous variable segmentation and exogenous variable inversion embedding: The enhanced endogenous variable feature representation is segmented into multiple local blocks along the time dimension and positional encoding is superimposed to initialize a learnable global representation vector; the exogenous variable sequence is generated into a variable-level representation vector using an inversion embedding strategy.
[0041] S5, hierarchical dependency modeling based on Transformer encoder: block-level temporal representation, global representation vector and variable-level representation vector are concatenated and input into Transformer encoder. Temporal dependencies and cross-variable interaction relationships are captured synchronously through multi-head self-attention to obtain the encoded endogenous variable feature representation.
[0042] S6, Linear mapping and result restoration based on residual dual-path decoding: The endogenous variable feature representation of the encoder output is simultaneously input into the linear main path and the auxiliary path containing the nonlinear activation function. The output of the auxiliary path is multiplied by a learnable scaling factor initialized to zero and then superimposed with the output of the main path to generate the dual-path fusion prediction result. Finally, the mean and standard deviation parameters saved in S2 are used to perform inverse normalization restoration and output the final prediction value.
[0043] This invention evaluates the results of a two-terminal time series prediction model based on multi-kernel adaptive feature enhancement and residual connection by using mean squared error (MSE) and mean absolute error (MAE).
[0044] The mean squared error (MSE) is the average of the sum of squares of the differences between the predicted and actual values from a regression model. Its formula is:
[0045] Where D represents the number of multivariate time series features, T represents the previous time of the multivariate time series for which prediction is being made, t represents the prediction window length, and y represents the true value from time step T+1 to T+t. This represents the model's predicted value from time step T+1 to T+t.
[0046] The Mean Absolute Error (MAE) is the average of the absolute values of the differences between the predicted and actual values of a regression model. Its formula is:
[0047] Where D represents the number of multivariate time series features, T represents the previous time of the multivariate time series being predicted, t represents the prediction window length, and y represents the true value from time step T+1 to T+t. This represents the model's predicted value from time step T+1 to T+t.
[0048] The experimental results on the ETTh1 dataset are shown in Table 1, where 96, 192, 336, and 720 represent the experimental prediction window sizes. All experiments were conducted on a computer running Windows 11 with a 12th Gen Intel(R) Core(TM) i7-12700H 2.30 GHz processor. The baseline models selected were Informer, Pyformer, and LogTrans (based on Transformer models), MICN (based on TCN), and GRU (based on RNN). The experimental results are shown in Table 1. Table 1 Experimental results on the ETTh1 dataset
[0049] The specific embodiments of this disclosure have been described in detail above with reference to the accompanying drawings. However, this disclosure is not limited to the specific details in the above embodiments. Within the scope of the technical concept of this disclosure, various simple modifications can be made to the technical solutions of this disclosure, and these simple modifications all fall within the protection scope of this disclosure.
[0050] It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction. In order to avoid unnecessary repetition, this disclosure will not describe the various possible combinations separately.
[0051] Furthermore, various different embodiments of the present invention can be combined in any way, as long as they do not violate the spirit of this disclosure, they should also be regarded as the content disclosed in this invention.
Claims
1. A two-terminal time series prediction method based on multi-kernel adaptive feature enhancement and residual connection, characterized in that, Includes the following steps: Step 1, Data Preprocessing and Two-End Separation: After inputting the original time series data into the model, first calculate the mean and standard deviation of its sequence dimension, then normalize the original time series data, and divide the normalized data into endogenous variable sequences and exogenous variable sequences; Step 2, Multi-kernel Adaptive Feature Enhancement and Residual Connection: After extracting features from the endogenous variable sequence obtained in Step 1 using a two-branch method and fusing them, the enhanced endogenous variable feature representation is obtained through residual connection, including the following sub-steps: Step 2.1, Multi-core parallel representation: The endogenous variable sequence obtained in Step 1 is simultaneously input into the first feature extraction branch and the second feature extraction branch; Step 2.2: Perform weighted fusion of the output of the first feature branch and the output of the second feature branch using a learnable scaling factor initialized to zero; Step 2.3: The weighted fused features are randomly deactivated and then directly passed to the residual stacking layer along with the endogenous variable sequence obtained in Step 1 through the Identity mapping path to generate the enhanced endogenous variable feature representation. Step 3, Endogenous variable segmentation and exogenous variable inversion embedding: The enhanced variable feature representation obtained in Step 2 is segmented into multiple local blocks along the time dimension, converted into block-level temporal representations through linear mapping, and positional encoding is superimposed to preserve the temporal order. At the same time, a learnable global representation vector is initialized as the cross-variable interaction center; the complete time series of each exogenous variable obtained in Step 1 is mapped as a whole into a variable-level representation vector. Step 4, Dependency modeling based on Transformer encoder: The block-level temporal representation, global representation vector and variable-level global representation vector obtained in Step 3 are concatenated along the sequence dimension and input into the stacked Transformer encoder to obtain the encoded endogenous variable feature representation; Step 5, linear mapping and result restoration based on residual dual-path decoding, includes the following sub-steps: Step 5.1: Input the encoded endogenous variable feature representation obtained in Step 4 into both the first and second decoding branches simultaneously; perform residual superposition of the outputs of the first and second decoding branches using a learnable scaling factor initialized to zero to obtain the dual-path fusion prediction result; Step 5.2: Based on the mean and standard deviation parameters obtained in Step 1, perform an inverse normalization operation on the dual-path fusion prediction results obtained in Step 5.1 to obtain the prediction results.
2. The two-terminal time series prediction method based on multi-kernel adaptive feature enhancement and residual connection as described in claim 1, characterized in that, In step 2.1, the first feature extraction branch uses a Chebyshev KAN kernel to perform a multi-order polynomial expansion mapping on the endogenous variable sequence obtained in step 1 along the variable dimension. It uses Chebyshev basis functions of different orders to approximate the high-order nonlinear coupling relationship between variables, generating cross-variable global features. The global output feature G of the first feature extraction branch... C Represented as: in, It is the first Chebyshev basis functions, acting on input variables ; It is the maximum set order. ; These are learnable weighting coefficients; Indicates the first The value of each input variable Substitute the first The basis function values obtained after the Chebyshev basis functions; The second feature extraction branch uses a one-dimensional convolution kernel to perform a sliding convolution operation along the time dimension with a preset window size to extract local temporal patterns in the sequence. The convolution method is as follows: in, Given the normalized sequence of endogenous variables as input. For learnable convolution kernel parameters, Sets the default window size.
3. The two-terminal time series prediction method based on multi-kernel adaptive feature enhancement and residual connection as described in claim 1, characterized in that, In step 1, let the input original time series be... The input data is normalized as follows: Where μ and σ are the mean and standard deviation of the sequence dimension, respectively.
4. The two-terminal time series prediction method based on multi-kernel adaptive feature enhancement and residual connection as described in claim 1, characterized in that, The combined expression for steps 2.2 and 2.3 is: in, To enhance the representation of endogenous variable characteristics, For an endogenous variable sequence, For learnable scaling factors initialized to zero, For random inactivation, , These are the outputs of the first feature branch and the second feature branch, respectively.
5. The two-terminal time series prediction method based on multi-kernel adaptive feature enhancement and residual connection as described in claim 1, characterized in that, In step 5.1, the first decoding branch is a linear main path, which directly maps the encoded endogenous variable feature representation to a preliminary prediction value of a preset prediction length through a fully connected layer, and serves as the decoding main signal; The second decoding branch is a nonlinear auxiliary path. It performs a nonlinear transformation on the encoded endogenous variable feature representation through a fully connected layer containing an activation function to generate an incremental correction term. Then, the incremental correction term is multiplied by a learnable scaling factor initialized to zero and superimposed with the preliminary prediction value to generate a dual-path fusion prediction result.
6. The two-terminal time series prediction method based on multi-kernel adaptive feature enhancement and residual connection as described in claim 1, characterized in that, In step 5.2, the dual-path fusion prediction result obtained in step 5.1 is multiplied by the standard deviation obtained in step 1 and the mean is added to restore it to the original data units. The final predicted value is then output, and the calculation formula is as follows: in, The normalized predicted value output by the model. and These are the standard deviation and mean of the original sequence, respectively. This indicates element-wise multiplication.