Time sequence prediction method based on multi-scale decomposition and gating fusion

Abstract

The invention belongs to the technical field of time sequence modeling and prediction, and particularly relates to a time sequence prediction method based on multi-scale decomposition and gating fusion. Performing multi-level down-sampling operation on the original meteorological time sequence data, and outputting a meteorological time sequence of a time scale after down-sampling; on the basis of the down-sampled meteorological time sequence, applying a neural Fourier trend decomposition mechanism to extract trend components and seasonal components of original meteorological time sequence data under different scales; carrying out fusion processing on the trend components and the seasonal components under different scales through a gating multi-layer sensing network to form meteorological time characteristics of a multi-periodic structure; normalizing the meteorological time characteristics of the multi-periodic structure, then generating a prediction result through each scale prediction branch, and finally obtaining multi-scale prediction output through weighted fusion; and combining various loss functions to form a total loss function for training optimization.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of time series modeling and prediction, and particularly relates to a time series prediction method based on multi-scale decomposition and gated fusion. BACKGROUND

[0002] With the rapid development of information technology and sensing devices, time series data is widely collected and applied in many fields such as meteorology, energy, transportation, finance, etc. Accurate prediction of future time series is of great significance to improve energy scheduling efficiency, optimize meteorological disaster warning, and improve resource utilization. However, in practical applications, time series often exhibit non-stationarity, periodic mixing, cross-scale variation, and multivariate coupling characteristics, which pose higher requirements on traditional modeling and prediction methods.

[0003] Traditional methods such as ARIMA (autoregressive integrated moving average model), exponential smoothing, etc. rely on the stationarity assumption of time series, and are difficult to handle complex nonlinear relationships and long-term dependence characteristics. In recent years, with the development of deep learning technology, models such as RNN (recurrent neural network), CNN (convolutional neural network), Transformer and its variants have been gradually applied to time series modeling tasks, significantly improving model performance. However, these models often have complex structures and consume large amounts of computing resources, and lack explicit modeling mechanisms when dealing with periodic mixing and multi-scale structures, resulting in insufficient model generalization ability and interpretability.

[0004] To solve the above problems, existing research has attempted to introduce sequence decomposition and periodic extraction mechanisms, such as Autoformer (self-coupling decomposition Transformer), FEDformer (frequency-enhanced decomposition Transformer) and other models that improve predictability by decomposing time series into trend and seasonal terms. Meanwhile, some research has adopted frequency domain modeling, sliding window linear regression, etc. to model information at different time granularities. However, these methods still have room for optimization in terms of handling multi-period mixing, multi-scale feature fusion, and model structure lightweight.

[0005] Therefore, there is an urgent need for a time series prediction method that can balance modeling performance and structure lightweight, with good periodic feature extraction ability and multi-scale fusion ability, to cope with real-world long sequence, high frequency, and complex pattern prediction tasks, especially in meteorological temperature, humidity, precipitation, and other time-varying factor prediction, a modeling framework with strong periodic recognition ability, good residual stability, and strong generalization performance is needed. SUMMARY

[0006] To overcome the problems in the prior art, the present application proposes a time series prediction method based on multi-scale decomposition and gated fusion.

[0007] The technical scheme for solving the above technical problems of the present application is as follows: The present application provides a time series prediction method based on multi-scale decomposition and gating fusion, comprising the following steps: Performing multi-level downsampling operation on the original meteorological time series data to output the meteorological time series of the downsampled time scale; Based on the downsampled meteorological time series, applying a neural Fourier trend decomposition mechanism to extract trend components and seasonal components at different scales of the original meteorological time series data; Fusing the trend components and seasonal components at different scales through a gated multi-layer perception network to form meteorological time features with multiple periodic structures; Splitting the meteorological time features with multiple periodic structures according to the channels, outputting independent time sequence features of each channel, and embedding high-dimensional representation; and after normalization, generating prediction results from each scale prediction branch, and finally obtaining multi-scale prediction output through weighted fusion; Combining multiple loss functions to form a total loss function for training and optimization, the loss functions including standard supervised loss, auxiliary loss and regularization loss, wherein a consistency regularization term is designed as auxiliary loss based on high-dimensional representation features.

[0008] Further, the original meteorological time series data is a multivariate sequence recorded at fixed time intervals.

[0009] Further, based on the downsampled meteorological time series, a neural Fourier trend decomposition mechanism is applied to extract trend components and seasonal components at different scales of the original meteorological time series data, comprising: Based on the downsampled meteorological time series, a fast Fourier transform is used to obtain a complex frequency domain representation of the meteorological time series; Performing channel dimension weighting on the complex frequency domain representation of the meteorological time series, the weight being a learnable parameter, to obtain a weighted result; Based on the weighted result, performing real and imaginary part separation and nonlinear enhancement processing, respectively extracting real and imaginary parts, and using a multilayer perception network with residual to perform nonlinear mapping to obtain nonlinearly mapped real and imaginary parts; Merging and enhancing the nonlinearly mapped real and imaginary parts, and performing inverse Fourier transform to obtain the result after inverse Fourier transform, i.e., the reconstructed seasonal component; Using the original meteorological time series data input to subtract the reconstructed seasonal component to obtain the trend component: For the seasonal component and the trend component, the neural Fourier trend decomposition mechanism is performed to obtain the trend component and the seasonal component.

[0010] Further, the trend component and the seasonal component at different scales are fused by a gated multi-layer perception network to form a multi-period structure meteorological time feature representation, including: combining the trend component and the seasonal component to form a fused feature representation at each scale; inputting the fused feature at each scale into a gated multi-layer perception network; aligning the unified time dimension of the fused output of all scales, ensuring consistency of different scale outputs in the prediction stage through interpolation or time alignment operation; stacking and summarizing the fused output of all scales to form the final multi-period structure meteorological time feature representation.

[0011] Further, the gated multi-layer perception network includes a plurality of gated perception units, each of which includes the following steps: map the fused feature to a high-dimensional space to obtain an upgraded fused feature; the upgraded fused feature calculates the backbone feature through the main branch, and the backbone feature is calculated through the gating branch to calculate the gating weight; multiply the backbone feature and the gating weight to form a weighted output, and connect the weighted output and the fused feature residual to perform LayerNorm standardization.

[0012] Further, the upgraded fused feature calculates the backbone feature through the main branch, and the backbone feature is calculated through the gating branch to calculate the gating weight, including: the upgraded fused feature calculates the backbone feature through the main branch: ; ; the backbone feature is calculated through the gating branch to calculate the gating weight: ; wherein, is a weight matrix; is a bias; denotes a Sigmoid activation function for generating a gating factor; H denotes a backbone feature; O denotes a fused feature after linear transformation of the original feature; G denotes a gating weight.

[0013] Further, a consistency regularization term is designed based on the high-dimensional representation feature as an auxiliary loss, including: meteorological time series data X based on neural network decomposition output trend item and seasonal item : ; And based on the high-dimensional representation characteristics, a consistency regularization term is designed as an auxiliary loss: ; In the above formula, The auxiliary loss is represented; E represents the high-dimensional representation characteristics.

[0014] Further, the total loss function : ; Wherein, The total loss function is represented; The regularization loss is represented; The auxiliary loss weight is represented; The regularization loss weight is represented; The standard supervised loss is represented.

[0015] Compared with the prior art, the present application has the following technical effects: The present application proposes a time series prediction method based on multi-scale decomposition and gated fusion, which has the significant characteristics of light structure, period sensitivity and strong adaptability. The method can fully capture long-term and short-term periodic structures by constructing multi-scale time representation through down-sampling mechanism, and significantly improve the modeling ability of complex periodic patterns. Unlike traditional single scale or linear modeling methods, the present application introduces a neural Fourier trend decomposition mechanism under multi-scale to extract trend components and seasonal components, thereby improving the diversity and robustness of feature expression. A gated multi-layer perception network is used to realize adaptive weighting and noise suppression between multi-scale outputs, further optimizing key feature selection. In addition, the method supports channel independent modeling strategy, which can effectively reduce the parameter quantity and is suitable for deployment in resource-constrained environments. In summary, the present application is particularly suitable for processing long sequence prediction tasks with mixed multi-period characteristics, especially meteorological observation tasks. While ensuring prediction accuracy, it has superior scalability and industrial practical value. BRIEF DESCRIPTION OF DRAWINGS

[0016] In order to more clearly illustrate the technical solutions and advantages of the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and those skilled in the art can obtain other drawings according to these drawings without creative labor.

[0017] Figure 1 The flow chart of the multi-scale decomposition and gated fusion time series prediction method of the present application; Figure 2 The prediction result fitting graph performed on the Weather weather data set; Figure 3 A prediction result graph made on the Weather weather dataset. DETAILED DESCRIPTION

[0018] In order to further clarify the technical means and effects taken by the present application to achieve the predetermined object of the application, the specific embodiments, structures, features and effects of the technical solutions proposed according to the present application are described in detail below in combination with the drawings and preferred embodiments. The specific features, structures or characteristics in one or more embodiments can be combined in any suitable form. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as understood by those skilled in the art to which the present application belongs.

[0019] In one embodiment of the present application, referring to Figure 1 , a time series prediction method based on multi-scale decomposition and gated fusion is provided, comprising the following steps: Step 100: performing multi-level downsampling operation on the original meteorological time series data, and outputting the downsampled time scale meteorological time series; Step 200: based on the downsampled meteorological time series, applying a neural Fourier trend decomposition mechanism to extract trend components and seasonal components at different scales of the original meteorological time series data; Step 300: fusing the trend components and seasonal components at different scales through a gated multi-layer perception network to form meteorological time features with multiple periodic structures; Step 400: splitting the meteorological time features with multiple periodic structures according to the channels, outputting independent time series features of each channel, and embedding high-dimensional representation; and generating prediction results from each scale prediction branch after normalizing the meteorological time features with multiple periodic structures, and finally obtaining multi-scale prediction output through weighted fusion; Step 500: training and optimizing by combining a plurality of loss functions to form a total loss function, the loss functions including standard supervised loss, auxiliary loss and regularization loss, wherein a consistency regularization term is designed as auxiliary loss based on high-dimensional representation features.

[0020] The above steps are described in detail as follows: Step 100: performing multi-level downsampling operation on the original meteorological time series data, and outputting the downsampled time scale meteorological time series.

[0021] In the present application, one specific implementation of step 100 can be: Step 1001: receiving the original meteorological time series data.

[0022] The original meteorological time series data is a multivariate sequence recorded at fixed time intervals, including any combination of meteorological observation values such as temperature, humidity, air pressure, wind speed, and precipitation.

[0023] Let the original meteorological time series data be , where B represents the batch size, T1 represents the time length, and F represents the feature dimension.

[0024] Step 1002: Perform multi-level downsampling operation on the original meteorological time series data along the time dimension to obtain the downsampled meteorological time series.

[0025] Each downsampling operation uses a specified sliding window with a length of w and a step size of s, forming a number of downsampled sequences , where L represents the number of downsampling layers, X (0) = X original , and the downsampled meteorological time series of the remaining each layer is , , where wL represents the length of the sliding window used in the Lth downsampling operation. .

[0026] Step 1003: For each downsampled meteorological time series, generate its corresponding timestamp label synchronously.

[0027] For each downsampled meteorological time series, generate its corresponding timestamp label synchronously , where D represents the time embedding dimension, which is used to assist the subsequent model in perceiving periodic and seasonal changes; , and T represents the number of time steps of the downsampled meteorological time series after the Lth layer. The timestamp label can include discrete time encodings such as hours, days, weeks, months, and whether it is a holiday, which will be combined with the sequence features through embedding in the subsequent stage. .

[0028] Step 1004: Output the list of downsampled meteorological time series and the corresponding list of time label information, forming a multi-scale meteorological time representation structure for subsequent decomposition module processing and fusion.

[0029] Output the list of downsampled meteorological time series at all scales : ; , and the corresponding list of time labels M: ; for use in the next stage of decomposition and fusion. ​​

[0030] Step 200: based on the down-sampled meteorological time series, a neural Fourier trend decomposition mechanism is applied to extract the trend component and seasonal component of the original meteorological time series data at different scales.

[0031] In the present application, a specific implementation of step 200 can be: Based on the down-sampled meteorological time series, a neural Fourier trend decomposition mechanism is applied to decompose the original meteorological observation value into trend component and seasonal component with physical meaning, so as to enhance the predictability and stability of sequence modeling.

[0032] Wherein, the neural Fourier trend decomposition (Neural Fourier Trend, NFT) adopts fast Fourier transform (FFT) to obtain complex spectrum, uses frequency selection mechanism with learnable weight to weight each channel frequency, and then obtains refined seasonal modeling result through nonlinear MLP mapping.

[0033] Step 2001: down-sampled meteorological time series , wherein B is the batch size, represents the sequence length after down-sampling of the i-th layer, and D is the feature dimension. l

[0034] Step 2002: based on the down-sampled meteorological time series, fast Fourier transform is adopted to obtain the complex frequency domain representation of the meteorological time series: ; In the above formula, represents the complex frequency domain representation of the meteorological time series.

[0035] Step 2003: weighting the complex frequency domain representation of the meteorological time series in channel dimension, and the weight is learnable parameter , to obtain the weighted result: ; In the above formula, represents the weighted result.

[0036] Step 2004: based on the weighted result, the real part and the imaginary part are separated and nonlinear enhancement processing is performed.

[0037] The real part and the imaginary part are extracted respectively, and a multilayer perception network with residual is used for nonlinear mapping: ; ; In the above formula, represents the real part after nonlinear mapping.​ represents the imaginary part after the nonlinear mapping.

[0038] Step 2005: The real part after the nonlinear mapping and the imaginary part after the nonlinear mapping are combined and enhanced, and an inverse Fourier transform is performed: ; ; ; wherein, is a learnable residual scaling factor; i represents an imaginary unit; , represents the combined and enhanced spectrum; represents the result after the inverse Fourier transform, i.e., the reconstructed seasonal component.

[0039] Step 2006: The reconstructed seasonal component is subtracted from the original meteorological time series data input to obtain a trend component: ; Finally, the decomposition results (the seasonal component) and (the trend component) are output for subsequent fusion modeling processes.

[0040] Step 2007: For the seasonal component and the trend component, a neural Fourier trend decomposition mechanism is performed to obtain a trend component and a seasonal component wherein, represents the number of down-sampling layers.

[0041] The trend component and the seasonal component corresponding to each scale are respectively reserved for subsequent cross-scale gating fusion operations and multi-path prediction tasks.

[0042] Step 300: The trend component and the seasonal component at different scales are fused through a gated multi-layer perception network to form a meteorological time feature representation with a multi-period structure.

[0043] In the present application, one specific implementation of step 300 can be: Step 3001: The trend component and the seasonal component are combined to form a fusion feature representation at each scale.

[0044] The independence of the trend and seasonal information is preserved by using a connection method rather than an additive method, i.e., the trend component and the seasonal component at each scale are combined in series to form a fusion feature at each scale: ; ; wherein, represents the dimension of the channel after expansion, represents the channel dimension of the original single component; represents the length of the time step corresponding to the scale after downsampling.

[0045] Step 3003: input the fusion feature of each scale into the gated multi-layer perception network.

[0046] The gated multi-layer perception network is composed of several gated perception units, and each gated perception unit includes the following steps: Step 30031: dimensionality transformation, mapping the fusion feature to a high-dimensional space to enhance the expression ability: ; In the above formula, represents the fusion feature after dimensionality expansion; represents a weight matrix; represents a bias term.

[0047] Step 30032: the fusion feature after dimensionality expansion is calculated through the main branch to obtain the backbone feature, and the backbone feature is calculated through the gating branch to obtain the gating weight.

[0048] The learnable gating branch is used to control the information flow of the main branch, so as to improve the recognition of key features and the noise suppression ability of the model.

[0049] The fusion feature after dimensionality expansion is calculated through the main branch to obtain the backbone feature: ; ; The backbone feature is calculated through the gating branch to obtain the gating weight: ; wherein, is a weight matrix; is a bias; represents a Sigmoid activation function for generating a gating factor; H represents a backbone feature; O represents a fusion feature after linear transformation of an original feature; and G represents a gating weight.

[0050] Step 30033: multiply the backbone feature H and the gating weight G to form a weighted output, and perform LayerNorm standardization on the weighted output H⊙G and the fusion feature after residual connection, so as to enhance the stability and training convergence of the model: ; In the above formula, denotes the normalized output.

[0051] Step 3004: Perform uniform time dimension alignment on the normalized outputs of all scales, and ensure that the outputs of different scales are aligned in time by interpolation or time alignment operation In the prediction stage, there is consistency: ; In the above formula, is the normalized output of the i-th scale, is the output after scale alignment, is an alignment operation function, is a target time dimension, is the normalized output of the i-th scale,

[0052] Step 3005: Stack the fusion outputs of all scales and summarize them as , and the stacking and summarizing formula is: form the final meteorological time feature of multiple periodic structures as the modeling and input of the subsequent prediction path.

[0053] Step 400: Split the meteorological time feature of multiple periodic structures according to the channel, output the independent time sequence feature of each channel, and embed a high-dimensional representation; and generate a prediction result from each scale prediction branch after normalizing the meteorological time feature of multiple periodic structures, and finally obtain a multi-scale prediction output through weighted fusion.

[0054] In the present application, one specific implementation of step 400 can be: Step 4001: Split the meteorological time feature of multiple periodic structures according to the channel, and output the independent time sequence feature of each channel.

[0055] In the case of enabling channel independence, for the meteorological time feature of multiple periodic structures, the channel is split, so that the temperature, humidity, wind speed and other data of each channel form independent one-dimensional time sequences. A shared univariate model is used to model these features, thereby outputting the independent time sequence feature of each channel: ; In the above formula, denotes the independent time sequence feature, ; Reshape is a reshaping operation, which adjusts the shape of the meteorological time feature of multiple periodic structures so that it can be processed in a channel-independent manner, and finally the independent time sequence feature is obtained. ​

[0056] The independent time sequence features are input into the embedding network, and high-dimensional representation features are output. The high-dimensional representation features provide better and abstract features for the subsequent multi-scale fusion and prediction module, helping the model to better make weather prediction. The high-dimensional features, through affecting the prediction result accuracy and the association with the output of the decomposition module, directly associate the quality of the high-dimensional features with the size of the standard supervised loss, consistency regularization term and regularization loss in the loss function optimization, and promote the model parameters to adjust to a better direction: ; In the above formula, represents the high-dimensional representation features; represents the embedding network; represents the timestamp label information.

[0057] The embedding network converts coarse-grained signal features into high-dimensional representations that are readable and optimizable by the model, which is the basis for the modeling accuracy of the subsequent multi-scale fusion and prediction module.

[0058] The embedding network can be a neural network composed of multiple fully connected layers or convolutional layers, which maps independent time sequence features from the original feature space to a higher-dimensional representation space.

[0059] Step 4002: using the standard normalization operation Normalize to the multi-period structure of the weather time feature The mean and variance calibration is performed to make the time representation under different scales have numerical consistency, which is beneficial to the training convergence: ; In the above formula, represents the normalized multi-period structure of the weather time feature; represents the mean of the multi-period structure of the weather time feature ; represents the variance of the multi-period structure of the weather time feature .

[0060] The normalized multi-period structure of the weather time feature generates a prediction result through the prediction module. The prediction module uses the original data to perform preliminary prediction through the above steps 100, 200 and 300, and is a functional unit for time sequence modeling and future value inference of the normalized multi-period weather time feature based on a deep learning architecture. By learning the time association and periodic pattern between features, the weather variable prediction result is output: ; And then it is restored to the true dimension through denormalization The real dimension plays a role in comparing the prediction results of the real dimension with the actual observation values to calculate evaluation indexes such as mean square error, accurately measure the prediction accuracy of the model, and judge the performance of the model: ; Step 4003: Map the features under different down-sampling windows to a fixed prediction length pred_len using the prediction branch corresponding to each scale: ; ; In the above formula, represents the prediction branch corresponding to the i-th scale, l is the processed feature under the i-th scale; represents the prediction result under this scale. l is the normalized meteorological time feature of the multi-period structure, which is decomposed into scale-specific features through down-sampling and scale adaptation processing

[0061] , wherein represents the down-sampling operation of the i-th scale, which ensures that inherits the multi-period feature of and adapts to the input requirements of the multi-scale prediction branch. represents the prediction result under this scale, which is also the output of the single-scale prediction branch, and is the prediction result after multi-scale fusion, which is generated by each

[0062] through fusion operations such as weighting and summation, and is the basis for subsequent de-normalization to obtain : ; , wherein represents the fusion operation, which reflects the relationship between and each .

[0063] After stacking and fusing the outputs of all scales, summation, attention weighting or residual correction can be used to finally obtain the multi-scale prediction fusion output: ; In the above formula, is the final prediction value, which is sent to the loss function for training supervision; represents a learnable weight or a fixed weight.

[0064] ​​​Step 500: Combine multiple loss functions to construct a total loss function for training and optimization. The loss function includes standard supervised loss, auxiliary loss, and regularization loss. A consistency regularization term based on high-dimensional representation features is designed as an auxiliary loss. Combining multiple loss functions optimizes the model, ensuring the accuracy and consistency of the prediction results.

[0065] In this invention, one specific implementation of step 500 can be: Step 5001: Analyze the multi-scale prediction results During training, a standard supervised loss is constructed by comparing the mean squared error (MSE) with the true label: ; In the above formula, This represents the single-scale prediction loss, which measures the error between the predicted value and the actual value. Indicates the batch size (the number of samples in one training / prediction session); Indicates the length of the prediction time step (the number of future time periods to be predicted). Indicates the feature dimension (the number of features in a meteorological time series). Indicates the first i The sample, the first t The first prediction time step, the first j Predicted values ​​of dimensional features; Indicates the first i The sample, the first t The first prediction time step, the first j The true value of the dimensional feature.

[0066] If a multi-scale prediction pathway is adopted, multi-scale weighted prediction error can be introduced: ; In the above formula, Indicates the first l Single-scale prediction loss at each scale; l Indicates scale index (traversing from 0 to ... L (single scale); L Indicates the total number of scales (the number of different scale levels in multi-scale prediction). This represents the total loss of multi-scale weighted prediction, which integrates the prediction errors of each scale. Indicates the first l The weights of each scale (can be learned or preset, adjusting the impact of errors at each scale).

[0067] Step 5002: For the trend term output by the neural network-based decomposition module and seasonal items The following refactoring consistency should be satisfied: ; And based on the high-dimensional representation characteristics, a consistency regularization term is designed as an auxiliary loss: ; In the above formula, represents the auxiliary loss.

[0068] Penalize large weights, make the model more concise, avoid overfitting, design Regularization loss: ; In the above formula, is the regularization strength hyperparameter, is the set of model learnable parameters, is the L2 norm square of the parameter .

[0069] The model training stage will minimize the total loss function composed of the above three loss combinations: ; Where, represents the total loss function; represents the regularization loss; represents the auxiliary loss weight; represents the regularization loss weight.

[0070] Experimental results analysis: Experiments are conducted on the publicly available weather dataset Weather. All time series are segmented into review windows L z = 336, prediction range H ∈ {96, 192, 336, 720}, step size 1, which means that each subsequent window is moved one step, and the resulting error results between 0.08 and 0.3.

[0071] By predicting the Weather weather time series dataset, the time series prediction method based on multi-scale decomposition and gating fusion produces good accurate prediction as Figure 2 shown, the prediction results of the embodiment of the present application on the weather dataset show the excellent fitting ability of the proposed model to the change trend of meteorological variables. The blue curve in the figure is the ground truth, and the orange curve is the prediction value (Prediction) output by the model. It can be seen that the two curves are highly consistent at the key turning points, and the model not only accurately captures the evolution law of the overall trend, but also maintains good prediction stability in multiple local fluctuation areas.

[0072] In order to further verify the effectiveness and practicality of the multi-scale time series prediction method proposed in the present application, refer to Figure 3, a bar chart of the mean square error (MSE) of the model on the weather dataset under different prediction lengths (96, 192, 336, 720) is drawn. The graph intuitively shows the trend of the prediction error gradually rising with the increase of the prediction step, reflecting the inherent challenge of long-term prediction in time series prediction tasks. In the bar chart, short-term prediction (such as 96 steps) has lower error, indicating that the model has high accuracy in capturing short-term dynamics; while in the medium and long-term range (such as 336, 720 steps), it still maintains stable growth, showing that the method has good generalization ability and multi-scale modeling ability. This visualization result helps to evaluate the performance of the model in different prediction tasks from a quantitative perspective, further demonstrating the wide applicability and stability of the invention in long sequence, multi-period tasks such as weather prediction.

[0073] Especially noteworthy is that in various typical weather fluctuations such as data rising, falling and platform oscillation, the prediction curve can follow the rhythm of the real data changes, effectively avoiding common problems such as delayed response or excessive smoothing. This shows that the method has good time sensitivity and cycle capture ability, and can provide stable and reliable prediction results in complex and variable weather scenarios, with important practical value and deployment prospects.

[0074] The use of devices on the device is less limited, and the devices relied on by time series prediction inventions to achieve the same effect are much higher than the invention. The invention involves fewer parameters and runs faster, and can achieve the same prediction accuracy at a faster speed.

[0075] The above examples are only used to illustrate the technical solutions of the present application, but not to limit them; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that they can modify the technical solutions recorded in the foregoing examples, or make equivalent substitutions for part of the technical features; and these modifications or substitutions do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

Claims

1. A time series prediction method based on multi-scale decomposition and gated fusion, characterized in that, Includes the following steps: Perform multi-level downsampling operations on the original meteorological time series data, and output the downsampled time-scale meteorological time series; Based on the downsampled meteorological time series, a neural Fourier trend decomposition mechanism is applied to extract the trend and seasonal components at different scales of the original meteorological time series data. Trend components and seasonal components at different scales are fused through a gated multilayer sensing network to form meteorological time features with multiple periodic structures. The meteorological time features of the multi-period structure are split into channels, and the independent time-series features of each channel are output and embedded into a high-dimensional representation. The meteorological time features of the multi-period structure are normalized and then the prediction results are generated by the prediction branches of each scale. Finally, the multi-scale prediction output is obtained by weighted fusion. The training optimization is achieved by combining multiple loss functions to form a total loss function, which includes standard supervised loss, auxiliary loss, and regularization loss. Among them, a consistency regularization term is designed as an auxiliary loss based on high-dimensional representation features.

2. The time series prediction method based on multi-scale decomposition and gating fusion according to claim 1, characterized in that, The original meteorological time series data is a multivariate sequence recorded at fixed time intervals.

3. The time series prediction method based on multi-scale decomposition and gating fusion according to claim 2, characterized in that, Based on the downsampled meteorological time series, a neural Fourier trend decomposition mechanism was applied to extract the trend and seasonal components at different scales of the original meteorological time series data, including: Based on the downsampled meteorological time series, a fast Fourier transform is used to obtain the meteorological time series in complex frequency domain representation; The meteorological time series represented by the complex frequency domain is weighted according to the channel dimension, and the weights are learnable parameters to obtain the weighted result; Based on the weighted results, the real and imaginary parts are separated and nonlinear enhancement is performed. The real and imaginary parts are extracted separately, and a multilayer sensing network with residuals is used for nonlinear mapping to obtain the real and imaginary parts after nonlinear mapping. After merging and enhancing the real and imaginary parts after nonlinear mapping, an inverse Fourier transform is performed to obtain the result after inverse Fourier transform, which is the reconstructed seasonal component. The trend component is obtained by subtracting the reconstructed seasonal component from the original meteorological time series data input: For seasonal and trend components, a neural Fourier trend decomposition mechanism is performed to obtain the trend and seasonal components.

4. The time series prediction method based on multi-scale decomposition and gating fusion according to claim 1, characterized in that, Trend components and seasonal components at different scales are fused using a gated multilayer sensing network to form a multi-periodic structure of meteorological temporal features, including: The trend component and the seasonal component are combined to form a fused feature representation at each scale; The fused features at each scale are input into a gated multilayer perceptron. The fusion outputs at all scales are aligned to a uniform temporal dimension. Interpolation or temporal alignment operations are used to ensure that the outputs at different scales are consistent during the prediction phase. The fused outputs from all scales are stacked and summarized to form the final multi-period structure of meteorological time features.

5. The time series prediction method based on multi-scale decomposition and gating fusion according to claim 4, characterized in that, The gated multilayer sensing network includes multiple gated sensing units, and each gated sensing unit includes the following steps: The fused features are mapped to a high-dimensional space to obtain the fused features after dimensionality enhancement. The fused features after dimensionality upgrade are used to calculate the backbone features through the main branch, and the backbone features are then used to calculate the gating weights through the gating branch. The backbone features are multiplied by the gating weights to form a weighted output. The weighted output is then concatenated with the fused feature residuals and standardized using LayerNorm.

6. The time series prediction method based on multi-scale decomposition and gating fusion according to claim 5, characterized in that, The fused features after dimensionality upgrade are used to calculate the backbone features through the main branch, and the backbone features are then used to calculate the gating weights through the gating branch, including: The fused features after dimensionality upgrade are used to calculate the backbone features through the main branch: ； ； The main features are processed through gated branches, and the gate weights are calculated: ； in, This is the weight matrix; For bias; represents the Sigmoid activation function, used to generate the gating factor; H represents the backbone feature; O represents the fused feature after linear transformation of the original features; G represents the gating weight.

7. The time series prediction method based on multi-scale decomposition and gating fusion according to claim 1, characterized in that, Based on high-dimensional representation features, a consistency regularization term is designed as an auxiliary loss, including: Meteorological time series data X decomposed and output trend term based on neural network and seasonal items : ； Based on the high-dimensional representation features, a consistency regularization term is designed as an auxiliary loss: ； In the above formula, denoted as auxiliary loss; E represents the high-dimensional representation feature.

8. The time series prediction method based on multi-scale decomposition and gating fusion according to claim 7, characterized in that, Total loss function : ； in, Represents the total loss function; Indicates the regularization loss; Indicates the auxiliary loss weight; Indicates the regularization loss weight; This indicates standard monitoring losses.

Images