Time sequence prediction method and system based on dual-channel multi-path interaction and disturbance optimization

Abstract

The invention discloses a time sequence prediction method and system based on dual-channel multi-path interaction and disturbance optimization, and relates to the technical field of time sequence prediction and artificial intelligence. In order to solve the problems of insufficient representation, gradient degradation and insufficient adaptive capacity to a complex disturbance environment in a high-dimensional scene in an existing time sequence prediction method, the method comprises the following steps: constructing two channels, and respectively capturing a time dimension long-time dependence feature and a feature dimension static structure feature for an initial input feature; sending the information to a multi-path interaction module to obtain fused multi-dimensional interaction features and potential components; fitting the potential component with the initial input feature, and taking a calculated difference value as a residual feature; and the multi-dimensional interaction features and the residual features are sent to a prediction head for prediction, and a prediction sequence of the target time interval is obtained. The method is mainly used for predicting the time sequence in a high-dimensional scene.

Description

Technical Field

[0001] This invention relates to the fields of time series forecasting and artificial intelligence, and in particular to a time series forecasting method and system based on dual-channel multi-path interaction and perturbation optimization. Background Technology

[0002] With the widespread deployment of industrial control systems, energy management systems, traffic monitoring networks, and smart terminal devices, operational decisions in many industries heavily rely on the predictive capabilities of time series data. For example, power load systems need to obtain trends in temperature and equipment operating status in advance; traffic monitoring platforms rely on future estimates of traffic flow sequences for signal regulation; and meteorological systems need to predict environmental changes based on historical multivariate sequences. In these scenarios, time series typically exhibit characteristics such as multivariability, high-frequency sampling, and fluctuations across multiple time scales, making models prone to decreased accuracy or instability when handling long sequences, strong disturbances, and multidimensional correlations.

[0003] Existing time series forecasting methods typically face the following limitations: On the one hand, traditional models mostly handle all dimensions and time steps simultaneously through a single structure, which cannot independently express "time dependency patterns" and "variable structure patterns", making it difficult to capture long-term dependencies and failing to fully model variable relationships; in high-dimensional scenarios, they are prone to insufficient representation or gradient degradation problems.

[0004] On the other hand, while some methods introduce multi-branch structures, the branches often lack independent structural responsibilities, making it difficult to effectively decouple long-term trend patterns, short-term disturbance patterns, and variable coupling patterns, resulting in insufficient fusion capabilities. Furthermore, complex deep structures have high hardware dependencies, making them difficult to deploy in industrial edge environments.

[0005] Furthermore, real-world data acquisition systems commonly suffer from noise, missing data points, and sudden fluctuations. Existing prediction models largely rely on standard training methods and lack robust optimization strategies for perturbation environments, making them susceptible to noise-induced performance degradation in real-world applications.

[0006] Therefore, there is a need for a time series forecasting method and system that can simultaneously express cross-time patterns, cross-variable structures, and multi-level fluctuation characteristics, and has strong anti-disturbance capabilities and efficient deployment performance. Summary of the Invention

[0007] To address the problems of insufficient representation, gradient degradation, and insufficient adaptability to complex perturbation environments in existing time series forecasting methods in high-dimensional scenarios, this invention proposes a time series forecasting method and system based on dual-channel multi-path interaction and perturbation optimization that can simultaneously express cross-time patterns, cross-variable structures, and multi-level fluctuation characteristics, and has strong anti-perturbation capabilities and efficient deployment performance.

[0008] The time series prediction method based on dual-channel multi-path interaction and perturbation optimization described in this invention includes the following steps: S1. Obtain multivariate time series data, perform layer normalization and embedding representation on the original input data to form initial input features; S2. Construct a dual-channel system to perform time path modeling and feature path modeling on the initial input features, respectively capturing long-term dependent features in the time dimension and static structural features in the feature dimension. S3. The long-term time-dependent features and static structural features of the feature dimension are fed into the multi-path interaction module to obtain the fused multi-dimensional interaction features. S4. Input the multidimensional interaction features into the regularized latent component regression model to obtain the latent components; S5. Fit the latent components to the initial input features, and use the calculated difference as the residual features; S6. Generate adversarial perturbations on the initial input features and train them together with the original input data to optimize the target; S7. The multidimensional interactive features and residual features are fed into the prediction head for prediction to obtain the prediction sequence for the target time interval.

[0009] Further: In S1, the multivariate time series data undergoes layer normalization and embedding representation of the original input data to form initial input features, including: S11. Construct input samples using a sliding window approach, taking a historical window of length L as the input features; S12. Perform layer normalization on the input features to ensure that the values ​​of each time step remain stable at the same feature scale. S13. Map the feature vector corresponding to each time step to a unified hidden space dimension to obtain the hidden space features, and use the hidden space features as the initial input features.

[0010] Further: In S2, the specific steps of constructing a dual-channel system and performing temporal path modeling and feature path modeling on the initial input features to capture long-term dependent features in the time dimension and static structural features in the feature dimension respectively include: S21. For the time path, the initial input features are sent into the main channel, and a multi-head self-attention mechanism is used to model the dependency relationship between different times to obtain the long-term dependency features of the time dimension of the time path output. S22. For the feature path, the initial input features are sent into the feature channel, transposed in the feature dimension, and a multi-head self-attention mechanism is used to model the static coupling relationship between different variables to calculate the static structural features of the feature dimension output by the feature path.

[0011] Further: In S3, the specific process of feeding the long-term time-dependent features and the static structural features of the feature dimension into the multi-path interaction module to obtain the fused multi-dimensional interaction features is as follows: S31. Input the long-term time-dependent features, static structural features, and initial input features into the multi-path interaction module. S32. Perform a non-linear weighted combination of the three source features according to the trained weights to obtain the weighted combination result; S33. Input the weighted combination result into the gated fusion layer, and perform path adaptive fusion through dynamic gating to obtain the fused multidimensional interactive features.

[0012] Further: In S4, the specific steps for inputting the multidimensional interaction features into the regularized latent component regression model to obtain the latent components include: S41. Centralize and standardize the multi-dimensional interactive features; S42. By performing covariance analysis and singular value decomposition on the input matrix and target matrix after centering and standardization, the potential direction that best matches the target matrix is ​​obtained, and the projection weights are obtained. S43. Project the features into a low-dimensional space through the latent direction to obtain the latent component score vector.

[0013] Further: In S5, the latent components are fitted with the initial input features, and the difference is calculated as the residual features; S51. Reconstruct the initial input features based on the latent component score vector by fitting a mapping; S52. Subtract the reconstruction result from the initial input features to obtain the residual features; S53. The multidimensional interactive features and the residual features are concatenated along the feature dimension to obtain the residual enhancement representation.

[0014] Further: In S6, the specific steps for optimizing the objective are as follows; S61. The initial input features, the multidimensional interaction features, and the residual features are fed into the prediction head to obtain a perturbation-free prediction output; mean square error is used as the perturbation-free loss function. S62. Input the perturbation sample, the multidimensional interaction feature and the residual feature into the prediction head to obtain the adversarial prediction output; use the mean squared error as the adversarial loss function. S63. Construct a joint loss based on the prediction error under unperturbed samples and the prediction error under perturbed samples; and update the model parameters by minimizing the joint loss.

[0015] Further: In S7, the specific steps for feeding the multidimensional interaction features and residual features into the prediction head for prediction to obtain the prediction sequence for the target time interval include: S71. Scale recovery is performed on the prediction results based on the scaling and bias parameters of the corresponding variables during normalization. S72. Perform a linear transformation on the restored prediction results that is the opposite of the normalization process, and map the prediction results from the normalization space back to the original variable space. S73. Reconstruct the predicted sequence according to the time step order to obtain the prediction result aligned with the original time series in the time dimension.

[0016] The time series prediction system described in this invention, used to implement the time series prediction method based on dual-channel multi-path interaction and perturbation optimization, includes a raw time series data processing module, a dual-channel feature extraction module, a multi-path interaction module, a latent component regression module, a fitting module, a perturbation optimization module, and a prediction module. The raw time series data processing module is used to acquire raw time series data and preprocess it to form initial input features; The dual-channel feature extraction module is used to construct the main channel and the feature channel, and to perform time path modeling and feature path modeling based on the initial input features, thereby capturing long-term dependent features in the time dimension and static structural features in the feature dimension, respectively. The multi-path interaction module is used to fuse the long-term time-dependent features and the static structural features of the feature dimension to obtain the fused multi-dimensional interaction features. The latent component regression module is used to input the multidimensional interaction features into the regularized latent component regression model to obtain latent components; The fitting module is used to fit the latent components with the initial input features and use the calculated difference as the residual features; The perturbation optimization module is used to generate adversarial perturbations on the initial input features and to participate in training together with the original input data to optimize the target. The prediction module is used to feed the multidimensional interactive features and residual features into the prediction head for prediction, so as to obtain the prediction sequence of the target time interval.

[0017] The beneficial effects of this invention are: This invention proposes a prediction method for complex time series data. By constructing a dual-channel multi-path structure, a latent component dimensionality reduction mechanism, a residual compensation system, and a perturbation optimization training strategy, it achieves high-precision prediction of multivariate, multi-scale, and multi-mode mixed time series. This invention systematically designs various aspects, including data structure representation, feature decoupling methods, feature interaction strategies, and training stability, representing a significantly different technical approach from the single-path structure or fixed-mode modeling methods of traditional time series models.

[0018] The overall goal of this invention is to build a time series prediction system that can adaptively handle long-term dependencies, intervariate relationships and local perturbation features, and is both lightweight and highly robust, so that it can operate stably in multiple industry scenarios such as energy, transportation, meteorology and industrial monitoring.

[0019] This invention constructs a dual-channel architecture, focusing independently on modeling time-dependent patterns and variable structure patterns, thereby effectively improving the ability to capture long-term dependencies and express multi-dimensional variable relationships. Simultaneously, a multi-path interaction mechanism is introduced to decouple and fuse long-term trends, short-term fluctuations, and variable coupling relationships at different time scales, further enhancing the model's ability to characterize multi-level fluctuation features. Furthermore, this method incorporates a perturbation optimization strategy, significantly improving the model's robustness in real-world scenarios through simulation and adversarial training against noise, missing data, and sudden fluctuations. This innovation not only addresses the limitations of traditional methods in high-dimensional scenarios but also ensures model efficiency, making it better suited for deployment in industrial edge environments. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall process of the method of the present invention; Figure 2 This is a structural diagram of the method of the present invention; Figure 3 This is a schematic diagram illustrating the prediction performance of the method of the present invention on the ETTh1 dataset; Figure 4 This is a schematic diagram of the predicted residual distribution of the method of the present invention; Figure 5 This is a schematic diagram of the geometric approach to the projection process of potential components. Detailed Implementation

[0021] The following are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. The embodiments described below are only for explaining the present invention and should not be construed as limiting the present invention. The scope of protection of the present invention should be determined by the scope of the claims. The embodiments of the present invention are described in detail below. In order to facilitate the description of the present invention and simplify the description, the technical terms used in the specification of the present invention should be interpreted broadly, including but not limited to conventional alternatives not mentioned in this application, and including both direct and indirect implementation methods.

[0022] Example 1 Combination Figures 1-5 This embodiment describes a time series forecasting method based on dual-channel multi-path interaction and perturbation optimization, comprising the following steps: S1. Obtain multivariate time series data and perform input preprocessing; The original time series data is used as input, denoted as X(B,D,T). In industrial and practical application scenarios, X can be a multivariate sequence such as load, temperature, vibration, traffic flow, weather, and multi-sensor data. Here, B is the sample batch size, D is the original feature dimension, and T is the time length.

[0023] This step may include: S11: Construct the input sample; Using a sliding window approach, a historical window of length L is used as the input feature, and the next H time steps are used as the prediction target, generating sample pairs in the following form: ; Where t is the starting time step index of the sliding window; Let L represent the historical input subsequence of length L starting from time step t, where This represents the multivariate feature vector corresponding to the original input sequence at the i-th time step; Let H represent the target prediction subsequence of length H starting from the (t+L)th time step, where This represents the target variable value or target feature vector at the corresponding time step. By sliding along the time dimension t, multiple overlapping input-output sample pairs can be constructed, enabling the model to fully learn the continuous change patterns and temporal dependencies of the time series during training.

[0024] S12, Perform multi-level normalization; To improve the model's adaptability to multivariate sequences, multi-level normalization is performed on the input features. This multi-level normalization includes global standardization, local window normalization, and reversible maximum / minimum scaling, so that different variables are within a comparable scale range.

[0025] S13. Through transformations such as linear projection or one-dimensional convolution, the feature vector corresponding to each time step is mapped to a unified hidden space dimension F to mitigate the impact of differences in the original data distribution on the modeling process; the hidden space features obtained by the above transformation are denoted as... This is called the initial input feature (i.e., the embedding representation of the standardized input feature sequence in the hidden space). The processed samples are uniformly denoted as the initial input features. This serves as the foundational input for subsequent time path modeling and feature path modeling.

[0026] S2. The initial input features are fed into a dual-channel system for time dimension modeling and feature dimension modeling, respectively. This embodiment constructs a dual-channel approach to the initial input features. Perform time path modeling and feature path modeling, and then perform multi-dimensional interactive fusion based on these.

[0027] S21. Construct the main channel and incorporate the initial input features. Send to the main channel (time path); In this embodiment, the initial input features are... Expanding along the time dimension, we obtain a form like The sequence representation is given by , where B is the batch size, T is the number of time steps, and F is the dimension of the hidden features. For the p-th layer time path (p = 1,2,…,P), the output of the previous layer is used. As input, construct the query matrix of the time path at layer p. Key matrix Sum matrix : ; in, It is the trainable weight matrix used to generate the query in the p-th layer time path. It is the trainable weight matrix used to generate the key in the p-th layer time path. This is the trainable weight matrix used to generate values ​​in the p-th layer time path; dividing the above matrix into h attention heads according to the feature dimension, for the i-th attention head, we have: ; Where h is the number of attention heads, For each attention head, define its feature dimensions. Concatenate the outputs of each attention head along these feature dimensions and linearly map them to obtain the output of the p-th layer's time path. : ; in, Let p be the output projection matrix of the p-th layer. Stacking multiple time paths allows for the full capture of long-term trends, periodicity, and cross-period dependencies. The output at time step t is denoted as... This is called a long-term dependent feature in the time dimension; Concat() represents the operation of concatenating the outputs of multiple attention heads along the feature dimension. This represents the attention output of the i-th attention head in the p-th time path at time step t. h represents the number of attention heads.

[0028] S22, The initial input features Send to the feature channel (feature path); Unlike the time path, the feature path focuses on the interactions between different variables. To this end, the input tensor is transposed along the feature dimension, treating the feature dimension as a "sequence" dimension, and multi-head self-attention modeling is performed on the interactions between different features. Specifically, the initial input features... This represents a tensor rearranged into shape (B,F,T). Where B is the batch size, D is the original feature dimension, T is the time length, and F is the hidden feature dimension. Construct the query matrix for the feature path of the p-th layer. Key matrix Sum matrix : ; in, It is the pth Output features of the first-level feature path It is the trainable weight matrix used to generate the query in the feature path of the p-th layer. It is the trainable weight matrix used to generate the key in the feature path of the p-th layer. It is the trainable weight matrix used to generate values ​​in the feature path of the p-th layer; The feature path output is obtained through a similar scaling dot product multi-head attention computation: ; in, The p-th layer feature path outputs a projection weight matrix. This indicates that the i-th attention head in the p-th layer feature path is based on , , The calculated attention output, where h represents the number of attention heads. Reflecting the correlation strength, redundancy, and cooperative patterns among variables, their representation at time step t is denoted as... This is called the static structural feature of the feature dimension. Finally, Transpose the shape back to match the time path in the time-feature dimension for subsequent fusion.

[0029] S3. The long-term time-dependent features output by the main channel, the static structural features of the feature dimension output by the feature channel, and the initial input features are sent to the multi-path interaction module to construct fused features; In this step, features from two channels (i.e., long-term dependent features in the time dimension) are... , Feature Dimension Static Structural Features The mapping features between the input and the input (i.e., the initial input features) Vector at time step t They are fed together into the multi-path interaction module to achieve non-linear fusion between features.

[0030] S31. Input the three-source features into the multi-path interaction module; the three-source feature input includes long-time dependent features in the time dimension. Static structural features with feature dimensions and initial input feature vector ; S32. Perform a non-linear weighted combination of the three source features according to the trained weights to obtain the weighted combination result: ; in, It is a non-linear activation function. This is the intermediate fused representation obtained at time step t from the three source features through a weighted nonlinear transformation. To act on long-term dependent features in the time dimension The trainable weight matrix, To act on static structural features of the feature dimension The trainable weight matrix, To act on the initial input features The trainable weight matrix.

[0031] S33. Input the weighted combination result into the gated fusion layer, and perform path adaptive fusion through dynamic gating; Gating is used to control the proportion of different feature paths, enabling the model to automatically adjust the fusion result according to the scene: ; in, These are the path fusion weights at time step t, with values ​​between (0,1), used to control the proportion of different feature paths. It is a sigmoid-gated activation function. These are the trainable weight vectors of the gating layer, used to... Mapped to scalar gating coefficients. The final fusion output is: ; in, It is the final fusion of multi-dimensional interactive features at time step t. The fusion coefficient is... It is the long-term dependency feature of the main channel time path in the time dimension at time step t; Obtain the multidimensional interactive features after fusion.

[0032] S4. Perform dimensionality reduction and structural compression of potential components; To reduce redundancy in high-dimensional features and highlight the most critical components for prediction, the multidimensional interactive features are input into a regularized latent component regression model. S41. Implement centralized and standardized processing; The multidimensional interaction features obtained in step S3 The stacked input feature matrix X and the target matrix Y, constructed in step S1 using a sliding window method and consisting of the predicted target sequence corresponding to the input feature matrix at the time step, are respectively centered and standardized to eliminate the influence of differences in the scale and numerical scale of different variables on the latent component extraction process.

[0033] S42. Calculate the covariance matrix and obtain the potential directions through singular value decomposition; By performing covariance analysis and singular value decomposition on the input matrix and target matrix after centering and standardization, the latent direction that best matches the target matrix is ​​obtained, and the projection weights are obtained. The steps are as follows: S421, Construction of the covariance matrix; The input feature matrix X and the target matrix Y are mapped along the sample dimension, and the covariance matrix between them is calculated. Its formula is:

[0034] Where N is the number of samples (i.e. the number of rows in the stacked matrix), used to characterize the linear correlation between multidimensional interactive features and the prediction target.

[0035] S422, Covariance Matrix Decomposition; For the obtained cross covariance matrix Performing singular value decomposition yields a set of P, R, Σ such that the following equation holds (this decomposition exhibits equivalence non-uniqueness within the symbol / subspace): ; Where P is the left singular vector matrix, R is the right singular vector matrix, and Σ is the singular value matrix. The singular values ​​are arranged in descending order to extract the feature directions that reflect the main correlation patterns between the input features and the prediction target.

[0036] S423, Selection of Potential Directions; Take the first K column vectors of the left singular vector matrix P as the input direction vector, denoted as ; in, This is the k-th potential direction vector, used for constructing subsequent potential components.

[0037] S43, forming potential component sequences; By projecting the features into a low-dimensional space using the latent orientation, we obtain the latent component score vector: ; in, The input direction (load vector) for the k-th potential component. Let be the corresponding latent component score vector (score in the sample dimension). For ease of alignment with time steps, let the latent component score corresponding to the t-th time step be: ; The latent component vector at time step t is denoted as: ; This allows for the compression of dimensionality while retaining key information relevant to the prediction objective.

[0038] like Figure 5 As shown, the latent component projection process used in this embodiment can be illustrated geometrically: the elliptical point cloud on the left represents the high-dimensional sample distribution composed of multi-dimensional interactive features, and the arrow... This represents the k-th latent direction obtained through singular value decomposition. Projecting each sample along this direction yields a one-dimensional coordinate, which represents the corresponding latent component score. By selecting several dominant latent components, the feature dimensionality can be significantly reduced while preserving the main predictive information.

[0039] S5. Perform residual compensation to enhance local fitting ability; S51. Let the initial input features obtained in step S1 be represented at time step t as follows: The latent component vector obtained in step S4 is The initial input features are reconstructed by fitting a mapping: ; in, To fit the mapping weight matrix, For bias terms, This is the initial input feature estimate obtained from the reconstruction of latent components.

[0040] S52. Subtract the reconstruction result from the initial input features to obtain the residual features: ; in, Characterizes local fluctuations and fine-grained information that the underlying components fail to explain.

[0041] S53. To enable residual features to participate in subsequent predictions, the multidimensional interactive features fused at time step t will be used. With residual characteristics By concatenating the features, we obtain the residual augmentation representation: ; Wherein, Concat() represents concatenation along the feature dimension; the This is one of the input features for the prediction head in step S7.

[0042] S6. Based on the adversarial optimization algorithm, the initial input features are processed... Perform robust training to generate adversarial examples; In this embodiment, the adversarial perturbation optimization algorithm is a perturbation generation method based on the gradient direction of the loss function. It constructs adversarial examples by applying restricted perturbations in the input feature space along the direction that increases the prediction error, thereby improving the robustness of the model in perturbation environments.

[0043] Specifically, during the model training phase, the input samples are first predicted forward using the current model parameters, and the loss function between the predicted result and the real target is calculated. Subsequently, based on the gradient information of the loss function relative to the initial input features, the perturbation direction is determined, and a perturbation with limited amplitude is applied in the direction to generate adversarial examples.

[0044] S61. During the training phase, the initial input features... The perturbationless input is used to obtain multidimensional interactive features through steps S2 to S5. and residual characteristic sequence The result is fed into the prediction head to obtain a perturbation-free prediction output: ; in, For the prediction head mapping function, This indicates a concatenation operation along the feature dimension.

[0045] The mean squared error (MSE) is used as the unperturbed loss function, defined as: ; in , Let N represent the predicted sequence and the actual target sequence corresponding to the i-th sample, respectively, and N be the number of samples (or the number of samples in the batch).

[0046] Subsequently, the perturbation-free loss function is calculated. Regarding initial input features gradient: ; in, By analyzing the loss function Obtained by performing backpropagation calculations.

[0047] Based on the gradient, generate adversarial perturbation increments: ; in, This represents the increment of the adversarial perturbation sample relative to the original sample. `sign()` represents the sign function for selecting the element's sign, and `ε` is a hyperparameter controlling the perturbation amplitude. Based on the above perturbation, adversarial samples are constructed: .

[0048] S62, Disturb the sample Inputting the same model, steps S2 to S5 yield the corresponding multidimensional interaction features and residual feature sequences, which are then fed into the prediction head to obtain the adversarial prediction output: ; in, , This indicates that the input is The corresponding feature sequence obtained under the given conditions (distinguished from the feature sequence under undisturbed input). , These are the eigenvectors at time t.

[0049] The mean squared error is used as the adversarial loss function, defined as: ; in , Let represent the predicted sequence and the actual target sequence corresponding to the i-th sample, respectively.

[0050] And construct joint loss : ; in, The prediction error is for unperturbed samples. Let λ represent the prediction error under perturbed samples, and λ be the importance coefficient that balances the two factors. This is achieved by minimizing the joint loss. The model parameters are updated to ensure that the model maintains stable predictive performance under both undisturbed and perturbed input conditions.

[0051] During the verification phase, noisy samples were further constructed. And samples with added jump peaks The errors under noisy conditions were obtained respectively. Errors in jump scenarios And calculate the error ratio: .

[0052] On test samples containing sudden anomaly segments, the error of the abnormal scene was obtained. and calculate ; in, It represents the error ratio under noise disturbance scenarios, and is used to measure the degree of change in the prediction error of the model under the condition of adding noise disturbance samples relative to the prediction error of the undisturbed samples. It represents the error ratio under abrupt disturbance scenarios, used to measure the degree of change in the prediction error of the model under the condition of adding sudden abrupt disturbance samples relative to the prediction error of undisturbed samples. This is the error ratio under abnormal scenarios, used to measure the degree of change in the model's prediction error under abnormal scenarios compared to the prediction error under undisturbed conditions. When the above ratio is closer to 1 or lower than a preset threshold, it indicates that the model's performance degradation is smaller and its robustness is stronger under the corresponding perturbation scenario.

[0053] S7. Perform reasoning and reverse normalization reduction; During the inference phase, the multidimensional interaction features and the residual features are fed into the prediction head for prediction, resulting in a predicted sequence of the target variable in the normalized space: ; in, It is the predicted sequence of the target variable obtained by the model in the normalized space. It is the prediction head mapping function. This indicates a concatenation operation along the feature dimension. It is the fused multidimensional interactive feature corresponding to time step t (including the latent space representation after residual compensation). This represents the residual feature corresponding to time step t.

[0054] Subsequently, based on the normalization parameters recorded during the normalization process of the original time series data in step S1, an inverse normalization restoration operation is performed on the prediction result. The specific process includes: S71. Scale recovery is performed on the prediction results based on the scaling and bias parameters of the corresponding variables during normalization. S72. Perform a linear transformation on the restored prediction results that is the opposite of the normalization process, and map the prediction results from the normalization space back to the original variable space. S73. Reconstruct the predicted sequence according to the time step order to obtain the prediction result aligned with the original time series in the time dimension.

[0055] Through the above inverse normalization restoration steps, the predicted sequence within the target time interval at the scale of the original variables is obtained.

[0056] Analysis of experimental results: In one specific embodiment, the present invention was systematically validated on multiple publicly available datasets, covering typical time series with different dimensions and structural characteristics, including the power transformer temperature datasets ETTh1 and ETTh2, the weather monitoring dataset Weather, the electricity load dataset Electricity, and the high-dimensional traffic sensor dataset Traffic. These datasets have different characteristics such as low-dimensional climate cycles, multivariate industrial monitoring, high-dimensional electricity demand, and ultra-high-dimensional traffic flow, which can comprehensively reflect the adaptability and stability of the method of the present invention in various practical scenarios. Furthermore, it was compared with the current mainstream time series prediction models itransformer, Dlinear, Crossformer, and TimesNet, and the comparison results are shown in Table 1 below.

[0057] In the experiment, the historical window length was set to L=512, and the prediction step size H was set to 96, 192, 336, and 720 to cover short-term, medium-term, and long-term prediction needs. The model performance was evaluated using mean squared error (MSE) and mean absolute error (MAE), and was compared with various comparative models under the same training configuration and input window conditions.

[0058] Table 1 Comparison results between our method and existing methods

[0059] As shown in Table 1, the dual-channel multi-path interaction and perturbation optimization method proposed in this invention achieves lower MSE and MAE in most scenarios, and its performance is consistently better than the baseline model without a dual-channel structure or without a perturbation optimization strategy. In particular, on high-dimensional datasets such as Electricity and Traffic, the error of this invention is significantly reduced in long prediction ranges (e.g., H=336 or H=720), indicating that the proposed time-feature dual-path modeling mechanism can effectively alleviate common problems such as information dilution, gradient degradation, and model overfitting when dealing with high-dimensional long sequences.

[0060] From the prediction curve Figure 3 and residual Figure 4 In terms of performance, this invention maintains stable trend-following capabilities across different time periods. In stable intervals, the predicted curve highly overlaps with the actual sequence; in periods of sharp fluctuations, the model still responds promptly and provides reasonable trend predictions. The residual distribution exhibits characteristics of near-zero mean, symmetrical concentration, and small variance, indicating that the model does not have significant systematic bias and has good overall fitting quality. Furthermore, without additional complex structures or large-scale parameters, this invention enhances the model's generalization ability through a dual-channel structure and perturbation training, enabling the model to maintain stable predictive performance even when facing noise, outliers, or local mutations.

[0061] The comparison of model resource consumption and computational efficiency shows that the method of this invention has significant lightweight advantages while maintaining high prediction accuracy. Table 2 shows the differences between this invention and representative deep time series models such as Informer, Autoformer, FEDformer, and PatchTST in terms of parameter size, computational cost (MACs), memory consumption, single-round training time, and inference latency.

[0062] Table 2 shows the differences in indicators between our method and existing models.

[0063] As shown in Table 2, traditional deep temporal models based on attention or hybrid structures typically have a large number of parameters (6M–18M), high computational complexity (approximately 4G–11GMACs), and significant GPU memory usage, while also exhibiting relatively slow inference speeds. For example, Informer and Autoformer both have around 12M parameters, FEDformer's GPU memory usage exceeds 1100MB, and PatchTST's computational cost reaches 11.21GMACs. In contrast, the method of this invention contains only approximately 1.12M parameters, significantly reducing structural complexity while maintaining complete feature extraction capabilities. Furthermore, on the same hardware platform, the training time of this invention is 33.9 seconds, far lower than PatchTST's 290.3 seconds; during the inference phase, the inference latency of this invention is 838.4ms, only about 1 / 46th of FEDformer's, also significantly better than other comparative models.

[0064] In summary, the proposed method is more lightweight in terms of model size, memory overhead, and computational load, with lower dependence on hardware computing power. It can be efficiently deployed in resource-constrained industrial environments, edge computing devices, and online prediction systems. Furthermore, the method achieves leading prediction accuracy while maintaining a smaller model size, demonstrating its excellent expressive power and computational efficiency. This indicates its broad application value and potential for widespread adoption in time series forecasting tasks across multiple industries.

Claims

1. A time series forecasting method based on dual-channel multi-path interaction and perturbation optimization, characterized in that, Includes the following steps: S1. Obtain multivariate time series data, perform layer normalization and embedding representation on the original input data to form initial input features; S2. Construct a dual-channel system to perform time path modeling and feature path modeling on the initial input features, respectively capturing long-term dependent features in the time dimension and static structural features in the feature dimension. S3. The long-term time-dependent features and static structural features of the feature dimension are fed into the multi-path interaction module to obtain the fused multi-dimensional interaction features. S4. Input the multidimensional interaction features into the regularized latent component regression model to obtain the latent components; S5. Fit the latent components to the initial input features, and use the calculated difference as the residual features; S6. Generate adversarial perturbations on the initial input features and train them together with the original input data to optimize the target; S7. The multidimensional interactive features and residual features are fed into the prediction head for prediction to obtain the prediction sequence for the target time interval.

2. The time series prediction method based on dual-channel multi-path interaction and perturbation optimization according to claim 1, characterized in that, In S1, the multivariate time series data undergoes layer normalization and embedding representation of the original input data to form initial input features, including: S11. Construct input samples using a sliding window approach, taking a historical window of length L as the input features; S12. Perform layer normalization on the input features to ensure that the values ​​of each time step remain stable at the same feature scale. S13. Map the feature vector corresponding to each time step to a unified hidden space dimension to obtain the hidden space features, and use the hidden space features as the initial input features.

3. The time series prediction method based on dual-channel multi-path interaction and perturbation optimization according to claim 2, characterized in that, In S2, the specific steps of constructing a dual-channel system and performing time path modeling and feature path modeling on the initial input features to capture long-term dependent features in the time dimension and static structural features in the feature dimension respectively include: S21. For the time path, the initial input features are sent into the main channel, and a multi-head self-attention mechanism is used to model the dependency relationship between different times to obtain the long-term dependency features of the time dimension of the time path output. S22. For the feature path, the initial input features are sent into the feature channel, transposed in the feature dimension, and a multi-head self-attention mechanism is used to model the static coupling relationship between different variables to calculate the static structural features of the feature dimension output by the feature path.

4. The time series prediction method based on dual-channel multi-path interaction and perturbation optimization according to claim 3, characterized in that, In S3, the specific process of feeding the long-term time-dependent features and the static structural features of the feature dimension into the multi-path interaction module to obtain the fused multi-dimensional interaction features is as follows: S31. Input the long-term time-dependent features, static structural features, and initial input features into the multi-path interaction module. S32. Perform a non-linear weighted combination of the three source features according to the trained weights to obtain the weighted combination result; S33. Input the weighted combination result into the gated fusion layer, and perform path adaptive fusion through dynamic gating to obtain the fused multidimensional interactive features.

5. The time series prediction method based on dual-channel multi-path interaction and perturbation optimization according to claim 4, characterized in that, In S4, the specific steps for inputting the multidimensional interaction features into the regularized latent component regression model to obtain the latent components include: S41. Centralize and standardize the multi-dimensional interactive features; S42. By performing covariance analysis and singular value decomposition on the input matrix and target matrix after centering and standardization, the potential direction that best matches the target matrix is ​​obtained, and the projection weights are obtained. S43. Project the features into a low-dimensional space through the latent direction to obtain the latent component score vector.

6. The time series prediction method based on dual-channel multi-path interaction and perturbation optimization according to claim 5, characterized in that, In S5, the latent components are fitted with the initial input features, and the difference is calculated as the residual features; S51. Reconstruct the initial input features based on the latent component score vector by fitting a mapping; S52. Subtract the reconstruction result from the initial input features to obtain the residual features; S53. The multidimensional interactive features and the residual features are concatenated along the feature dimension to obtain the residual enhancement representation.

7. The time series prediction method based on dual-channel multi-path interaction and perturbation optimization according to claim 6, characterized in that, In S6, the specific steps for optimizing the objective are as follows: S61. The initial input features, the multidimensional interaction features, and the residual features are fed into the prediction head to obtain a perturbation-free prediction output; mean square error is used as the perturbation-free loss function. S62. Input the perturbation sample, the multidimensional interaction feature and the residual feature into the prediction head to obtain the adversarial prediction output; use the mean squared error as the adversarial loss function. S63. Construct a joint loss based on the prediction error under unperturbed samples and the prediction error under perturbed samples; and update the model parameters by minimizing the joint loss.

8. The time series prediction method based on dual-channel multi-path interaction and perturbation optimization according to claim 7, characterized in that, In S7, the specific steps for feeding the multidimensional interaction features and residual features into the prediction head for prediction to obtain the prediction sequence for the target time interval include: S71. Scale recovery is performed on the prediction results based on the scaling and bias parameters of the corresponding variables during normalization. S72. Perform a linear transformation on the restored prediction results that is the opposite of the normalization process, and map the prediction results from the normalization space back to the original variable space. S73. Reconstruct the predicted sequence according to the time step order to obtain the prediction result aligned with the original time series in the time dimension.

9. A time series prediction system for implementing the time series prediction method based on dual-channel multi-path interaction and perturbation optimization as described in any one of claims 1-8, characterized in that, It includes a raw time series data processing module, a dual-channel feature extraction module, a multi-path interaction module, a latent component regression module, a fitting module, a perturbation optimization module, and a prediction module; The raw time series data processing module is used to acquire raw time series data and preprocess it to form initial input features; The dual-channel feature extraction module is used to construct the main channel and the feature channel, and to perform time path modeling and feature path modeling based on the initial input features, thereby capturing long-term dependent features in the time dimension and static structural features in the feature dimension, respectively. The multi-path interaction module is used to fuse the long-term time-dependent features and the static structural features of the feature dimension to obtain the fused multi-dimensional interaction features. The latent component regression module is used to input the multidimensional interaction features into the regularized latent component regression model to obtain latent components; The fitting module is used to fit the latent components with the initial input features and use the calculated difference as the residual features; The perturbation optimization module is used to generate adversarial perturbations on the initial input features and to participate in training together with the original input data to optimize the target. The prediction module is used to feed the multidimensional interactive features and residual features into the prediction head for prediction, so as to obtain the prediction sequence of the target time interval.

Images