Multivariate singular spectrum analysis hybrid-based weather multivariate time series prediction method
Patent Information
- Application Number
- CN202611274498.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-21
- Publication Date
- 2026-11-03
AI Technical Summary
Existing meteorological multivariate time series forecasting methods suffer from problems such as loss of correlation and shape mismatch when dealing with multivariate meteorological variables, which affect the accurate modeling of meteorological evolution patterns.
A multivariate singular spectrum analysis hybrid approach is adopted to extract equal-length multi-scale components, preserve the correlation between meteorological variables, and combine trend-seasonal decomposition and stratified mixing to achieve efficient meteorological element prediction.
It effectively maintains the physical dependencies between meteorological variables, avoids causal confusion, and improves the accuracy and computational efficiency of long-term meteorological forecasts. It is applicable to fields such as regional temperature change forecasting, extreme weather warnings, and wind energy resource assessment.
Smart Images
Figure CN123030883A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of meteorological prediction and deep learning, and particularly relates to a meteorological multivariate time series prediction method based on multivariate singular spectrum analysis mixing. BACKGROUND
[0002] Meteorological time series prediction is a key technology for modern meteorological services and disaster warning, and is widely used in temperature prediction, precipitation prediction, wind speed estimation and other fields. With the development of deep learning, meteorological prediction models have gradually evolved from RNN, TCN, LSTM to Transformer and MLP architecture. In recent years, the research focus has gradually shifted from short-term meteorological prediction to long-term meteorological time series prediction, aiming to predict the changes of meteorological elements in the future hundreds of time steps to support long-term decision-making such as agricultural planning, energy scheduling, disaster prevention and mitigation.
[0003] Long-term meteorological time series prediction faces two major challenges: one is the non-stationarity of meteorological data, that is, the statistical characteristics of temperature, humidity and other variables change over time (such as seasonal fluctuations and climate change trends); the second is the architecture problem of whether to model the physical dependence relationship between each meteorological variable (temperature, humidity, wind speed, air pressure, etc.) in multivariate meteorological prediction. To solve non-stationarity, decomposition technology has become an important paradigm, which can decompose complex meteorological changes into structured components such as trends and seasonality. Early methods mainly rely on statistical techniques (such as STL decomposition), while modern methods integrate decomposition into deep learning frameworks.
[0004] However, existing decomposition methods have two fundamental limitations in meteorological prediction: one is that most methods are based on the channel-independent assumption, and each meteorological variable is processed separately, resulting in the loss of variable correlation (for example, the coupling relationship between temperature and humidity, wind speed and air pressure) in the representation learning stage; the second is that traditional decomposition operations such as down-sampling will produce sub-sequences of different lengths, causing shape mismatch problems, and then causing causal confusion when mixing information, affecting accurate modeling of meteorological evolution rules.
[0005] Therefore, how to extract multi-scale features in multivariate meteorological time series prediction while maintaining the physical dependence relationship between meteorological variables and avoiding the confusion caused by shape inconsistency has become a technical problem to be solved. SUMMARY
[0006] This invention aims to address the issues of variable dependency loss and shape mismatch in multivariate long-term meteorological time series forecasting. To this end, this invention proposes SpectralMixer, a multivariate meteorological time series forecasting method based on multivariate singular spectral analysis. This invention extracts equal-length multi-scale components through multivariate singular spectral analysis, preserves cross-variable correlations, and combines trend-seasonal decomposition and hierarchical mixing to achieve efficient meteorological element forecasting. It is applicable to high-dimensional time series forecasting tasks in meteorological fields such as regional temperature change forecasting, extreme weather warnings, and wind energy resource assessment.
[0007] To achieve the above-mentioned objectives, the present invention specifically adopts the following technical solution:
[0008] In a first aspect, the present invention provides a meteorological multivariate time series forecasting method based on multivariate singular spectral analysis, comprising the following steps:
[0009] S1. Obtain the time series of multiple meteorological variables. Each time series consists of the values of the corresponding meteorological variable at multiple historical time steps. Normalize the time series of each meteorological variable. The normalized time series of all meteorological variables form a meteorological multivariate time series.
[0010] S2. Input the meteorological multivariate time series into the trained prediction model. First, the time series of each meteorological variable after normalization is averaged and pooled. The time series of all meteorological variables after average pooling are used to form the low-frequency trend component. Perform dimension-aligned residual extraction operation on the meteorological multivariate time series and the low-frequency trend component to obtain the seasonal component.
[0011] S3. Perform multivariate singular spectrum analysis on the seasonal components to obtain the primary and secondary fluctuation components;
[0012] S4. Input the low-frequency trend component, the main fluctuation component, and the secondary fluctuation component into a prediction channel to obtain the prediction results of each prediction channel; after adding the prediction results of the three prediction channels, the normalized prediction values of each meteorological variable at multiple future time steps are obtained. The normalized prediction results composed of the normalized prediction values of all meteorological variables are inversely normalized to obtain the prediction time series.
[0013] As a preferred embodiment of the first aspect above, in S2, obtaining the seasonal component specifically involves: subtracting the normalized time series belonging to the same meteorological variable from the average pooled time series step by step to obtain the residual series of the meteorological variable, and the seasonal component is composed of the residual series of all meteorological variables.
[0014] As a preferred option in the first aspect mentioned above, in S3, the multivariate singular spectrum analysis of the seasonal component is performed as follows:
[0015] S31. Obtain the trajectory matrix for each meteorological variable. The trajectory matrices of each meteorological variable are combined into a unified trajectory matrix. ;
[0016] S32. Perform truncated singular value decomposition on the unified trajectory matrix to obtain the low-rank approximation matrix corresponding to the first r singular values with the largest singular values and the low-rank approximation matrix corresponding to the first 3r singular values, where r is a hyperparameter.
[0017] S33. Take the low-rank approximation matrix corresponding to the first r singular values as the main fluctuation subspace. The secondary fluctuation subspace is obtained by subtracting the low-rank approximation matrix corresponding to the first 3r singular values element-wise from the primary fluctuation subspace. ;
[0018] S34. Divide the main fluctuation subspace according to the meteorological variable dimension to obtain the principal subspace of each meteorological variable. The secondary fluctuation subspace is divided according to the meteorological variable dimension to obtain the secondary subspace of each meteorological variable. , making exist The position occupied in the middle exist The position occupied in the middle exist The positions occupied by all three are the same.
[0019] S35. Perform a diagonal average on the principal subspace of all meteorological variables to obtain the principal fluctuation sequence of each meteorological variable, and then the principal fluctuation sequence of all meteorological variables constitutes the principal fluctuation component; perform a diagonal average on the secondary subspace of all meteorological variables to obtain the secondary fluctuation sequence of each meteorological variable, and then the secondary fluctuation sequence of all meteorological variables constitutes the secondary fluctuation component.
[0020] As a preferred embodiment of the first aspect mentioned above, in S31, obtaining the trajectory matrix of each meteorological variable specifically involves performing Hankelization on the residual sequence of the c-th meteorological variable in the seasonal component, and then mapping the residual sequence into a trajectory matrix composed of time-delay vectors.
[0021] As a preferred embodiment of the first aspect mentioned above, in S32, according to Determine the number of singular values to be truncated, where k is the width of the Hankel window. This is for rounding up.
[0022] As a preferred embodiment of the first aspect above, in S4, the three prediction paths do not share parameters; each prediction path sequentially executes channel embedding, a linear layer for time projection, a GELU activation function, a linear layer for time projection, and a channel decoder to obtain the prediction result of that prediction path.
[0023] As a preferred embodiment of the first aspect described above, the channel decoder employs a linear layer.
[0024] In a second aspect, the present invention provides a computer program product, including a computer program / instruction, which, when executed by a processor, can realize the meteorological multivariate time series forecasting method based on multivariate singular spectrum analysis as described in any of the solutions of the first aspect above.
[0025] Thirdly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the meteorological multivariate time series forecasting method based on multivariate singular spectrum analysis as described in any of the embodiments of the first aspect above.
[0026] Fourthly, the present invention provides a computer electronic device, which includes a memory and a processor;
[0027] The memory is used to store computer programs;
[0028] The processor is configured to, when executing the computer program, implement the meteorological multivariate time series forecasting method based on multivariate singular spectrum analysis as described in any of the above-described first aspects.
[0029] Compared with the prior art, the present invention has the following advantages:
[0030] This invention supports both channel-mixed and channel-independent modes through Multivariate Singular Spectrum Analysis (MSSA), adapting to the variable dependency characteristics of different meteorological datasets (e.g., strong correlation between temperature and humidity in coastal areas, but potentially weaker in inland areas). It achieves shape-aligned multi-scale decomposition, avoids causal confusion, and better reflects the multi-scale evolution patterns of meteorological elements. This invention requires no strong prior assumptions, adaptively extracts data-driven meteorological temporal patterns, and its prediction model has a small parameter size and high computational efficiency, achieving state-of-the-art performance on multiple real meteorological datasets. Attached Figure Description
[0031] Figure 1 This is a schematic diagram of the overall architecture of the present invention;
[0032] Figure 2 This is a schematic diagram comparing the parameters of the present invention with those of other methods;
[0033] Figure 3 This diagram illustrates a comparison of the computational efficiency of the present invention with other methods. Detailed Implementation
[0034] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below. Technical features in the various embodiments of the present invention can be combined accordingly without mutual conflict.
[0035] In the description of this invention, it should be understood that the terms "first" and "second" are used only for descriptive purposes and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined with "first" and "second" may explicitly or implicitly include at least one of those features.
[0036] This embodiment takes the long-term prediction of multiple meteorological elements such as temperature, humidity, wind speed, and air pressure at a meteorological station in a certain area as an example, but the present invention is also applicable to other meteorological prediction scenarios such as precipitation probability prediction and solar radiation prediction.
[0037] like Figure 1 As shown, in a preferred embodiment of the present invention, the above-mentioned meteorological multivariate time series forecasting method based on multivariate singular spectral analysis includes the following steps S1 to S4. The specific implementation process of each step will be described in detail below.
[0038] S1. Obtain time series of multiple meteorological variables. Each time series consists of the values of the corresponding meteorological variable at multiple historical time steps. Normalize the time series of each meteorological variable. The normalized time series of all meteorological variables are combined to form a multivariate meteorological time series.
[0039] It should be noted that the meteorological variables used in step S1 of this invention include temperature, humidity, wind speed, air pressure, and rainfall. Subsequently, this invention calculates the mean value for each meteorological variable. with standard deviation This allows for the normalization of the time series of the meteorological variable, resulting in a normalized time series. The value of the c-th meteorological variable at historical time step t is then used as an example. For example, its normalized representation is:
[0040]
[0041] In the formula, Let c be the normalized value of the c-th meteorological variable at historical time step t; To prevent the denominator from being divided by a small constant.
[0042] After completing the above normalization process, the present invention further cleans the missing and outlier values in the time series in chronological order, thereby forming a meteorological multivariate time series from the normalized time series of meteorological variables. , For the number of meteorological variables, This represents the normalized length of the time series.
[0043] S2. Input the meteorological multivariate time series into the trained prediction model. First, the time series of each meteorological variable after normalization is averaged and pooled. The low-frequency trend component is composed of the time series of all meteorological variables after average pooling. Perform dimension-aligned residual extraction operation on the meteorological multivariate time series and the low-frequency trend component to obtain the seasonal component.
[0044] It should be noted that, in step S2 of the present invention, obtaining the seasonal component specifically involves: subtracting the normalized time series belonging to the same meteorological variable from the average pooled time series step by step to obtain the residual series of the meteorological variable, and the seasonal component is composed of the residual series of all meteorological variables.
[0045] In step S2 of this embodiment, if the normalized time series of each meteorological variable are arranged in rows, i.e., one row corresponds to one meteorological variable, to form a multivariate meteorological time series, then the low-frequency trend component... Similarly, the time series of each meteorological variable after average pooling are arranged row by row. Then, the meteorological multivariate time series is subtracted from the low-frequency trend component row by row; that is, the normalized time series of the same meteorological variable is always subtracted from its average pooled time series. This yields the residual series corresponding to that meteorological variable. After arranging the residual series of all meteorological variables row by row, the seasonal component is obtained. Furthermore, the seasonal component is consistent with the low-frequency trend component and the dimensions of the meteorological multivariate time series.
[0046] S3. Perform multivariate singular spectrum analysis on the seasonal components to obtain the primary and secondary fluctuation components.
[0047] It should be noted that in step S3 of this invention, the multivariate singular spectrum analysis (MSSA) of the seasonal component is performed as follows:
[0048] S31. Obtain the trajectory matrix of each meteorological variable and combine the trajectory matrices of each meteorological variable into a unified trajectory matrix.
[0049] In step S31 of the present invention, obtaining the trajectory matrix of each meteorological variable specifically involves performing Hankelization on the residual sequence of the c-th meteorological variable in the seasonal component, and then mapping the residual sequence into a trajectory matrix composed of time delay vectors.
[0050] In step S31 of this embodiment, during the construction of the trajectory matrix, the residual sequence of the c-th meteorological variable in the seasonal component is first... The variable is divided into several time-delay vectors, each corresponding to a column of the trajectory matrix. Then, the trajectory matrix of the c-th meteorological variable is constructed as shown in the following formula. Also known as the Henkel matrix:
[0051]
[0052] Where k is the width of the Hankel window; The total number of time delay vectors obtained from the segmentation; The residual sequence of the c-th meteorological variable is respectively the first... Each residual value.
[0053] After obtaining the trajectory matrices of all meteorological variables, this embodiment combines the trajectory matrices of different meteorological variables into a unified trajectory matrix. , For matrix transpose, The first The trajectory matrix of each meteorological variable.
[0054] S32. Perform truncated singular value decomposition (SVD) on the unified trajectory matrix to obtain the low-rank approximation matrices corresponding to the first r singular values with the largest singular values and the low-rank approximation matrices corresponding to the first 3r singular values, where r is a hyperparameter.
[0055] The truncated singular value decomposition in step S32 of this embodiment is an existing technology used to obtain key features required for analyzing multivariate meteorological time series. In the truncated singular value decomposition, the unified trajectory matrix is first decomposed into... .in, For a left singular matrix, it can be represented by a symmetric matrix. The feature vectors are obtained; A right singular matrix can be obtained through a symmetric matrix. The feature vectors are obtained; It is a diagonal matrix. Then, sort it according to the size of the singular values, truncate the first r singular values and the eigenvectors corresponding to these singular values, and obtain the low-rank approximation matrix corresponding to the first r singular values and the low-rank approximation matrix corresponding to the first 3r singular values.
[0056] The low-rank approximation matrix corresponding to the previous r singular values For example, it can be represented as For example, among which, This is the truncated left singular matrix; This is the truncated diagonal matrix; This is the truncated right singular matrix.
[0057] Furthermore, this embodiment is based on Determine the number of singular values to be truncated. This is for rounding up.
[0058] S33. Take the low-rank approximation matrix corresponding to the first r singular values as the main fluctuation subspace. The secondary fluctuation subspace is obtained by subtracting the low-rank approximation matrix corresponding to the first 3r singular values element-wise from the primary fluctuation subspace. .
[0059] S34. Divide the main fluctuation subspace according to the meteorological variable dimension to obtain the principal subspace of each meteorological variable. The secondary fluctuation subspace is divided according to the meteorological variable dimension to obtain the secondary subspace of each meteorological variable. , making exist The position occupied in the middle exist The position occupied in the middle exist They occupy the same position in the middle.
[0060] S35. Perform diagonal averaging on the principal subspace of each meteorological variable to obtain the principal fluctuation series of each meteorological variable. The main fluctuation components are composed of the main fluctuation sequences of all meteorological variables. Diagonal averages are applied to the secondary subspaces of all meteorological variables to obtain the secondary fluctuation sequences of each variable. Therefore, the secondary fluctuation components are composed of the secondary fluctuation sequences of all meteorological variables. .
[0061] In practical applications, by flexibly selecting different values of r, a series of primary and secondary fluctuation components containing information at different scales can be obtained. Taking the primary fluctuation component as an example, the primary fluctuation sequence of the c-th meteorological variable... Value Represented as:
[0062]
[0063] in, for The Middle Line 1 Values in the column; for The Middle Line 1 Values in the column; and For time step indexing.
[0064] In this embodiment, the key steps of MSSA decomposition are as follows: Figure 1 As shown, multivariate singular spectrum analysis can more accurately capture hidden trends, fluctuations, and noise in meteorological multivariate time series, which not only reduces noise interference but also provides better forecasting performance.
[0065] S4. Input the low-frequency trend component, the main fluctuation component, and the secondary fluctuation component into a prediction channel to obtain the prediction results of each prediction channel; after adding the prediction results of the three prediction channels, the normalized prediction values of each meteorological variable at multiple future time steps are obtained. The normalized prediction results composed of the normalized prediction values of all meteorological variables are inversely normalized to obtain the prediction time series.
[0066] It should be noted that in step S4 of this invention, the three prediction paths do not share parameters; each prediction path sequentially executes channel embedding, a linear layer for time projection, a GELU activation function, a linear layer for time projection, and a channel decoder to obtain the prediction result of that prediction path.
[0067] Furthermore, the channel decoder employs a linear layer.
[0068] In step S4 of this embodiment, the above-mentioned prediction pathways employ a hierarchical MLP architecture for feature mixing. The parameter size increases quadratically with the sequence length, resulting in high computational efficiency and suitability for long-term prediction of high-frequency data from meteorological observation stations. After obtaining the prediction results of each prediction pathway, this embodiment further combines the prediction results of the three prediction pathways. Add them together to get the normalized prediction result. Then, the mean value saved in step S1 is used. with standard deviation Perform inverse normalization to obtain the predicted time series in the original dimensions. , This represents the number of future time steps to be predicted.
[0069] It's also worth noting that during the prediction model training process, mean squared error or mean absolute error is used as the prediction loss, and adaptive momentum estimation (Adam) is used as the optimization algorithm. Instead of stochastic gradient descent (SGD), it jointly optimizes the parameters of the three prediction paths through backpropagation. Adam dynamically adjusts the learning rate of each parameter based on the average of past gradients and the average of past squared gradients. By tracking the first and second momentum of the gradients, Adam exhibits better performance and faster convergence in the task. Average pooling, Hankelization, singular value decomposition, subspace partitioning, and diagonal averaging constitute a deterministic data-driven decomposition process that does not require parameter updates.
[0070] As described above, steps S2-S3 belong to the hierarchical decoupling stage. Average pooling is used to extract slowly changing low-frequency trend components, while Hankelization and singular value decomposition are used to discover the dominant structure of seasonal components in the delayed coordinate space. Since all reconstructed components are restored to a sequence of the same length as the original input through diagonal averaging, shape mismatch and causal confusion during subsequent mixing caused by traditional downsampling decomposition are avoided. Step S4 belongs to the multi-path prediction stage. Low-frequency trend components, major fluctuation components, and minor fluctuation components correspond to different meteorological evolution models. Therefore, this invention uses channel embedding with non-shared parameters, a linear layer for time projection, and a channel decoder for modeling. This design enables each prediction path to learn the intervariate dependencies in low-frequency trends, stable periodic structures, and local fluctuations separately, reducing the risk of spurious correlations caused by heterogeneous model mixing.
[0071] To better demonstrate the specific implementation and technical effects of the present invention, the meteorological multivariate time series forecasting method based on multivariate singular spectrum analysis mixture shown in steps S1 to S4 of the above preferred implementation is applied to a specific example.
[0072] Example
[0073] The specific implementation process of the meteorological multivariate time series forecasting method based on multivariate singular spectral analysis mixture used in this embodiment is as described above and will not be repeated here. For the sake of simplicity, the above-mentioned meteorological multivariate time series forecasting method based on multivariate singular spectral analysis mixture will be referred to as SpectralMixer below.
[0074] This embodiment was performed on a desktop workstation equipped with an AMD Ryzen 9 5900HX CPU and an NVIDIA GeForce RTX 3080 GPU. The workstation ran a Linux operating system and was implemented using the Python 3.8.18 programming language. The prediction model was developed using the PyTorch 1.7.1 framework, and communication between the local host and the central server was established using sockets. The results of this embodiment were visualized using Matplotlib 3.3.1 and Microsoft Visio Professional 2021.
[0075] This embodiment uses the Weather dataset (containing 21 meteorological indicators such as temperature, humidity, wind speed, and air pressure from nearly 1600 weather stations in a certain country) and regional meteorological observation data for validation. Furthermore, to verify the generalization ability of this method in different fields, this embodiment also conducted comparative experiments on other public datasets (such as the ETT power transformer temperature dataset). The results are shown in Tables 1, 2, and 3, demonstrating that this method also achieves good performance.
[0076] Detailed parameter information used in this embodiment: The input time window is set to 96 (the number of historical observation points, corresponding to a certain number of past hours / days), and the prediction length is set to 96, 192, 336, and 720 (corresponding to meteorological element predictions for different future durations). The epoch and learning rate are set to 10 and 0.01, respectively. The mean squared error (MSE) is used as the loss function to measure the difference between the actual meteorological observations and the predicted values, aiming to minimize errors when training the prediction model.
[0077] Table 1. Comparison of different methods on the ETTh1 and ETTh2 datasets
[0078]
[0079] Table 2. Comparison of different methods on ETTm1 and ETTm2 datasets
[0080]
[0081] Table 3. Comparison of different methods on the Exchange and Weather datasets
[0082]
[0083] On three datasets, the method of this invention outperforms the baseline models (including TimeMixer, PatchTST, Transformer, etc.) in both MSE and MAE metrics, with an average improvement of approximately 8% to 15%, while the number of parameters is only 30% to 70% of the comparative models. This demonstrates that the method of this invention effectively solves the problems of shape mismatch and channel dependency loss in multivariate long-term time series forecasting, and has advantages such as high prediction accuracy, high parameter efficiency, and strong adaptability, making it suitable for intelligent forecasting tasks in the meteorological field.
[0084] like Figure 2 As shown, models closer to the bottom left corner have both lower prediction errors and fewer parameters. SpectralMixer is located in the lower MSE region and has significantly fewer parameters than many models with millions of parameters, indicating that this method maintains good prediction accuracy while having a lightweight model size.
[0085] like Figure 3 As shown, models closer to the bottom left corner generally perform better in terms of both prediction accuracy and computational efficiency. SpectralMixer is in the region with lower MSE and lower FLOPs, showing a better efficiency advantage compared to some models with higher computational costs; while some models, although with similar errors, have significantly higher computational overhead.
[0086] It is understood that the meteorological multivariate time series forecasting method based on multivariate singular spectral analysis mixture described in S1-S4 above can essentially be implemented by a computer program. Therefore, based on the same inventive concept, another preferred embodiment of the present invention also provides a computer program product corresponding to the meteorological multivariate time series forecasting method based on multivariate singular spectral analysis mixture provided in the above embodiments. This product includes a computer program / instruction, which, when executed by a processor, can implement the meteorological multivariate time series forecasting method based on multivariate singular spectral analysis mixture as described in the above embodiments.
[0087] Similarly, based on the same inventive concept, another preferred embodiment of the present invention also provides a computer electronic device corresponding to the meteorological multivariate time series forecasting method based on multivariate singular spectrum analysis provided in the above embodiment, which includes a memory and a processor;
[0088] The memory is used to store computer programs;
[0089] The processor is configured to implement the meteorological multivariate time series forecasting method based on multivariate singular spectrum analysis mixture in the above embodiments when executing the computer program.
[0090] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.
[0091] Therefore, based on the same inventive concept, another preferred embodiment of the present invention also provides a computer-readable storage medium corresponding to the meteorological multivariate time series forecasting method based on multivariate singular spectral analysis hybrid provided in the above embodiments. The storage medium stores a computer program, which, when executed by a processor, can realize the meteorological multivariate time series forecasting method based on multivariate singular spectral analysis hybrid provided in the above embodiments.
[0092] Specifically, in the computer-readable storage medium of the above three embodiments, the stored computer program is executed by a processor, which can perform the aforementioned steps S1 to S4.
[0093] It is understood that the aforementioned storage media may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Furthermore, the storage media may also be various media capable of storing program code, such as USB flash drives, external hard drives, magnetic disks, or optical discs.
[0094] It is understood that the processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.
[0095] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the invention. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the invention. Therefore, all technical solutions obtained through equivalent substitution or transformation fall within the protection scope of the present invention.
Claims
1. A meteorological multivariate time series forecasting method based on multivariate singular spectral analysis, characterized in that, Includes the following steps: S1. Obtain the time series of multiple meteorological variables. Each time series consists of the values of the corresponding meteorological variable at multiple historical time steps. Normalize the time series of each meteorological variable. The normalized time series of all meteorological variables form a meteorological multivariate time series. S2. Input the meteorological multivariate time series into the trained prediction model. First, the time series of each meteorological variable after normalization is averaged and pooled. The time series of all meteorological variables after average pooling are used to form the low-frequency trend component. Perform dimension-aligned residual extraction operation on the meteorological multivariate time series and the low-frequency trend component to obtain the seasonal component. S3. Perform multivariate singular spectrum analysis on the seasonal components to obtain the primary and secondary fluctuation components; S4. Input the low-frequency trend component, the main fluctuation component, and the secondary fluctuation component into a prediction channel to obtain the prediction results of each prediction channel; after adding the prediction results of the three prediction channels, the normalized prediction values of each meteorological variable at multiple future time steps are obtained. The normalized prediction results composed of the normalized prediction values of all meteorological variables are inversely normalized to obtain the prediction time series.
2. The meteorological multivariate time series forecasting method based on multivariate singular spectral analysis as described in claim 1, characterized in that, In S2, the seasonal component is obtained by subtracting the normalized time series belonging to the same meteorological variable from the average pooled time series step by step to obtain the residual series of the meteorological variable. The seasonal component is composed of the residual series of all meteorological variables.
3. The meteorological multivariate time series forecasting method based on multivariate singular spectral analysis as described in claim 2, characterized in that, In S3, the multivariate singular spectrum analysis of the seasonal components is performed as follows: S31. Obtain the trajectory matrix for each meteorological variable. The trajectory matrices of each meteorological variable are combined into a unified trajectory matrix. ; S32. Perform truncated singular value decomposition on the unified trajectory matrix to obtain the low-rank approximation matrix corresponding to the first r singular values with the largest singular values and the low-rank approximation matrix corresponding to the first 3r singular values, where r is a hyperparameter. S33. Take the low-rank approximation matrix corresponding to the first r singular values as the main fluctuation subspace. The secondary fluctuation subspace is obtained by subtracting the low-rank approximation matrix corresponding to the first 3r singular values element-wise from the primary fluctuation subspace. ; S34. Divide the main fluctuation subspace according to the meteorological variable dimension to obtain the principal subspace of each meteorological variable. The secondary fluctuation subspace is divided according to the meteorological variable dimension to obtain the secondary subspace of each meteorological variable. , making exist The position occupied in the middle exist The position occupied in the middle exist The positions occupied by all three are the same. S35. Perform a diagonal average on the principal subspace of all meteorological variables to obtain the principal fluctuation sequence of each meteorological variable, and then the principal fluctuation sequence of all meteorological variables constitutes the principal fluctuation component; perform a diagonal average on the secondary subspace of all meteorological variables to obtain the secondary fluctuation sequence of each meteorological variable, and then the secondary fluctuation sequence of all meteorological variables constitutes the secondary fluctuation component.
4. The meteorological multivariate time series forecasting method based on multivariate singular spectral analysis as described in claim 3, characterized in that, In S31, obtaining the trajectory matrix of each meteorological variable specifically involves performing Hankelization on the residual sequence of the c-th meteorological variable in the seasonal component, and then mapping the residual sequence into a trajectory matrix composed of time-delay vectors.
5. The meteorological multivariate time series forecasting method based on multivariate singular spectral analysis as described in claim 3, characterized in that, In S32, according to Determine the number of singular values to be truncated, where k is the width of the Hankel window. This is for rounding up.
6. The meteorological multivariate time series forecasting method based on multivariate singular spectral analysis as described in claim 1, characterized in that, In S4, the three prediction paths do not share parameters; each prediction path sequentially executes channel embedding, a linear layer for time projection, a GELU activation function, a linear layer for time projection, and a channel decoder to obtain the prediction result of that prediction path.
7. The meteorological multivariate time series forecasting method based on multivariate singular spectral analysis as described in claim 6, characterized in that, The channel decoder uses a linear layer.
8. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the processor, it can realize the meteorological multivariate time series forecasting method based on multivariate singular spectrum analysis as described in any one of claims 1 to 7.
9. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the meteorological multivariate time series forecasting method based on multivariate singular spectrum analysis as described in any one of claims 1 to 7.
10. A computer electronic device, characterized in that, Including memory and processor; The memory is used to store computer programs; The processor is configured to, when executing the computer program, implement the meteorological multivariate time series forecasting method based on multivariate singular spectrum analysis as described in any one of claims 1 to 7.