Covariant time sequence prediction method based on multi-scale decoupling

Abstract

The invention relates to a covariable time sequence prediction method based on multi-scale decoupling. The method comprises the following steps: acquiring a target variable historical sequence X; carrying out average pooling down-sampling processing to obtain target variable input of multiple sampling scales; performing multi-scale analysis to obtain target variable multi-scale mixed representation; and obtaining a covariable Z, performing time sequence decomposition on the multi-scale mixed representation of the target variable to obtain multi-scale periodic term and trend term representation, modeling the influence of the covariable Z on the historical sequence X of the target variable, and obtaining a prediction result Y of a future sequence. The method has the beneficial effects that a multi-scale analysis and time sequence decomposition technology is utilized, a time sequence multi-level dominant change mode is accurately modeled, time sequence components are decoupled to refine the difference influence of a covariable on different change modes of a target variable, the modeling bottleneck of an existing method on time sequence modeling complexity and covariable influence heterogeneity is broken through, and the time sequence modeling complexity and covariable influence heterogeneity are improved. The defects of an existing covariable time sequence prediction method in target variable and covariable information modeling are overcome.

Description

Technical Field

[0001] The present invention relates to the technical field of time series prediction, and more specifically, to a covariate time series prediction method based on multi-scale decoupling. Background Art

[0002] Time series forecasting refers to the process of predicting future changes based on the historical changes in the time series. It is an important time series analysis technology that is widely used in finance, climate, energy, transportation and other fields to provide a scientific basis for application decisions. Changes in time series data are often affected by external factors. For example, electricity prices are highly dependent on the supply and demand relationship in the market. At this time, it is very difficult to predict future prices based on historical data alone. There are usually multiple variables in a time series system, among which the target variable is the variable that needs to be predicted, and the covariate is a time series variable that does not need to be predicted but can provide auxiliary information for the prediction of the target variable. There is a correlation between the changes in the covariate and the target variable. The covariate time series prediction model can take into account the unequal role of the covariate and the target variable in the prediction, and can effectively use the covariate information to assist the prediction of the target variable. It is an important time series prediction technology.

[0003] The limitation of the existing covariate time series forecasting methods is that they only focus on the selection of covariates and the information interaction modeling in the variable dimension, ignoring the inherent complex change characteristics of time series data. Specifically, time series data presents mixed change patterns at different sampling scales (such as small scale focuses on periodicity and large scale focuses on trend), and the existing methods do not decompose the intrinsic structure of the time series (such as periodic terms and trend terms), resulting in the blurring of the differential effects of covariates on the short-term fluctuations and long-term trends of the target variable; at the same time, the correlation between the covariate and the target variable has dynamic differences in different sampling scales and decomposed trend terms / periodic terms (for example, the impact of weather on the periodic disturbance of daily electricity consumption is stronger than the effect on the annual trend), but the traditional covariate time series forecasting method does not make a targeted distinction in the single variable granularity modeling, resulting in the homogenization of the covariate impact, which limits the model performance. Summary of the invention

[0004] The purpose of the present invention is to address the deficiencies of the prior art and propose a covariate time series prediction method based on multi-scale decoupling.

[0005] First, a covariate time series prediction method based on multi-scale decoupling is provided, including:

[0006] S1. Obtain the historical sequence X of the target variable;

[0007] S2, performing average pooling downsampling processing on the target variable history sequence X to obtain target variable inputs of multiple sampling scales;

[0008] S3, performing multi-scale analysis on the target variable input of the multi-sampling scales to obtain a multi-scale mixed representation of the target variable;

[0009] S4. Obtain the covariate Z, perform time series decomposition on the multi-scale mixed representation of the target variable to obtain multi-scale periodic term and trend term representations, model the impact of the covariate Z on the historical sequence X of the target variable, and obtain the prediction result Y of the future sequence.

[0010] Preferably, S4 includes:

[0011] S401, obtaining a covariate Z, and performing time series decomposition on a multi-scale mixed representation of a target variable to obtain a multi-scale period term and trend term representation;

[0012] S402, using the trend item of the minimum sampling scale to characterize and predict the trend item of the corresponding future sequence;

[0013] S403, embedding the multi-scale periodic term representation in the Patch dimension, dividing the multi-scale periodic term representation into non-overlapping time segments Patch, then extracting local features and encoding position information; and embedding the covariate Z in the Patch dimension;

[0014] S404, through the stackable hybrid attention mechanism of L layers, the influence of the covariate Z on the historical sequence X of the target variable is modeled to obtain the prediction information of the future sequence period items;

[0015] S405, aggregating the prediction results corresponding to different scales to obtain the final prediction results of future sequence period items;

[0016] S406: Add the trend item prediction result and the period item prediction result of the future sequence to obtain the prediction result Y of the future sequence.

[0017] Preferably, S404 includes:

[0018] S4041, the embedding of covariate Z and multi-scale target variables learns their own change pattern information through different self-attention mechanisms;

[0019] S4042. Through the cross-attention mechanism, with the target variable embedding Ps as the query and the covariate embedding Pz as the key and value, the influence of the covariate Z on the target variable historical sequence X is modeled.

[0020] As a preference, it also includes:

[0021] S5. Calculate the loss function, where the loss function is the error between the predicted result Y of the future sequence and the true value.

[0022] Preferably, in S3, the target variable input of multiple sampling scales is decomposed into periodic terms and trend terms, and the periodic terms are mixed from bottom to top and the trend terms are mixed from top to bottom, the periodic terms and trend terms corresponding to different scales are combined, and a convolutional layer is used to obtain a multi-scale mixed representation of the target variable.

[0023] In a second aspect, a covariate time series prediction system based on multi-scale decoupling is provided, which is used to execute any method described in the first aspect, including:

[0024] The first acquisition module is used to obtain the historical sequence X of the target variable;

[0025] An average pooling downsampling module is used to perform average pooling downsampling processing on the target variable history sequence X to obtain target variable inputs of multiple sampling scales;

[0026] A multi-scale analysis module, used to perform multi-scale analysis on the target variable input of the multi-sampling scales to obtain a multi-scale mixed representation of the target variable;

[0027] The second acquisition module is used to obtain the covariate Z, perform time series decomposition on the multi-scale mixed representation of the target variable to obtain multi-scale periodic term and trend term representation, model the impact of the covariate Z on the historical sequence X of the target variable, and obtain the prediction result Y of the future sequence.

[0028] According to a third aspect, a computer storage medium is provided, wherein a computer program is stored in the computer storage medium; when the computer program is executed on a computer, the computer executes any method described in the first aspect.

[0029] In a fourth aspect, an electronic device is provided, including:

[0030] Memory, used to store computer programs;

[0031] A processor is used to execute the computer program to implement any method as described in the first aspect.

[0032] The beneficial effects of the present invention are as follows: the present invention utilizes multi-scale analysis and time series decomposition technology to accurately model multi-level dominant change patterns of time series, and decouples time series components to refine the differential impacts of covariates on different change patterns of target variables (such as short-term cycles and long-term trends), breaking through the modeling bottlenecks of existing methods for the complexity of time series modeling and the heterogeneity of covariate influences, improving the deficiencies of existing covariate time series prediction methods in modeling target variables and covariate information, thereby improving time series prediction performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 A flowchart of the covariate time series prediction method based on multi-scale decoupling provided in this application;

[0034] Figure 2 A schematic diagram of the structure of the past mixed decomposition module provided for this application;

[0035] Figure 3 A schematic diagram of the structure of the covariate cross-attention module provided for this application;

[0036] Figure 4 Schematic diagram of the stackable hybrid attention mechanism provided for this application. DETAILED DESCRIPTION

[0037] The present invention is further described below in conjunction with embodiments. The description of the following embodiments is only used to help understand the present invention. It should be noted that for ordinary persons in the art, without departing from the principle of the present invention, the present invention can also be modified in some ways, and these improvements and modifications also fall within the scope of protection of the claims of the present invention.

[0038] Embodiment 1:

[0039] Before introducing the technical solution of the present application, it is necessary to explain the definitions of relevant key terms and the prior art. Specifically, time series data refers to a series of data points arranged in chronological order. In time series data, time is usually a separate variable, and the order of data points reflects the temporal development of events or observations. Time series data is usually used in fields such as statistical analysis, predictive models, and system monitoring.

[0040] There are usually multiple variables in a time series system, where the target variable is the variable that needs to be predicted, and the covariate is a time series variable that does not need to be predicted but can provide auxiliary information for the prediction of the target variable. The covariate time series prediction method takes into account the unequal role of covariates and target variables in prediction, and can effectively use covariate information to assist the prediction of target variables. It is an important time series prediction paradigm.

[0041] Time Series Decompositions is a technology used for time series forecasting. It can decompose time series into several different components, such as trend terms, periodic terms, and random terms, which respectively represent different time series change information such as long-term evolution rules, repetitive fluctuations, and random noise, thereby improving the adaptability of the forecasting model to complex time series patterns. By decomposing time series, we can better understand the data structure and improve the time series forecasting performance. For example, using the trend decomposition module in the Autoformer model, the multi-scale time series set can be decomposed:

[0042]

[0043] Where s represents the periodic term under the current sampling scale, and t represents the trend term.

[0044] Time series multi-scale analysis technology is a time series data modeling technology. By hierarchically modeling the dynamic change characteristics of time series at different time granularities (such as seconds, hours, days, and months), it can capture mixed change patterns such as short-term cycles and long-term trends implicit in the data. Its core lies in avoiding over-simplification of complex time series patterns by single-scale modeling through scale-adaptive feature extraction (such as sliding window averaging and hierarchical attention mechanism).

[0045] For example, the process of obtaining time series of different sampling scales by downsampling is as follows:

[0046]

[0047] Where T is the length of the original time series, so the time series set of different sampling scales obtained according to the original time series is:

[0048]

[0049] Different elements represent the change information of the original time series at different sampling scales, and M is the number of downsampling times.

[0050] TimeXer is a recently proposed covariate time series prediction model based on the Transformer architecture. It enhances the prediction ability of the target variable by introducing covariates. The input of this model is the historical sequence X of the target variable and the covariate Z, the specific size is:

[0051]

[0052]

[0053] Where T and Tex are the lengths of the sequences, and C is the number of covariates. The purpose of the model is to predict the target variable in the next H steps based on X and Z:

[0054]

[0055] Considering the different effects of target variables and covariates on the prediction results in the covariate time series prediction task, TimeXer adopts different embedding strategies for the target variable historical sequence X and covariate Z. For covariate Z, since it only plays an auxiliary role in the prediction of the target variable, only variable-level embedding is performed, and each covariate is directly embedded into the D dimension to reduce the modeling complexity:

[0056]

[0057] For the target variable X, considering the dominant role of the change information it contains in the prediction task, the variable dimension and the patch dimension are embedded respectively. The first is the variable dimension embedding similar to the covariate Z:

[0058]

[0059] Then the Patch dimension is embedded. First, X is divided into non-overlapping time segments Patch, and then local features are extracted and position information is encoded:

[0060]

[0061] Patchify is a slicing operation that divides the T-step sequence into N=T / L segments, where L is the segment length. Embed is implemented through linear layers and activation function layers to map each segment to a D-dimensional space. PE is a cosine position encoding in time order to help the model capture time changes. For the tth segment, the corresponding D-dimensional cosine position encoding is:

[0062]

[0063] Where i represents the i-th element of the D-dimensional position encoding.

[0064] TimeXer then uses two different attention mechanisms to build stackable attention modules to model covariate information and target variable information. For the l+1th attention module, the self-attention mechanism is first used to model the interaction between the local change pattern of the target variable X and the global trend information:

[0065]

[0066] Then we use cross attention to explicitly model the covariate influence, with Ex as Query, Ez as Key and Value, we have:

[0067]

[0068] Thus, the covariate information can be introduced into Ex, and the covariate information can be introduced into Px through the interaction between Ex and Px in the next layer of attention module. Here, Ex can be regarded as a global token, serving as an intermediary for information transmission between Px and Ez.

[0069] After the target variable X and covariate Z are processed by L layers of stacked attention modules, the prediction result can be obtained by splicing patches and then transforming:

[0070]

[0071] Projection is the projection layer, which is usually implemented using two linear layers and one activation function layer. During the training process, the loss function of the model is:

[0072]

[0073] That is, the L2 norm of the error between the real future sequence and the predicted value is used as the loss function for training.

[0074] The existing technology TimeXer achieves improved prediction performance through covariate embedding and hybrid attention mechanism when modeling covariate information. However, it does not consider that the target variable may have different correlations with the covariate at different sampling scales, and only models it at the original scale, making it difficult to distinguish the differential impact of covariates at different scales; and does not explicitly distinguish between local cycle and long-term trend change information, which increases the modeling complexity of the model and reduces the room for performance improvement. The problem arises because the existing hybrid attention mechanism and embedding strategy cannot meet the challenges of the model in multi-scale analysis and time series decomposition, and it is necessary to introduce relevant technologies and design the corresponding model structure.

[0075] In order to solve the problem that the existing technology is limited to the covariate selection and interaction modeling of the variable dimension, it fails to effectively capture the complex dynamic patterns of the target variable time series at multiple sampling scales (such as hours and months) (small scale focuses on periodicity, large scale focuses on trend), and due to the lack of decoupling of the intrinsic structure of the time series (such as undecomposed periodic terms and trend terms), the differentiated impact of the covariate on the short-term fluctuations and long-term trends of the target variable is blurred, thereby restricting the problem of prediction accuracy. Example 1 of the present application provides a covariate time series prediction (MDF) method based on multi-scale decoupling. By introducing multi-scale analysis and time series decomposition technology, it can accurately identify the multi-level dominant patterns of the time series, decompose the time series components to refine the dynamic impact of the covariate on different time granularities, and finally achieve the improvement of prediction accuracy in complex time series scenarios.

[0076] Specifically, Figure 1 As shown, the method includes:

[0077] S1. Obtain the historical sequence X of the target variable.

[0078] In the actual implementation of the technical solution, thanks to the parallel computing design of existing deep learning frameworks (such as Pytorch and Tensorflow), the actual shape of the target variable X is [B, T], and the actual shape of the covariate Z obtained by S4 is [B, C, Tex], that is, B input samples are input in parallel. Since the calculation process of the B input samples is the same, the following is a simplified description, and only a single sample is used to introduce the technical solution of this application.

[0079] S2. Perform average pooling downsampling processing on the target variable history sequence X to obtain target variable inputs of multiple sampling scales.

[0080] Specifically, the formula of S2 is expressed as:

[0081]

[0082] S3. Perform multi-scale analysis on the target variable input at multiple sampling scales to obtain a multi-scale mixed representation of the target variable.

[0083] Specifically, the past mixed decomposition module PDM is used to mix the change information in the historical sequence of the multi-scale target variable to obtain the preliminary representation of the target variable X. The structure of a single-layer PDM is as follows Figure 2 shown.

[0084] Since PDM is a stackable structure, a total of N layers of PDM modules are used in the method provided in this application. For the nth layer of PDM module, the sequence decomposition technology in Autoformer is first used to decompose the multi-sampling scale target variable input into a period term s and a trend term t:

[0085]

[0086] Then, considering that the smaller sampling scale sequence contains more micro-periodic item information, and the larger periodic information can also be regarded as the aggregation of periods, the periodic items are mixed from bottom to top:

[0087]

[0088] Bottom-Up-Mixing processes the time dimension through two linear layers, including an activation function layer. Smaller sampling scales contain more periodic information, which may introduce noise when the model captures the macro trend information, while larger sampling scale sequences are more likely to provide clear trend change information, so the trend items are mixed from top to bottom:

[0089]

[0090] The structures of Top-Down-Mixing and Bottom-Up-Mixing are similar. After the mixing is completed, PDM combines the periodic terms and trend terms corresponding to different scales again, and uses the convolution layer to enhance the representation ability of the multi-scale target variable sequence:

[0091]

[0092] FeedForward consists of two one-dimensional convolutional layers, including an activation function layer, which extracts representation information by performing convolution operations in the time dimension. After the target variable of multiple sampling scales is input and processed by the N-layer PDM module, the multi-scale mixed representation of the target variable is finally output:

[0093]

[0094] S4. Obtain the covariate Z, perform time series decomposition on the multi-scale mixed representation of the target variable to obtain multi-scale periodic term and trend term representations, model the impact of the covariate Z on the historical sequence X of the target variable, and obtain the prediction result Y of the future sequence.

[0095] Embodiment 2:

[0096] On the basis of Example 1, Example 2 of the present application provides a more specific covariate time series prediction method based on multi-scale decoupling, including:

[0097] S1. Obtain the historical sequence X of the target variable.

[0098] S2. Perform average pooling downsampling processing on the target variable history sequence X to obtain target variable inputs of multiple sampling scales.

[0099] S3. Perform multi-scale analysis on the target variable input at multiple sampling scales to obtain a multi-scale mixed representation of the target variable.

[0100] S4. Obtain the covariate Z, perform time series decomposition on the multi-scale mixed representation of the target variable to obtain multi-scale periodic term and trend term representations, model the impact of the covariate Z on the historical sequence X of the target variable, and obtain the prediction result Y of the future sequence.

[0101] In S4, this application uses the covariate cross attention module (EXCA for short) to model the impact of covariates on the target variable. The structure of this module is as follows Figure 3 Compared with the existing covariate time series prediction methods, this structure innovatively combines multi-scale analysis and time series decomposition technology, and designs an efficient network structure based on the attention mechanism. It models the differential impact of covariates on target variables at different levels in different scales and periodic trend items, thereby better utilizing covariate information to assist target variable prediction and achieving advanced covariate time series prediction performance.

[0102] Specifically, S4 includes:

[0103] S401, obtain the covariate Z, and perform time series decomposition on the multi-scale mixed representation of the target variable to obtain multi-scale periodic term and trend term representation, which is expressed as:

[0104]

[0105] S402. Considering that the change information contained in the trend item is relatively simple, the trend item with the minimum sampling scale is directly used to represent the trend item of the corresponding future sequence, which is expressed as:

[0106]

[0107] S403, embed the multi-scale periodic term representation in the Patch dimension, divide the multi-scale periodic term representation into non-overlapping time segments Patch, then extract local features and encode position information, expressed as:

[0108]

[0109] Where PE is the cosine position encoding. The same is true for the covariate Z:

[0110]

[0111] S404. After obtaining the multi-scale target variable periodic term representation and the embedding Ps and Pz of the covariate Z in the Patch dimension, the influence of the covariate on the target variable can be modeled through an L-layer stackable mixed attention mechanism (Mix-Attention for short) to obtain the prediction information of the future sequence periodic term.

[0112] like Figure 4 As shown, S404 includes:

[0113] S4041, the embedding of covariate Z and multi-scale target variables learns their own change pattern information through different self-attention mechanisms;

[0114] S4042. Through the cross-attention mechanism, the target variable embedding Ps is used as the query (Query), and the covariate embedding Pz is used as the key (Key) and value (Value), to model the impact of the covariate Z on the target variable historical sequence X. For the Mix-Attention mechanism of the lth layer, there are:

[0115]

[0116]

[0117]

[0118] S405: Aggregate the prediction results corresponding to different scales to obtain the final prediction results of the future sequence period items.

[0119] It should be noted that in the EXCA module, after the Patch dimension embedding of the target variable and the covariate has passed through the L-layer hybrid attention mechanism, the target variable embedding Ps contains prediction information that can be used to predict future sequence period items. In order to utilize the complementary prediction capabilities of multi-scale time series, multiple prediction components are used to obtain prediction results corresponding to different scales, and the final prediction results of future sequence period items can be obtained by aggregating them, which is expressed as:

[0120]

[0121] S406, adding the trend item prediction result and the period item prediction result of the future sequence to obtain the prediction result Y of the future sequence, which is expressed as:

[0122]

[0123] In addition, since the purpose of the model is to accurately predict the true value of the future series, this method also includes:

[0124] S5. Calculate the loss function, which is the error between the predicted result Y of the future sequence and the true value, expressed as:

[0125]

[0126] By combining multi-scale analysis and time series decomposition technology and designing a reasonable network structure, the MDF model proposed in the present invention achieves better time series prediction performance than previous covariate time series prediction methods, and has certain scientificity and practicality. On the short-term electricity price forecasting dataset EPF, MDF is compared with various advanced time series prediction methods, including two Transformer-based models: Autoformer and iTransformer, two MLP-based models: TimeMixer and SOFTS, and two covariate prediction models: TimeXer and TiDE. Except for the two covariate prediction methods TimeXer and TiDE, the other four are multivariate prediction methods. Although the multivariate prediction paradigm treats all time series variables equally, these four models all have modeling of variable dependencies, so they can also use covariate information to assist target variable prediction to a certain extent, and are comparable. Using MSE as the evaluation indicator, the experimental results are shown in Table 1 below:

[0127] Table 1

[0128]

[0129] It can be found that MDF is always better than other advanced covariate / non-covariate time series forecasting models in all data sets, achieving advanced forecasting performance.

[0130] It should be noted that the parts in this embodiment that are the same or similar to those in Embodiment 1 can be referenced to each other and will not be described in detail in this application.

[0131] Embodiment 3:

[0132] On the basis of embodiments 1 and 2, embodiment 3 of the present application provides a covariate time series prediction system based on multi-scale decoupling, including:

[0133] The first acquisition module is used to obtain the historical sequence X of the target variable;

[0134] An average pooling downsampling module is used to perform average pooling downsampling processing on the target variable history sequence X to obtain target variable inputs of multiple sampling scales;

[0135] A multi-scale analysis module, used to perform multi-scale analysis on the target variable input of the multi-sampling scales to obtain a multi-scale mixed representation of the target variable;

[0136] The second acquisition module is used to obtain the covariate Z, perform time series decomposition on the multi-scale mixed representation of the target variable to obtain multi-scale periodic term and trend term representation, model the impact of the covariate Z on the historical sequence X of the target variable, and obtain the prediction result Y of the future sequence.

[0137] It should be noted that the system provided in this embodiment is a system corresponding to the method provided in Embodiments 1 and 2. Therefore, the parts in this embodiment that are the same or similar to Embodiments 1 and 2 can be referenced to each other and will not be repeated in this application.

[0138] In summary, this application improves the problem of previous covariate time series prediction methods ignoring the differences in covariate influences under complex time series change patterns by combining multi-scale analysis and time series decomposition technology and designing a reasonable network structure, and better utilizes covariate information to assist target variable prediction, achieving good prediction performance.

Claims

1. A covariate time series prediction method based on multi-scale decoupling, characterized in that: include: S1. Obtain the historical sequence X of the target variable; S2, performing average pooling downsampling processing on the target variable history sequence X to obtain target variable inputs of multiple sampling scales; S3, performing multi-scale analysis on the target variable input of the multi-sampling scales to obtain a multi-scale mixed representation of the target variable; S4. Obtain the covariate Z, perform time series decomposition on the multi-scale mixed representation of the target variable to obtain multi-scale periodic term and trend term representations, model the impact of the covariate Z on the historical sequence X of the target variable, and obtain the prediction result Y of the future sequence.

2. The method for predicting covariate time series based on multi-scale decoupling according to claim 1, characterized in that S4 include: S401, obtaining a covariate Z, and performing time series decomposition on a multi-scale mixed representation of a target variable to obtain a multi-scale period term and trend term representation; S402, using the trend item of the minimum sampling scale to characterize and predict the trend item of the corresponding future sequence; S403, embedding the multi-scale periodic term representation in the Patch dimension, dividing the multi-scale periodic term representation into non-overlapping time segments Patch, then extracting local features and encoding position information; and embedding the covariate Z in the Patch dimension; S404, through the stackable hybrid attention mechanism of L layers, the influence of the covariate Z on the historical sequence X of the target variable is modeled to obtain the prediction information of the future sequence period items; S405, aggregating the prediction results corresponding to different scales to obtain the final prediction results of future sequence period items; S406: Add the trend item prediction result and the period item prediction result of the future sequence to obtain the prediction result Y of the future sequence.

3. The covariate time series prediction method based on multi-scale decoupling according to claim 2 is characterized in that: S404 includes: S4041, the embedding of covariate Z and multi-scale target variables learns their own change pattern information through different self-attention mechanisms; S4042. Through the cross-attention mechanism, with the target variable embedding Ps as the query and the covariate embedding Pz as the key and value, the influence of the covariate Z on the target variable historical sequence X is modeled.

4. The covariate time series prediction method based on multi-scale decoupling according to claim 3 is characterized in that: Also includes: S5. Calculate the loss function, where the loss function is the error between the predicted result Y of the future sequence and the true value.

5. The covariate time series prediction method based on multi-scale decoupling according to claim 4 is characterized in that: In S3, the target variable input of multiple sampling scales is decomposed into periodic terms and trend terms, and the periodic terms are mixed from bottom to top and the trend terms are mixed from top to bottom. The periodic terms and trend terms corresponding to different scales are combined, and the convolutional layer is used to obtain the multi-scale mixed representation of the target variable.

6. A covariate time series prediction system based on multi-scale decoupling, characterized in that: The method for executing any one of claims 1 to 5 comprises: The first acquisition module is used to obtain the historical sequence X of the target variable; An average pooling downsampling module is used to perform average pooling downsampling processing on the target variable history sequence X to obtain target variable inputs of multiple sampling scales; A multi-scale analysis module, used to perform multi-scale analysis on the target variable input of the multi-sampling scales to obtain a multi-scale mixed representation of the target variable; The second acquisition module is used to obtain the covariate Z, perform time series decomposition on the multi-scale mixed representation of the target variable to obtain multi-scale periodic term and trend term representation, model the impact of the covariate Z on the historical sequence X of the target variable, and obtain the prediction result Y of the future sequence.

7. A computer storage medium, characterized in that: The computer storage medium stores a computer program; when the computer program is executed on a computer, the computer executes any one of the methods described in claims 1 to 5.

8. An electronic device, characterized in that: include: Memory, used to store computer programs; A processor, configured to execute the computer program to implement the method according to any one of claims 1 to 5.

Images