Power data time sequence interpolation method and system based on improved TimesNet

Abstract

The invention relates to an improved TimesNet-based power data time sequence interpolation method and system. The method comprises the following steps of performing missing value identification, anomaly detection and standardized preprocessing on power data, and extracting time sequence features; according to the method, an improved TimesNet model is constructed, a 1D time sequence is converted into a 2D tensor, and features are extracted through a Transformer-CNN hybrid model: an axial attention branch establishes periodic internal and external dependencies in row and column directions, a multi-scale convolution branch captures local features, and a gating mechanism fuses output of the axial attention branch and the multi-scale convolution branch. A deformable convolution kernel is adopted to dynamically adjust a receptive field to adapt to a multi-scale time mode. And in training, time sequence enhancement and GAN are combined to generate data, so that the generalization of the model is improved. The method solves the problem that the traditional interpolation technology cannot effectively model the complex time sequence dependence of the electric power data, remarkably improves the missing value interpolation precision, and provides reliable data support for dispatching and prediction of an electric power system.

Description

Technical Field

[0001] This application relates to the technical field of power data interpolation, and more specifically, to a method and system for time series interpolation of power data based on an improved TimesNet. Background Art

[0002] Power data sets usually contain a large amount of time series data, which record the changes of various parameters in the power system, such as power load, voltage, frequency, and electricity price. Missing data not only affects the accuracy of data analysis but may also lead to incorrect decisions, especially in scenarios that require high-precision prediction and real-time scheduling. Therefore, how to effectively fill in the missing data has become an important issue in power system data analysis. Traditional time series interpolation methods (such as mean interpolation, linear interpolation, forward and backward value filling, etc.) often fail to effectively capture the complex time dependencies and periodic characteristics in power data. With the development of deep learning and machine learning technologies, models based on deep neural networks have gradually been applied to time series interpolation, especially in the field of power data with strong time series characteristics. However, existing methods still have certain limitations and cannot fully utilize the periodic and seasonal characteristics of power data and the complex dependencies between multiple variables. Therefore, the present invention proposes a method for time series interpolation of power data based on the Transformer and its variant models, aiming to provide an efficient and accurate interpolation algorithm by combining the characteristics of power data, so as to improve the data support capabilities for tasks such as intelligent scheduling, load prediction, and fault diagnosis in the power system.

[0003] The prior art, such as the Chinese patent application with the publication number "CN115984281A", discloses a multi-task completion method for time series sea surface temperature images based on local specificity deepening, including: extracting local specific information at each moment through a local information extraction network, then extracting global information through a global information extraction network, and finally fusing the global information and local specific information through an image completion network to obtain fused information, and outputting the completed image after being processed by an image decoder; this invention retains the local specificity at each moment, avoids the local information being masked by the global information during the fusion process, and improves the quality of the image completion result.

[0004] The problems existing in the above prior art are that this method uses GRU / LSTM units to fuse global and local information, does not introduce deformable convolution or axial attention, and is difficult to dynamically adapt to dependencies at different time scales; it relies on weekly means as global information and does not explicitly model long-term periodic characteristics, resulting in insufficient capture of complex periodic patterns. Summary of the Invention

[0005] To solve the above technical problems, the present invention proposes a method and system for time series interpolation of power data based on an improved TimesNet.

[0006] The technical solution of the present invention is as follows:

[0007] The present invention proposes a time series interpolation method for power data based on improved TimesNet, including the following steps:

[0008] Collect power data with missing values, and preprocess the collected power data. The preprocessing includes missing value identification, outlier detection, and normalization; extract time series features from the preprocessed power data;

[0009] Construct an improved TimesNet model, use the time series features of the power data as the input of the improved TimesNet model, and predict and interpolate the missing values in the power data;

[0010] After the improved TimesNet model converts the 1D time series into a 2D tensor, it uses a Transformer-CNN hybrid model containing an axial attention mechanism to extract local features and periodic information in the time series; uses deformable convolutional kernels to dynamically adjust the receptive field and capture multi-scale time pattern features.

[0011] As a preferred implementation, during the training process of the improved TimesNet model, the diversity and data volume of the power data are enhanced through data augmentation techniques; the data augmentation techniques include: time series augmentation and generating power data using a generative adversarial network.

[0012] As a preferred implementation, the Transformer-CNN hybrid model includes:

[0013] A 2D convolution branch and an axial attention branch arranged in parallel. Among them, the 2D convolution branch uses a multi-scale convolutional kernel of the Inception structure to perform convolution operations on the input 2D tensor to obtain a multi-scale feature map; the axial attention branch performs sparse attention calculations on the input 2D tensor in both the row and column directions to obtain attention-weighted features; the attention-weighted features and the multi-scale feature map are fused through a gating mechanism to obtain a gated fusion feature.

[0014] As a preferred implementation, the axial attention branch performs sparse attention calculations on the input 2D tensor in both the row and column directions, where:

[0015] The row direction attention calculation is specifically: calculate the self-attention for each row of the 2D tensor separately to establish the correlation relationship between the same phase points of different periods;

[0016] The column direction attention calculation is specifically: calculate the self-attention for each column of the 2D tensor separately to establish the dependence relationship between different time points of the same period.

[0017] As a preferred embodiment, the deformable convolution kernel is used to dynamically adjust the receptive field, specifically:

[0018] Based on the local temporal features of the input data, dynamic offset parameters of the convolution kernel sampling points are generated through a prediction network;

[0019] The dynamic offset parameters are superimposed on the sampling coordinates of the standard convolution kernel to achieve adaptive adjustment of the spatial deformation parameters;

[0020] Among them, the spatial deformation parameters include the dynamic offset parameters and are used to control the deformation trajectory of the convolution kernel in the multi-dimensional temporal space.

[0021] On the other hand, the present invention also provides a power data temporal interpolation system based on the improved TimesNet, including:

[0022] A data collection and processing module that collects power data with missing values, preprocesses the collected power data, and the preprocessing includes missing value identification, outlier detection, and normalization processing; extracts time series features from the preprocessed power data;

[0023] A model construction and interpolation module that constructs an improved TimesNet model, uses the time series features of the power data as the input of the improved TimesNet model, and predicts and interpolates the missing values in the power data;

[0024] After the improved TimesNet model converts the 1D time series into a 2D tensor, a Transformer-CNN hybrid model containing an axial attention mechanism is used to extract local features and periodic information in the time series; a deformable convolution kernel is used to dynamically adjust the receptive field to capture multi-scale time pattern features.

[0025] As a preferred embodiment, during the training process of the improved TimesNet model, the diversity and data volume of the power data are enhanced through data augmentation techniques; the data augmentation techniques include: time series augmentation and generating power data using a generative adversarial network.

[0026] As a preferred embodiment, in the model construction and interpolation module, the Transformer-CNN hybrid model includes:

[0027] A 2D convolutional branch and an axial attention branch are arranged in parallel. The 2D convolutional branch uses a multi-scale convolutional kernel of the Inception structure to perform a convolution operation on the input 2D tensor to obtain a multi-scale feature map. The axial attention branch performs sparse attention calculations on the input 2D tensor in both the row and column directions to obtain an attention-weighted feature. The attention-weighted feature and the multi-scale feature map are fused through a gating mechanism to obtain a gated fusion feature.

[0028] On the other hand, the present invention also provides an electronic device, on which a computer program is stored. When the computer program is executed by a processor, it implements a method for interpolating power data time series based on an improved TimesNet as described in any embodiment of the present invention.

[0029] On the other hand, the present invention also provides a computer-readable medium for storing one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors are caused to implement a method for interpolating power data time series based on an improved TimesNet as described in any embodiment of the present invention.

[0030] The present invention has the following beneficial effects:

[0031] Multi-scale feature fusion: By axial attention and deformable convolution, short-term local features and long-term periodic dependencies are captured simultaneously;

[0032] Dynamic adaptability: The deformable convolution kernel adjusts the receptive field according to the input data, improving the modeling ability for complex time series patterns;

[0033] Noise and missing robustness: Combining data augmentation and adversarial training to enhance the stability of the model in data missing or noisy scenarios;

[0034] High-precision interpolation: The gating mechanism fuses global periodicity and local details, significantly improving the interpolation accuracy of power data. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings required to be used in the embodiments of the present application. It should be understood that the following drawings only show some embodiments of the present application, and therefore should not be regarded as limiting the scope. For those of ordinary skill in the art, other related drawings can be obtained based on these drawings without creative efforts.

[0036] Figure 1 It is a schematic flow chart of the method of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0037] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0038] It should be understood that the step numbers used in the text are only for convenient description and do not limit the execution order of the steps.

[0039] It should be understood that the terms used in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. As used in the specification of the present invention and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include the plural forms.

[0040] The terms "comprising" and "including" indicate the presence of the described features, wholes, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or their combinations.

[0041] The term "and / or" refers to any combination and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0042] Embodiment 1:

[0043] To make the purpose, technical solutions and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with specific embodiments of the present application and with reference to the attached Figure 1 , to clearly and completely describe the technical solutions of the present invention.

[0044] To solve the problems of the prior art, the present invention provides a power data time series interpolation method based on improved TimesNet, including the following steps:

[0045] Collect power data with missing values, preprocess the collected power data, and the preprocessing includes missing value identification, outlier detection and normalization processing; extract time series features from the preprocessed power data;

[0046] The data preprocessing step includes:

[0047] Missing value identification: In the power dataset, missing values may occur due to sensor failures, communication delays, etc. These missing values need to be identified in the data preprocessing stage.

[0048] Outlier Detection: There may be outliers in power data, which may be caused by equipment failures or data acquisition errors. Through statistical analysis and machine learning methods (such as Isolation Forest, Z-score, etc.), these outliers can be identified and processed to avoid affecting subsequent analysis.

[0049] Data Standardization Processing: It includes data standardization and normalization processing. Among them, standardization processing: converts the data into a distribution with a mean of 0 and a variance of 1, which is suitable for most machine learning algorithms, especially distance-based algorithms (such as KNN). Normalization processing: scales the data to a specific range (such as [0,1]), which helps to accelerate the convergence speed of the model, especially when using neural networks.

[0050] Time Series Feature Extraction:

[0051] Power data usually has strong time dependence. Therefore, in the preprocessing stage, relevant features of the time series need to be extracted:

[0052] Timestamp Processing: Convert the timestamp into features that can be used by the model, such as extracting information about year, month, day, hour, day of the week, etc., to capture seasonal and periodic changes.

[0053] Sliding Window: Use the sliding window technique to generate input features of the time series, which helps the model learn the trends and periodic changes in the time series.

[0054] During the training process of the improved TimesNet model, due to the limitation of the number of samples, the diversity and quantity of power data can be enhanced through data augmentation techniques; data augmentation techniques include: time series augmentation and generating power data using generative adversarial networks. Among them:

[0055] Time Series Augmentation: Generate new training samples by performing operations such as translation, scaling, and adding noise to the time series data to increase data diversity.

[0056] Generating Power Data using Generative Adversarial Networks: Use generative adversarial network (GANs) technology to generate synthetic data samples to supplement the shortage of real data.

[0057] Construct an improved TimesNet model, use the time series features of power data as the input of the improved TimesNet model, and predict and impute the missing values in the power data;

[0058] After converting the 1D time series into a 2D tensor, the improved TimesNet model uses a Transformer-CNN hybrid model with an axial attention mechanism to extract local features and periodic information in the time series; a deformable convolution kernel is used to dynamically adjust the receptive field to capture multi-scale time pattern features.

[0059] The Transformer-CNN hybrid model includes:

[0060] A 2D convolution branch and an axial attention branch arranged in parallel. The 2D convolution branch uses a multi-scale convolution kernel of the Inception structure to perform convolution operations on the input 2D tensor to obtain a multi-scale feature map; the axial attention branch performs sparse attention calculations on the input 2D tensor in both the row and column directions to obtain attention-weighted features; the attention-weighted features and the multi-scale feature map are fused through a gating mechanism to obtain gating fusion features.

[0061] The axial attention branch performs sparse attention calculations on the input 2D tensor in both the row and column directions, where:

[0062] The row-direction attention calculation is specifically: calculate the self-attention for each row of the 2D tensor separately to establish the correlation relationship between the same-phase points of different periods;

[0063] The column-direction attention calculation is specifically: calculate the self-attention for each column of the 2D tensor separately to establish the dependence relationship between different time points in the same period.

[0064] Compared with the traditional Inception module that only uses CNN for feature extraction and outputs a 2D tensor, the improved Inception module uses Transformer-CNN to extract local features, improving the ability to capture short-term patterns. Axial attention (Column-wise Axial Attention) is introduced in the column direction (within the period), and different time points within the same period can influence each other, enhancing the modeling ability of short-term time patterns. Axial attention (Row-wise Axial Attention) is introduced in the row direction (between periods), and time points of different periods but the same phase can influence each other, enhancing the modeling ability of long-term dependencies. After the axial attention between periods and within the period, a 2D tensor is output. This mechanism enables the improved TimesNet model to effectively extract local features and periodic information in the time series.

[0065] The use of a deformable convolution kernel to dynamically adjust the receptive field is specifically:

[0066] Based on the local temporal features of the input data, dynamic offset parameters of the convolution kernel sampling points are generated through a prediction network; the specific calculation formula is:

[0067] Δp = f θ (x local );

[0068] In the formula: Δp is the dynamic offset, representing the position adjustment amount of each sampling point of the convolution kernel; x local is the local temporal feature; f θ is a prediction network with parameter θ (such as a small neural network), which generates an offset according to the local temporal feature;

[0069] Superimpose the dynamic offset parameter on the sampling coordinates of the standard convolution kernel to achieve adaptive adjustment of the spatial deformation parameter; the specific calculation formula is:

[0070] T(p) = p + Δp(x local );

[0071] In the formula: T(p) is the spatial deformation parameter, representing the adjusted sampling point coordinates of the convolution kernel; p is the original sampling coordinate of the standard convolution kernel;

[0072] Among them, the spatial deformation parameter includes the dynamic offset parameter, which is used to control the deformation trajectory of the convolution kernel in the multi-dimensional temporal space.

[0073] After being processed by the Transformer-CNN hybrid model, the obtained 2D representation will be converted back to the 1D format for subsequent classification or regression tasks. This process compresses the multi-scale information back to one dimension to ensure that the improved TimesNet model can retain important temporal features in the final prediction output.

[0074] In the missing value imputation task of power data, indicators such as the mean square error (MSE) and the mean absolute error (MAE) can be used to systematically evaluate the imputation results to verify the effectiveness of the imputation method used. The following are the detailed implementation steps:

[0075] Definition of evaluation indicators: The mean square error MSE is the average of the squares of the differences between the actual value and the predicted value, and the formula is:

[0076]

[0077] where y i is the actual value, is the predicted value, and n is the number of samples. MSE emphasizes the penalty for large errors, so it can effectively reflect the accuracy of the model in the imputation task when evaluating the model performance.

[0078] The mean absolute error MAE is the average of the absolute differences between the actual value and the predicted value, and the formula is:

[0079]

[0080] MAE provides the average error between the predicted result and the actual result, which is suitable for evaluating the overall performance of the model.

[0081] Example Two:

[0082] This example provides a power data time series interpolation system based on the improved TimesNet, including:

[0083] A data collection and processing module that collects power data with missing values, preprocesses the collected power data, and the preprocessing includes missing value identification, outlier detection, and normalization processing; extracts time series features from the preprocessed power data;

[0084] A model construction and interpolation module that constructs an improved TimesNet model, takes the time series features of the power data as the input of the improved TimesNet model, and predicts and interpolates the missing values in the power data;

[0085] After converting the 1D time series into a 2D tensor, the improved TimesNet model uses a Transformer-CNN hybrid model containing an axial attention mechanism to extract local features and periodic information in the time series; uses a deformable convolutional kernel to dynamically adjust the receptive field and capture multi-scale time pattern features.

[0086] Example Three:

[0087] This example provides an electronic device with a computer program stored thereon, and when the computer program is executed by a processor, it implements a power data time series interpolation method based on the improved TimesNet as described in any embodiment of the present invention.

[0088] Example Four:

[0089] This example provides a computer-readable medium for storing one or more programs, and when the one or more programs are executed by the one or more processors, the one or more processors are caused to implement a power data time series interpolation method based on the improved TimesNet as described in any embodiment of the present invention.

[0090] In the embodiments of the present application, "at least one" means one or more, and "a plurality" means two or more. "And / or" describes the association relationship of associated objects, indicating that three relationships can exist. For example, A and / or B can represent the situations of A existing alone, A and B existing simultaneously, and B existing alone. Where A and B can be singular or plural. The character " / " generally represents an "or" relationship between the associated objects before and after. "At least one of the following" and its similar expressions refer to any combination of these items, including any combination of single items or plural items. For example, at least one of a, b, and c can represent: a, b, c, a and b, a and c, b and c, or a and b and c, where a, b, and c can be single or plural.

[0091] Those of ordinary skill in the art can realize that the various units and algorithm steps described in the embodiments disclosed herein can be implemented by a combination of electronic hardware, computer software, and electronic hardware. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present application.

[0092] Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working processes of the systems, devices, and units described above can refer to the corresponding processes in the foregoing method embodiments and will not be elaborated herein.

[0093] In several embodiments provided by the present application, if any function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, in essence, or the part that contributes to the prior art or a part of this 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 for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (ROM), random access memories (RAM), magnetic disks, or optical discs that can store program codes.

[0094] The above are only the embodiments of the present invention and do not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the content of the specification and drawings of the present invention, or directly or indirectly applied in other related technical fields, shall be equally included in the patent protection scope of the present invention.

Claims

1. An improved TimesNet-based power data time series interpolation method, characterized in that, It includes the following steps: Collect power data with missing values, and preprocess the collected power data. The preprocessing includes missing value identification, outlier detection, and normalization processing; Extract time series features from the preprocessed power data; Construct an improved TimesNet model, use the time series features of the power data as the input of the improved TimesNet model, and predict and impute the missing values in the power data; After converting the 1D time series into a 2D tensor, the improved TimesNet model uses a Transformer-CNN hybrid model containing an axial attention mechanism to extract local features and periodic information in the time series; a deformable convolution kernel is used to dynamically adjust the receptive field to capture multi-scale time pattern features.

2. The power data time series interpolation method based on the improved TimesNet according to claim 1, wherein: During the training process of the improved TimesNet model, enhance the diversity and data volume of the power data through data augmentation techniques; The data augmentation techniques include: time series augmentation and generating power data using a generative adversarial network.

3. A method for interpolating power data time series based on improved TimesNet according to claim 1, characterized in that: The Transformer-CNN hybrid model includes: A 2D convolution branch and an axial attention branch arranged in parallel. The 2D convolution branch uses a multi-scale convolution kernel of the Inception structure to perform a convolution operation on the input 2D tensor to obtain a multi-scale feature map; the axial attention branch performs sparse attention calculations on the input 2D tensor in both the row and column directions to obtain attention-weighted features; the attention-weighted features and the multi-scale feature map are fused through a gating mechanism to obtain gating fusion features.

4. An interpolation method for power data time series based on improved TimesNet according to claim 3, characterized in that: The axial attention branch performs sparse attention calculations on the input 2D tensor in both the row and column directions, where: The row direction attention calculation is specifically: calculate the self-attention for each row of the 2D tensor separately to establish the correlation relationship between the same phase points of different periods; The column direction attention calculation is specifically: calculate the self-attention for each column of the 2D tensor separately to establish the dependence relationship between different time points of the same period.

5. A method for interpolating time series of power data based on improved TimesNet according to claim 1, characterized in that: The use of a deformable convolution kernel to dynamically adjust the receptive field is specifically: Based on the local temporal features of the input data, generate dynamic offset parameters for the sampling points of the convolution kernel through a prediction network; Superimpose the dynamic offset parameters on the sampling coordinates of the standard convolution kernel to achieve adaptive adjustment of the spatial deformation parameters; Among them, the spatial deformation parameters include the dynamic offset parameters and are used to control the deformation trajectory of the convolution kernel in the multi-dimensional temporal space.

6. A power data time series interpolation system based on improved TimesNet, characterized in that, It includes: A data collection and processing module that collects power data with missing values and preprocesses the collected power data. The preprocessing includes missing value identification, outlier detection, and normalization processing; Extract time series features from the preprocessed power data; A model construction and imputation module that constructs an improved TimesNet model, uses the time series features of the power data as the input of the improved TimesNet model, and predicts and imputes the missing values in the power data; After converting the 1D time series into a 2D tensor, the improved TimesNet model uses a Transformer-CNN hybrid model with an axial attention mechanism to extract local features and periodic information in the time series; a deformable convolutional kernel is used to dynamically adjust the receptive field to capture multi-scale time pattern features.

7. An improved TimesNet-based power data time series interpolation system according to claim 6, characterized in that: During the training process of the improved TimesNet model, data augmentation techniques are used to enhance the diversity and volume of power data; The data augmentation techniques include: time series augmentation and generating power data using a generative adversarial network.

8. The power data time series interpolation system based on the improved TimesNet according to claim 6, characterized in that: In the model construction and imputation module, the Transformer-CNN hybrid model includes: A 2D convolution branch and an axial attention branch arranged in parallel, where the 2D convolution branch performs convolution operations on the input 2D tensor using a multi-scale convolutional kernel with an Inception structure to obtain a multi-scale feature map; the axial attention branch performs sparse attention calculations on the input 2D tensor in both the row and column directions to obtain attention-weighted features; the attention-weighted features and the multi-scale feature map are fused through a gating mechanism to obtain gated fusion features.

9. An electronic device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the program, it implements a method for time series imputation of power data based on the improved TimesNet as described in any one of claims 1 to 5.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements a method for time series imputation of power data based on the improved TimesNet as described in any one of claims 1 to 5.

Images