Short-term trend prediction method and system for multidimensional time series data
Through the combination of timing feature encoder and auxiliary information encoder, the N-BEATS architecture and global attention mechanism are used to solve the problem of inaccurate prediction of multi-dimensional timing data, and the accurate prediction of power load is achieved.
Patent Information
- Application Number
- CN202310182120.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-02-24
- Publication Date
- 2025-08-26
- Estimated Expiration
- 2043-02-24
AI Technical Summary
The prior art is difficult to effectively capture the correlation characteristics of multi-dimensional time series data, resulting in inaccurate short-term predictions.
The multi-dimensional timing data prediction method based on timing feature encoder, auxiliary information encoder and prediction decoder is adopted to obtain the characteristics of multi-dimensional timing data through the N-BEATS architecture and global attention mechanism, decompose trends and seasonal features, and predict through feature fusion.
Accurate short-term trend prediction of multidimensional timing data is achieved, especially in power load prediction, which shows high efficiency and accuracy.
Smart Images
Figure CN116258262B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of multidimensional time series data prediction, and in particular to a short-term trend prediction method, system, storage medium and electronic device for multidimensional time series data. Background Art
[0002] A time series is a sequence of data points arranged in chronological order. Multidimensional time series data has relationships of mutual influence between factors, also known as factor correlation. Multidimensional time series data prediction technology uses historical data to construct a time series predictor to map the changing patterns of the data. The established predictor is then used to extrapolate to predict future data trends. A representative sequence is set as the target sequence (the sequence to be predicted), and all other sequences that change over time are called feature sequences. Multidimensional time series data is complex, making it very difficult to build a model that can fully exploit the relevant features of the data to achieve accurate predictions.
[0003] Traditional time series forecasting methods primarily rely on statistical models based on data autocorrelation. While these models have rigorous theoretical derivations, they lack sufficient parameters to map the nonlinear relationships between the multidimensional features of multidimensional time series data, making them difficult to apply in complex scenarios. Furthermore, these models primarily focus on exploring temporal correlations, neglecting the spatial distribution of data and the interplay between multiple factors, thus losing crucial information.
[0004] This shows that effectively capturing the correlation characteristics of time series data is key to learning the temporal variation patterns in historical data and achieving accurate short-term forecasts. At the same time, comprehensively capturing and learning the temporal dependencies within time series data has always been a challenging problem in forecasting research and has important theoretical research significance. Summary of the Invention
[0005] (1) Technical problems solved
[0006] In response to the deficiencies in the prior art, the present invention provides a method, system, storage medium, and electronic device for short-term trend prediction of multidimensional time series data, which solves the technical problem of inaccurate prediction of short-term multidimensional time series data.
[0007] (2) Technical solution
[0008] To achieve the above objectives, the present invention is implemented through the following technical solutions:
[0009] A short-term trend prediction method for multidimensional time series data is based on a time series feature encoder, an auxiliary information encoder, and a prediction decoder; the method comprises:
[0010] S1. Acquire multidimensional time series data, where the multidimensional time series data includes target sequence data and related characteristic factor data;
[0011] S2. Using the target sequence data as input of a temporal feature encoder, and obtaining a first feature based on an N-BEATS architecture;
[0012] S3. Using the feature factor data as input to an auxiliary information encoder, and obtaining a second feature based on a global attention mechanism;
[0013] S4. Fusion of the first feature and the second feature;
[0014] S5. Use the fusion result as the input of the prediction decoder to obtain the prediction result of the multi-dimensional time series data in the future target period.
[0015] Preferably, the overall structure of the temporal feature encoder is composed of the trend stack and seasonal stack of the N-BEATS model in sequence;
[0016] The S2 specifically includes:
[0017] S21, the target sequence data is used as input of the trend stack to obtain trend features;
[0018] S22, using the remaining target sequence data component after removing the trend feature as the input of the seasonal stack to obtain seasonal features;
[0019] S23. Obtain the first feature according to the trend feature and the seasonal feature.
[0020] Preferably, in said S21, the target sequence data is input into the first basic block of the first trend stack, and the partial trend features extracted by the lth block of the sth trend stack are As shown below:
[0021]
[0022] Wherein, the time vector t=[0,1,2,...,H-2,H-1] T / H is defined on a discrete grid from 0 to (H-1) / H, and each prediction is H time steps long; are the polynomial coefficients generated by the prediction of the fully connected network of the s-th stack and l blocks;
[0023] The output of the final s-th layer trend stack is
[0024] Preferably, in said S22, the remaining target sequence data component after removing the trend feature is input into the first basic block of the first layer seasonal trend stack, and the partial seasonal features extracted by the lth block of the kth layer seasonal stack are As shown below:
[0025]
[0026] in, are the Fourier coefficients produced by the fully connected network prediction of the lth block of the kth seasonal stack
[0027] The output of the final k-th seasonal stack is
[0028] Preferably, the first feature expression is obtained in S23 as follows:
[0029] Preferably, the process of obtaining the multi-dimensional time series data prediction result within the future target period in S5 is as follows:
[0030] o t =σ(W o ·[h t-1 ,z s ])
[0031] r t =σ(W r ·[h t-1 ,z s ])
[0032]
[0033]
[0034] Among them, z s represents the output of the feature fusion module, z s =Concat(f e ,f a ), f e 、f a Represent the first and second features respectively; predict the final hidden state h of the decoder t As the final prediction result That is, the power load forecast result in the future H time period.
[0035] A short-term trend prediction system for multi-dimensional time series data is based on a time series feature encoder, an auxiliary information encoder, and a prediction decoder; the system comprises:
[0036] An acquisition module, configured to acquire multidimensional time series data, wherein the multidimensional time series data includes target sequence data and related characteristic factor data;
[0037] A first encoding module, configured to use the target sequence data as input of a temporal feature encoder and obtain a first feature based on an N-BEATS architecture;
[0038] A second encoding module, configured to use the feature factor data as input to an auxiliary information encoder and obtain a second feature based on a global attention mechanism;
[0039] A fusion module, configured to fuse the first feature and the second feature;
[0040] The decoding module is used to use the fusion results as the input of the prediction decoder to obtain the prediction results of the multi-dimensional time series data in the future target period.
[0041] A storage medium stores a computer program for short-term trend prediction of multi-dimensional time series data, wherein the computer program enables a computer to execute the short-term trend prediction method as described above.
[0042] An electronic device, comprising:
[0043] One or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by the one or more processors, the programs including a method for executing the short-term trend forecasting method as described above (III) Beneficial Effects
[0044] The present invention provides a method, system, storage medium, and electronic device for short-term trend prediction of multidimensional time series data. Compared with the existing technology, it has the following advantages:
[0045] This paper proposes a new short-term multidimensional time series data prediction model framework. This architecture first employs an interpretable time series feature encoder to decompose the target sequence data into trend and seasonal features. An auxiliary information encoder is then used to encode the feature factor data into a latent information matrix, and a multi-head self-attention mechanism is employed to extract high-dimensional autocorrelation features. The model's time series feature encoder possesses powerful nonlinear modeling capabilities, meeting the requirements for target sequence feature extraction. Finally, a feature fusion module fuses the extracted features of different types, and a decoder capable of time series feature extraction is used to perform the final multidimensional time series data prediction. BRIEF DESCRIPTION OF THE DRAWINGS
[0046] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.
[0047] Figure 1 A framework diagram of a short-term load forecasting model provided by an embodiment of the present invention;
[0048] Figure 2 A block diagram of a short-term trend prediction method for multidimensional time series data provided by an embodiment of the present invention;
[0049] Figure 3 A schematic diagram of the basic block structure of an N-BEATS architecture provided in an embodiment of the present invention;
[0050] Figure 4 A schematic structural diagram of an auxiliary information encoder provided by an embodiment of the present invention;
[0051] Figure 5 A schematic diagram of feature fusion provided by an embodiment of the present invention;
[0052] Figure 6 A schematic diagram of the structure of a prediction decoder provided by an embodiment of the present invention. DETAILED DESCRIPTION
[0053] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention are clearly and completely described. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.
[0054] The embodiments of the present application solve the technical problem of inaccurate prediction of short-term multidimensional time series data by providing a short-term trend prediction method, system, storage medium and electronic device for multidimensional time series data.
[0055] The technical solution in the embodiments of the present application is to solve the above technical problems, and the overall idea is as follows:
[0056] like Figure 1As shown in the figure, the present invention proposes a new short-term multidimensional time series data prediction model framework; the structure first uses an interpretable time series feature encoder to decompose the trend and seasonal characteristics of the target sequence data. Then, an auxiliary information encoder is used to encode the feature factor data into a hidden information matrix, and a multi-head self-attention mechanism is used to obtain high-dimensional autocorrelation features. The time series feature encoder of this model has powerful nonlinear modeling capabilities and can meet the requirements of target sequence feature extraction. Finally, the extracted different types of features are fused through a feature fusion module, and the final multidimensional time series data prediction is performed through a decoder that can extract time series features.
[0057] In order to better understand the above technical solution, the above technical solution will be described in detail below with reference to the accompanying drawings and specific implementation methods.
[0058] Example:
[0059] like Figure 2 As shown, an embodiment of the present invention provides a short-term trend prediction method for multi-dimensional time series data, based on a time series feature encoder, an auxiliary information encoder and a prediction decoder; the method includes:
[0060] S1. Acquire multidimensional time series data, where the multidimensional time series data includes target sequence data and related characteristic factor data;
[0061] S2. Using the multidimensional time series data as input to a time series feature encoder, and obtaining a first feature based on an N-BEATS architecture;
[0062] S3. Using the feature factor data as input to an auxiliary information encoder, and obtaining a second feature based on a global attention mechanism;
[0063] S4. Fusion of the first feature and the second feature;
[0064] S5. Use the fusion result as the input of the prediction decoder to obtain the prediction result of the multi-dimensional time series data in the future target period.
[0065] The embodiment of the present invention realizes accurate prediction of short-term power load by providing the above solution.
[0066] Next, we will take the multi-dimensional time series data as an example to describe the various steps of the solution in detail:
[0067] In step S1 , multidimensional time series data is acquired, where the multidimensional time series data includes historical power load data and related characteristic factor data.
[0068] The historical power load is expressed as The related characteristic factors are expressed as a matrix
[0069] In step S2, the historical power load data is used as input of a time series feature encoder, and a first feature is obtained based on an N-BEATS architecture.
[0070] First, it is necessary to introduce the relevant content of the N-BEATS (NEURAL BASIS EXPANSION ANALYSIS FOR INTERPRETABLE TIME SERIES) model:
[0071] The basic building block of the N-BEATS architecture can be viewed as a fully connected nonlinear regressor, which can receive historical time series data of a certain length and learn the trend and seasonality features contained in the data.
[0072] The N-BEATS architecture connects multiple blocks in series. Each block learns a subset of features contained in the input historical time series data. The input to the next block is the remaining features after removing the features learned in the previous block. In other words, each block only learns features that have not been learned by the previous module. Finally, the learning results of each block are added together to obtain the final output. The following describes its core functions in detail:
[0073] like Figure 3 As shown, there are two branches in the basic block structure. Among them, the backward prediction branch is designed to reconstruct the input, and the forward prediction branch performs partial prediction operations. For the lth block, its input y l It is y l-1 The residual between the reconstructed output of the previous block is:
[0074] y l =y l-1 -Φ(y l-1 )
[0075] where Φ(·) represents the loss function of the previous block. The reconstruction branch removes redundant information for that block. Each block consists of two parts: the first part generates the dilation coefficients, and the second part maps the coefficients to the output through the base layer.
[0076] The first part is a fully connected neural network, which consists of 4 fully connected (FC) layers and 2 linear projection (LP) layers, which generate the forward expansion coefficient θ l f and the reverse expansion coefficient θ l b . Use h l,m Represents the output vector form of the mth layer. Then the operation of the fully connected layer (FC) is expressed as:
[0077] h l,m=ReLu(W l,m h l,m-1 +b l,m )
[0078] Where W l,m represents the weight matrix vector, b l,m ReLu(·) represents the activation function.
[0079] In particular, the input to the first layer is y l ,Right now:
[0080] h l,1 =ReLu(W l,1 y l +b l,1 )
[0081] Then h l,4 The output of the network layer is used as the input of two linear projection layers to obtain the expansion coefficient θ l b and θ l f ,Right now:
[0082] θ l b =W l b h l,4 ,θ l f =W l f h l,4
[0083] The second part contains two base layers, respectively using g l f and g l b They will θ l f and θ l b As input, we get the output and It can be expressed as a mathematical formula:
[0084]
[0085] in represents the output of the backward prediction branch, Represents some features extracted by this module.
[0086] The overall structure of the temporal feature encoder in the embodiment of the present invention is composed of the trend stack and seasonal stack of the N-BEATS model in sequence;
[0087] Correspondingly, the S2 specifically includes:
[0088] S21, using the historical power load data as input to a trend stack to obtain trend features; including:
[0089] Input the historical power load data into the first basic block of the first-layer trend stack, and then the partial trend features extracted by the lth block of the sth-layer trend stack are As shown below:
[0090]
[0091] Wherein, the time vector t=[0,1,2,...,H-2,H-1] T / H is defined on a discrete grid from 0 to (H-1) / H, and each prediction is H time steps long; are the polynomial coefficients generated by the prediction of the fully connected network of the s-th stack and l blocks;
[0092] The output of the final s-th layer trend stack is
[0093] S22. Using the remaining historical power load data components after removing the trend features as input to a seasonal stack to obtain seasonal features; including:
[0094] The remaining historical power load data components after removing the trend features are input into the first basic block of the first layer seasonal trend stack, and the partial seasonal features extracted by the lth block of the kth layer seasonal stack are As shown below:
[0095]
[0096] in, are the Fourier coefficients produced by the fully connected network prediction of the lth block of the kth seasonal stack
[0097] The output of the final k-th seasonal stack is
[0098] S23. Obtain the first feature according to the trend feature and the seasonal feature; expressed as:
[0099]
[0100] In step S3, the feature factor data is used as input of the auxiliary information encoder, and the second feature is obtained based on the global attention mechanism.
[0101] In this step, the auxiliary information encoder E a The specific structure of Figure 4As shown, it is used to encode additional feature factors into the hidden information matrix, with the aim of mining the autocorrelation between features through the global attention mechanism and obtaining relevant auxiliary information features f a =E a (x t ).
[0102] The encoding layer of the auxiliary information encoder primarily consists of a multi-head attention mechanism and a feedforward neural network layer. The multi-head attention mechanism has more powerful feature extraction capabilities than a single attention mechanism with the same number of parameters. The feedforward neural network here consists of two linear layers, activated using the ReLU function, primarily to enhance the model's nonlinear fitting capabilities.
[0103] In step S4, the first feature and the second feature are fused.
[0104] like Figure 5 As shown, this step extracts the feature f from the temporal feature encoder. e and the auxiliary information f extracted by the auxiliary information encoder a Fusion is performed to improve prediction accuracy. The specific operation can be expressed as follows:
[0105] z s =Concat(f e ,f a ).
[0106] In step S5, the fusion result is used as the input of the prediction decoder to obtain the power load prediction result in the future target period.
[0107] In order to analyze the feature vectors output by the feature fusion module and better learn the deep coupling relationship and time dependency between historical load data and auxiliary information, a prediction decoder D is constructed. p , to perform the prediction task within the specified future time period H; the process is as follows:
[0108] o t =σ(W o ·[h t-1 ,z s ])
[0109] r t =σ(W r ·[h t-1 ,z s ])
[0110]
[0111]
[0112] Among them, z srepresents the output of the feature fusion module, z s =Concat(f e ,f a ), f e 、f a Represent the first and second features respectively; predict the final hidden state h of the decoder t The final prediction result is That is, the power load forecast result in the future H time period.
[0113] In particular, the prediction decoder is an improved model based on the LSTM network with two gates, namely the update gate and the reset gate. It can effectively alleviate the gradient vanishing problem in RNN, and has a small number of parameters and a short training time. The basic structure of the decoder network is as follows Figure 6 As shown. Among them, × represents the scalar multiplication of the matrix, σ represents the Sigmoid activation function, tanh represents the hyperbolic tangent activation function, and I- represents that the data transmitted on this link is Io t .
[0114] An embodiment of the present invention provides a short-term trend prediction system for multi-dimensional time series data, based on a time series feature encoder, an auxiliary information encoder, and a prediction decoder; the system includes:
[0115] An acquisition module, configured to acquire multidimensional time series data, wherein the multidimensional time series data includes target sequence data and related characteristic factor data;
[0116] A first encoding module, configured to use the multidimensional time series data as input of a time series feature encoder and obtain a first feature based on an N-BEATS architecture;
[0117] A second encoding module, configured to use the feature factor data as input to an auxiliary information encoder and obtain a second feature based on a global attention mechanism;
[0118] A fusion module, configured to fuse the first feature and the second feature;
[0119] The decoding module is used to use the fusion results as the input of the prediction decoder to obtain the prediction results of the multi-dimensional time series data in the future target period.
[0120] An embodiment of the present invention provides a storage medium storing a computer program for short-term trend prediction of multi-dimensional time series data, wherein the computer program enables a computer to execute the short-term trend prediction method described above.
[0121] An embodiment of the present invention provides an electronic device, including:
[0122] one or more processors;
[0123] Memory; and
[0124] One or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including instructions for executing the short-term trend prediction method as described above.
[0125] It can be understood that the short-term trend prediction system, storage medium and electronic device for multidimensional time series data provided in the embodiments of the present invention correspond to the short-term trend prediction method for multidimensional time series data provided in the embodiments of the present invention. The explanation, examples and beneficial effects of the relevant contents can refer to the corresponding parts in the short-term trend prediction method for multidimensional time series data, and will not be repeated here.
[0126] In summary, compared with the existing technology, the present invention has the following beneficial effects:
[0127] 1. An embodiment of the present invention proposes a new short-term multidimensional time series data prediction model framework; this structure first uses an interpretable time series feature encoder to decompose the target sequence data into trend and seasonal features. Then, an auxiliary information encoder is used to encode the feature factor data into a hidden information matrix, and a multi-head self-attention mechanism is used to obtain high-dimensional autocorrelation features. The time series feature encoder of this model has powerful nonlinear modeling capabilities and can meet the requirements of target sequence feature extraction. Finally, the extracted features of different types are fused through a feature fusion module, and the final multidimensional time series data prediction is performed through a decoder that can extract time series features.
[0128] 2. When the multi-dimensional time series data is power data, the embodiment of the present invention can accurately predict short-term power load.
[0129] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply the existence of any such actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or device comprising the element.
[0130] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.
Claims
1. A short-term trend prediction method for multidimensional time series data, characterized in that: Based on temporal feature encoder, auxiliary information encoder and prediction decoder; The short-term trend prediction method of multi-dimensional time series data is used for short-term power load prediction, and the method comprises: S1. Acquire multidimensional time series data, where the multidimensional time series data includes target sequence data and related characteristic factor data; S2. Using the target sequence data as input to a time series feature encoder, and obtaining a first feature based on an N-BEATS architecture; wherein the overall structure of the time series feature encoder is sequentially composed of a trend stack and a seasonal stack of an N-BEATS model; S2 specifically includes: S21, the target sequence data is used as input of the trend stack to obtain trend features; S22, using the remaining target sequence data component after removing the trend feature as the input of the seasonal stack to obtain seasonal features; S23. Obtain the first feature according to the trend feature and the seasonal feature; S3. Using the feature factor data as input to an auxiliary information encoder, and obtaining a second feature based on a global attention mechanism; S4. Fusion of the first feature and the second feature; S5. Use the fusion result as the input of the prediction decoder to obtain the prediction result of the multi-dimensional time series data in the future target period; Wherein, in said S21, the target sequence data is input into the first basic block of the first layer trend stack, and the partial trend features extracted by the lth block of the sth layer trend stack are As shown below: Wherein, the time vector t=[0,1,2,...,H-2,H-1] T / H is defined on a discrete grid from 0 to (H-1) / H, and each prediction is H time steps long; are the polynomial coefficients generated by the prediction of the fully connected network of the s-th stack and l blocks; The output of the final s-th layer trend stack is In the step S22, the remaining target sequence data components after removing the trend features are input into the first basic block of the first seasonal trend stack. Then, the seasonal features extracted from the lth block of the kth seasonal stack are As shown below: in, are the Fourier coefficients produced by the fully connected network prediction of the lth block of the kth seasonal stack; The output of the final k-th seasonal stack is 2. The short-term trend prediction method for multidimensional time series data according to claim 1, characterized in that: The first feature expression obtained in S23 is:
3. The short-term trend prediction method for multidimensional time series data according to claim 1, characterized in that: In S5, the multi-dimensional time series data prediction results within the future target period are obtained, and the process is as follows: about t =σ(W o ·[h t-1 ,With s ]) r t =σ(W r ·[h t-1 ,z s ]) Among them, z s represents the output of the feature fusion module, z s =Concat(f e ,f a ), f e 、f a Represent the first and second features respectively; predict the final hidden state h of the decoder t As the final prediction result That is, the power load forecast result in the future H time period.
4. A short-term trend prediction system for multidimensional time series data, characterized by: Based on temporal feature encoder, auxiliary information encoder and prediction decoder; The short-term trend prediction system of multi-dimensional time series data is used for short-term power load prediction, and the system includes: An acquisition module, configured to acquire multidimensional time series data, wherein the multidimensional time series data includes target sequence data and related characteristic factor data; a first encoding module, configured to use the target sequence data as input to a time series feature encoder and obtain a first feature based on an N-BEATS architecture; wherein the overall structure of the time series feature encoder is sequentially composed of a trend stack and a seasonal stack of the N-BEATS model; The first encoding module is specifically configured to perform the following steps: S21, the target sequence data is used as input of the trend stack to obtain trend features; S22, using the remaining target sequence data component after removing the trend feature as the input of the seasonal stack to obtain seasonal features; S23. Obtain the first feature according to the trend feature and the seasonal feature; A second encoding module, configured to use the feature factor data as input to an auxiliary information encoder and obtain a second feature based on a global attention mechanism; A fusion module, configured to fuse the first feature and the second feature; The decoding module is used to use the fusion results as the input of the prediction decoder to obtain the prediction results of the multi-dimensional time series data in the future target period; Wherein, in said S21, the target sequence data is input into the first basic block of the first layer trend stack, and the partial trend features extracted by the lth block of the sth layer trend stack are As shown below: Wherein, the time vector t=[0,1,2,...,H-2,H-1] T / H is defined on a discrete grid from 0 to (H-1) / H, and each prediction is H time steps long; are the polynomial coefficients generated by the prediction of the fully connected network of the s-th stack and l blocks; The output of the final s-th layer trend stack is In the step S22, the remaining target sequence data components after removing the trend features are input into the first basic block of the first seasonal trend stack. Then, the seasonal features extracted from the lth block of the kth seasonal stack are As shown below: in, are the Fourier coefficients produced by the fully connected network prediction of the lth block of the kth seasonal stack; The output of the final k-th seasonal stack is 5. A storage medium, characterized in that It stores a computer program for short-term trend prediction of multi-dimensional time series data, wherein the computer program enables a computer to execute the short-term trend prediction method according to any one of claims 1 to 3.
6. An electronic device, characterized in that: include: one or more processors; Memory; as well as One or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including instructions for executing the short-term trend forecasting method according to any one of claims 1 to 3.