Self-adaptive pooling method for dynamic sensing of electric power data characteristics

Abstract

The invention belongs to the technical field of data processing, and particularly relates to a self-adaptive pooling method for dynamic sensing of power data features. According to the method, multi-source historical and real-time power data are collected, a self-adaptive pooling model is constructed after preprocessing, and the model comprises an input layer, a dynamic sensing layer, a pooling operation layer and an output layer. Wherein the dynamic sensing layer can dynamically adjust the size of a window according to data features, and the pooling operation layer adopts a multi-dimensional constraint pooling strategy including window feature mapping, physical constraint pooling and feature sensitive fusion. And after the model is trained, real-time power data are detected, and a detection result is compared with a normal threshold value of a power grid to evaluate an operation state. Compared with a traditional method, the power data can be accurately processed, the operation state of the power grid can be comprehensively and accurately reflected, the defects of the traditional technology in the aspects of data pooling, noise processing and comprehensive analysis are effectively overcome, and stable and efficient operation of the power grid is guaranteed.

Description

Technical Field

[0001] The present invention belongs to the technical field of data processing, and in particular relates to an adaptive pooling method for dynamic perception of power data characteristics. Background Art

[0002] In today's booming power industry, power data processing plays a key role in ensuring the stable operation of the power grid and improving the management efficiency of the power system. With the continuous expansion of the scale of the power grid and the improvement of its intelligence, power data presents the characteristics of massive, complex and dynamic changes. Multi-source power data not only contains rich information, but also has different time scales and characteristic laws. However, the existing power data processing technology has many defects in dealing with these complex data. On the one hand, traditional data pooling methods often use fixed pooling windows, which cannot be flexibly adjusted according to the dynamic characteristics of power data, and are prone to losing key information or introducing redundant data, making it difficult to accurately capture the changing trend of data. On the other hand, in the data preprocessing link, conventional methods have poor processing effects on noise. While removing noise, they may destroy the original characteristic structure of the data, affecting the accuracy of subsequent analysis. In addition, the existing methods are insufficient in comprehensively considering the physical characteristics of the power system and integrating the characteristics of multi-source data, and cannot fully and accurately reflect the actual operating status of the power grid. Summary of the invention

[0003] In view of the technical problems existing in the above-mentioned background technology, the present invention proposes an adaptive pooling method for dynamically sensing power data characteristics.

[0004] In order to achieve the above object, the technical solution adopted by the present invention comprises the following steps:

[0005] S1. Data acquisition: Collect multi-source historical power data and multi-source real-time power data; the multi-source historical power data is recorded as first power data, and the multi-source real-time power data is recorded as second power data;

[0006] S2, data preprocessing: preprocessing the collected first power data to obtain third power data;

[0007] S3, model construction: build an adaptive pooling model;

[0008] The adaptive pooling model includes a data input layer, a dynamic perception layer, a pooling operation layer and an output layer;

[0009] Input layer: receives pre-processed power data;

[0010] Dynamic perception layer: dynamically adjusts the perception window size according to the characteristics of the input data;

[0011] Pooling operation layer: Based on the window size output by the dynamic perception layer, a multi-dimensional constrained pooling strategy is adopted;

[0012] Output layer: output the result after pooling;

[0013] S4, model training: inputting the third power data into the adaptive pooling model for training, and obtaining the trained model by continuously adjusting the model parameters;

[0014] S5, real-time detection: pre-processing the second power data and inputting it into the trained adaptive pooling model to obtain a detection result;

[0015] S6, analysis and evaluation: compare the output results with the normal threshold of the grid data characteristics to evaluate the operating status of the current power data;

[0016] The pooling operation layer adopts a multi-dimensional constrained pooling strategy including window feature mapping, physical constraint pooling and feature sensitive fusion;

[0017] Window feature mapping: maps the window size determined by the dynamic perception layer to the pooling dimension parameter;

[0018] Physical constraint pooling: Calculate the constraint violation degree for the power data in each window, set the violation degree threshold, retain the data points with a value greater than the violation degree threshold, and perform average pooling on the remaining points;

[0019] Frequency-sensitive fusion: Perform short-time Fourier transform on the pooled data, extract the frequency components, and calculate the weighted sum of the frequency weights to obtain the final features.

[0020] Preferably, the multi-source historical power data and multi-source real-time power data include power generation data, power transmission data, power distribution data, power consumption data, and power trading data.

[0021] Preferably, the specific implementation of obtaining the third power data is:

[0022] S21, the first power data Perform multi-scale decomposition and improve the wavelet packet basis function to meet the following requirements: , where j is the scale factor, k is the translation factor, and m is the branch index of the decomposition. , is the traditional basis function, is the improved basis function;

[0023] S22, after wavelet packet transformation, the coefficients of different scales and branches are obtained : ;

[0024] S23, perform soft threshold processing on the wavelet packet data to obtain the processed coefficients , calculated as: ;

[0025] S24, finally use the same improved wavelet packet basis function as in decomposition to process the wavelet packet coefficients Reconstruct and obtain the third power data , calculated as: .

[0026] Preferably, the soft threshold in step S23 The calculation method is: , where n is the number of wavelet packet coefficients of this scale and branch, is the noise estimate for this scale and branch.

[0027] Preferably, the input layer and the dynamic perception layer of the model constructed in step S3 need to undergo a special extraction operation, including data union, feature screening, feature weighting and feature fusion;

[0028] Joint feature extraction: For the pre-processed power data in the input layer, convolutional neural network and gated recurrent unit are used to extract spatial features and time series features respectively;

[0029] Feature screening: Calculate the mutual information between features, construct a mutual information matrix, and set a mutual information threshold. When the correlation coefficient of two features exceeds the threshold, retain the feature that has a more critical impact on the operating status of the power system and remove redundant features with too high correlation.

[0030] Feature weighting: Use the attention mechanism to assign weights to the selected features. The weight assignment is dynamically adjusted according to different operating scenarios and the importance of data features.

[0031] Feature fusion: Use feature splicing to fuse different types of features to form a comprehensive feature vector, and then perform dimensionality reduction to retain key data information.

[0032] Preferably, in the dynamic perception layer, the specific steps of dynamically adjusting the perception window size according to the characteristics of the input data include data feature quantification, fluctuation level division, window mapping size and dynamic adjustment;

[0033] Data feature quantification: Quantitatively analyze the comprehensive feature vector obtained through feature extraction, calculate the variance of the statistical index of each feature, and evaluate the degree of data fluctuation;

[0034] Volatility level classification: set three levels of low volatility, medium volatility and high volatility, corresponding to different volatility ranges, set thresholds and classify according to the degree of data volatility;

[0035] Window size mapping: pre-set the corresponding perception window size for each fluctuation level;

[0036] Dynamic adjustment: When the volatility level of the data changes, the perception window size is adjusted accordingly.

[0037] Preferably, the violation degree calculation of the physical constraint pooling includes reconstruction error, calculation residual, and calculation violation degree;

[0038] Reconstruction error: Build an autoencoder model and calculate the Euclidean distance between the original data and the reconstructed data as the reconstruction error;

[0039] Calculate the residual: Based on the power node admittance matrix, calculate the absolute difference between the actual value and the theoretical value of the active power and reactive power as the residual;

[0040] Calculate the violation degree: Based on the weighted fusion of the reconstruction error and the calculation residual, a violation degree indicator is formed.

[0041] Compared with the prior art, the advantages and positive effects of the present invention are that the dynamic perception layer dynamically adjusts the perception window size according to the data characteristics, divides the levels by quantifying the data fluctuations, matches different windows, flexibly adapts to data changes, and avoids information loss or redundancy. The multi-dimensional constrained pooling strategy of the pooling operation layer combines window feature mapping, physical constraint pooling and feature sensitive fusion, comprehensively considers physical characteristics and frequency characteristics, and improves data processing accuracy. The improved wavelet packet basis function is used to pre-process the data, and a variety of algorithms are combined for feature extraction and fusion, which effectively suppresses noise, retains key information, and accurately reflects the operation status of the power grid. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0043] Figure 1 It is a structural flow chart of the whole invention; Figure 2 It is the sub-flow chart of pre-processing; Figure 3 A sub-process diagram constructed for the model. DETAILED DESCRIPTION

[0044] In order to more clearly understand the above-mentioned purpose, features and advantages of the present invention, the present invention is further described below in conjunction with the accompanying drawings and embodiments. It should be noted that the embodiments of the present application and the features in the embodiments can be combined with each other without conflict.

[0045] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways than those described herein. Therefore, the present invention is not limited to the specific embodiments of the following disclosure.

[0046] Embodiment: In modern power systems, the scale of power grids continues to expand, the degree of intelligence continues to improve, the amount of power data is exploding, and is characterized by multi-source, complexity, and dynamic changes. Traditional data processing technology is difficult to adapt to such changes, and has deficiencies in data pooling, noise processing, and comprehensive reflection of power grid operation status, and cannot meet the needs of stable power grid operation and efficient management. In order to achieve the effect of accurately processing power data, accurately assessing the power grid operation status, ensuring the stable and efficient operation of the power grid, and solving the defects of traditional power data processing technology, the present invention adopts an adaptive pooling method for dynamic perception of power data characteristics. The specific implementation process is as follows: Figure 1 shown.

[0047] First, sensors and data acquisition equipment distributed in various power stations, substations, transmission lines and user terminals are used to collect multi-source historical power data and multi-source real-time power data. Multi-source historical power data (first power data) covers power generation data over the past period of time, including power generation data, transmission data, distribution data, power consumption data, and power trading data. Multi-source real-time power data (second power data) is the above-mentioned data collected in real time to reflect the current operating status of the power grid.

[0048] Then, the collected first power data is preprocessed to improve the data quality and obtain the third power data, which provides a reliable basis for subsequent analysis.

[0049] Specifically, for the first power data Perform multi-scale decomposition and improve the wavelet packet basis function to meet the following requirements: , where j is the scale factor, k is the translation factor, and m is the branch index of the decomposition. , is the traditional basis function, is an improved basis function. After the wavelet packet transformation, the coefficients of different scales and branches are obtained. : . Perform soft threshold processing on wavelet packet data. , where n is the number of wavelet packet coefficients of this scale and branch, is the noise estimate of the scale and branch. The processed coefficients are obtained , calculated as: Finally, the processed wavelet packet coefficients are processed by using the same improved wavelet packet basis function as in the decomposition process. Reconstruct and obtain the third power data , calculated as: After preprocessing, the noise of the data is effectively suppressed and the characteristic structure is retained, providing high-quality data for subsequent model training.

[0050] The adaptive pooling model is constructed as follows. Figure 3 As shown, it includes data input layer, dynamic perception layer, pooling operation layer and output layer. Input layer: receives pre-processed power data; dynamic perception layer: dynamically adjusts the perception window size according to the characteristics of the input data; pooling operation layer: adopts multi-dimensional constrained pooling strategy based on the window size output by the dynamic perception layer; output layer: outputs the pooled result.

[0051] Specifically, the data input layer adopts a parallel processing architecture, which can simultaneously receive and process data from multiple channels, improve data processing efficiency, and ensure rapid access and transmission of data. The dynamic perception layer first quantifies the data features of the comprehensive power data transmitted from the input layer. For example, the variance of statistical indicators of different types of power data is calculated, such as calculating the variance of current data over a period of time, to evaluate the degree of data fluctuation. Three levels of low fluctuation, medium fluctuation and high fluctuation are set, corresponding to different variance range thresholds. If the current data variance is less than the set low fluctuation threshold, it is judged as a low fluctuation level; if it is within the medium fluctuation threshold range, it is a medium fluctuation level; if it is greater than the high fluctuation threshold, it is a high fluctuation level. The corresponding perception window size is pre-set for each fluctuation level. For example, a low fluctuation level corresponds to a smaller window, which can finely capture local data features; a high fluctuation level corresponds to a larger window, which is used to obtain more extensive trend information. When the fluctuation level of the data changes, such as from low fluctuation to medium fluctuation, the dynamic perception layer will adjust the perception window size accordingly to adapt to data changes in a timely manner.

[0052] Special extraction operations are required between the input layer and the dynamic perception layer, such as Figure 2 As shown: including data union, feature screening, feature weighting and feature fusion.

[0053] Joint feature extraction: For the pre-processed power data in the input layer, convolutional neural networks and gated recurrent units are used to extract spatial features and time series features respectively; specifically, the pre-processed power data is input into the convolutional neural network CNN, and the convolution layer of CNN slides on the data through multiple convolution kernels to extract the local spatial features of the data. Nonlinear factors are introduced through activation functions to enhance the expressive power of the model. The pooling layer downsamples the feature map to reduce the data dimension and improve computational efficiency. At the same time, the data is input into the gated recurrent unit GRU. GRU controls the flow of information by updating gates and resetting gates to effectively capture the time series features of power data, such as the change trend of load over time. Finally, the spatial features extracted by CNN and the time series features extracted by GRU are spliced ​​to complete the joint feature extraction.

[0054] Feature screening: Calculate the mutual information between each feature, construct a mutual information matrix, set a mutual information threshold, and when the correlation coefficient of two features exceeds the threshold, retain the features that have a more critical impact on the operating status of the power system, and remove redundant features with too high correlation; specifically, after completing the joint extraction of features, carry out feature screening for the obtained spatial features and time series features. First, calculate the mutual information of each two features, consider the correlation between them in detail, and integrate these calculation results to construct a mutual information matrix. Then, set a reasonable mutual information threshold based on the characteristics of the power system and actual operation requirements. Then, traverse the mutual information matrix to find feature pairs whose correlation coefficients exceed the threshold. By consulting the operating rules of the power system, historical fault cases, etc., evaluate the degree of influence of each feature on the operating status of the power system, retain the features that have a more critical impact, and remove redundant features.

[0055] Feature weighting: The attention mechanism is used to assign weights to the screened features, and the weight assignment is dynamically adjusted according to different operating scenarios and the importance of data features; specifically, in the feature weighting stage, a risk situation assessment index system is constructed based on the real-time operating parameters of the power system, such as load rate, voltage deviation, equipment health status, etc. For different operating scenarios, such as peak load, valley load, fault repair, etc., the weights of each risk indicator in the current scenario are determined by the fuzzy comprehensive evaluation method. Then, the screened features are correlated with the risk indicators, and the initial weights are assigned to the features based on the degree of correlation. During operation, the dynamic changes of the power system are monitored in real time. When the risk situation changes, the feature weights are dynamically adjusted using the reinforcement learning algorithm.

[0056] Feature fusion: Use feature splicing to fuse different types of features to form a comprehensive feature vector, and then perform dimensionality reduction to retain key data information. Specifically, the different types of features after feature weighting are spliced ​​in a specific order to construct a comprehensive feature vector. In order to further optimize the data, an innovative dimensionality reduction method based on quantum-inspired principal component analysis is adopted. This method introduces the superposition and entanglement state characteristics of quantum bits, and extracts the main components of the data more efficiently on the basis of traditional principal component analysis. First, the comprehensive feature vector is quantum encoded and mapped to the quantum space. The parallelism and entanglement characteristics of the quantum state are used to quickly calculate the projection of the feature vector on the new quantum principal component. By setting appropriate quantum revolving gate parameters, the extraction direction of the principal component is dynamically adjusted, so as to maximize the retention of the most critical information for power system analysis during the dimensionality reduction process.

[0057] Then, in the dynamic perception layer, the specific steps of dynamically adjusting the perception window size according to the characteristics of the input data include data feature quantification, fluctuation level division, window mapping size and dynamic adjustment. The dynamic perception layer quantitatively analyzes the comprehensive feature vector obtained through the feature extraction operation, calculates the variance of the statistical indicators of each feature, and evaluates the degree of data fluctuation. Three levels of low fluctuation, medium fluctuation and high fluctuation are set, corresponding to different fluctuation ranges, and divided according to the degree of data fluctuation. The corresponding perception window size is pre-set for each fluctuation level, and the perception window size is adjusted accordingly when the fluctuation level of the data changes. For example, when the power load fluctuation is small for a period of time and is at a low fluctuation level, the perception window is set to a small value to focus on local details; when the load fluctuation increases and enters a high fluctuation level, the perception window is expanded to capture a wider range of change trends.

[0058] The pooling operation layer includes a multi-dimensional constrained pooling strategy including window feature mapping, physical constraint pooling and feature sensitive fusion.

[0059] Window feature mapping: Map the window size determined by the dynamic perception layer to the pooling dimension parameters; specifically, according to preset rules, the window time length and width determined by the dynamic perception layer are mapped to the step size and kernel size of the pooling operation in the time and feature dimensions, respectively.

[0060] Physical constraint pooling: Calculate the constraint violation degree for the power data in each window, set the violation threshold, retain the data points greater than the violation threshold, and perform average pooling on the remaining points. The violation degree calculation of the physical constraint pooling includes reconstruction error, calculation residual, and calculation violation degree. Reconstruction error: Build an autoencoder model and calculate the Euclidean distance between the original data and the reconstructed data as the reconstruction error; Calculate residual: Based on the power node admittance matrix, calculate the absolute difference between the actual value and the theoretical value of active power and reactive power as the residual; Calculate violation degree: Based on the weighted fusion of reconstruction error and calculation residual, form a violation degree indicator.

[0061] Frequency-sensitive fusion: Perform short-time Fourier transform on the pooled data, extract the frequency components, and calculate the weighted sum of the frequency weights to obtain the final features.

[0062] Real-time detection: The second power data is preprocessed and then input into the trained adaptive pooling model to obtain the detection result; specifically, the same preprocessing process as that for processing the first power data is first adopted. Subsequently, the preprocessed second power data is input into the trained adaptive pooling model. The input layer in the model receives the data, the window size is adjusted by the dynamic perception layer, and after processing by the pooling operation layer, the detection result is output by the output layer, which contains the key feature information of the power data.

[0063] Analysis and evaluation: The output results are compared with the normal threshold of the grid data characteristics to evaluate the current operating status of the power data; specifically, the detection results output by the model are compared with the pre-set normal threshold of the grid data characteristics. If it is detected that the current characteristic value of a certain transmission line exceeds the normal current threshold range, or the voltage fluctuation characteristics are abnormal, it is judged that the power data operating status of the line is abnormal. Based on these comparison results, the operating status of each part of the current power grid can be comprehensively evaluated, potential fault hazards can be discovered in a timely manner, and strong support can be provided for power grid operation and maintenance decisions.

[0064] The above description is only a preferred embodiment of the present invention and does not limit the present invention in other forms. Any technician familiar with the profession may use the technical content disclosed above to change or modify it into an equivalent embodiment with equivalent changes and apply it to other fields. However, any simple modification, equivalent change and modification made to the above embodiment based on the technical essence of the present invention without departing from the content of the technical solution of the present invention still falls within the protection scope of the technical solution of the present invention.

Claims

1. An adaptive pooling method for dynamic perception of power data characteristics, characterized in that: The following steps are involved: S1. Data acquisition: Collect multi-source historical power data and multi-source real-time power data; The multi-source historical power data is recorded as first power data, and the multi-source real-time power data is recorded as second power data; S2, data preprocessing: preprocessing the collected first power data to obtain third power data; S3, model construction: build an adaptive pooling model; The adaptive pooling model includes a data input layer, a dynamic perception layer, a pooling operation layer and an output layer; Input layer: receives pre-processed power data; Dynamic perception layer: dynamically adjusts the perception window size according to the characteristics of the input data; Pooling operation layer: Based on the window size output by the dynamic perception layer, a multi-dimensional constrained pooling strategy is adopted; Output layer: output the result after pooling; S4, model training: inputting the third power data into the adaptive pooling model for training, and obtaining the trained model by continuously adjusting the model parameters; S5, real-time detection: pre-processing the second power data and inputting it into the trained adaptive pooling model to obtain a detection result; S6, analysis and evaluation: compare the output results with the normal threshold of the grid data characteristics to evaluate the operating status of the current power data; The pooling operation layer includes adopting a multi-dimensional constrained pooling strategy including window feature mapping, physical constraint pooling and feature sensitive fusion; Window feature mapping: maps the window size determined by the dynamic perception layer to the pooling dimension parameter; Physical constraint pooling: Calculate the constraint violation degree for the power data in each window, set the violation degree threshold, retain the data points with a value greater than the violation degree threshold, and perform average pooling on the remaining points; Frequency-sensitive fusion: Perform short-time Fourier transform on the pooled data, extract the frequency components, and calculate the weighted sum of the frequency weights to obtain the final features.

2. According to claim 1, the adaptive pooling method for dynamic perception of power data characteristics is characterized in that: The multi-source historical power data and multi-source real-time power data include power generation data, power transmission data, power distribution data, power consumption data, and power transaction data.

3. The adaptive pooling method for dynamic perception of power data characteristics according to claim 1 is characterized in that: The specific implementation of obtaining the third power data is: S21, the first power data Perform multi-scale decomposition and improve the wavelet packet basis function to meet the following requirements: , where j is the scale factor, k is the translation factor, and m is the branch index of the decomposition. , is the traditional basis function, is the improved basis function; S22, after wavelet packet transformation, the coefficients of different scales and branches are obtained : ; S23, perform soft threshold processing on the wavelet packet data to obtain the processed coefficients , calculated as: ; S24, finally use the same improved wavelet packet basis function as in decomposition to process the wavelet packet coefficients Reconstruct and obtain the third power data , calculated as: .

4. The adaptive pooling method for dynamic perception of power data characteristics according to claim 3 is characterized in that: The soft threshold in step S23 The calculation method is: , where n is the number of wavelet packet coefficients of this scale and branch, is the noise estimate for this scale and branch.

5. The adaptive pooling method for dynamic perception of power data characteristics according to claim 1 is characterized in that: The input layer and the dynamic perception layer of the model constructed in step S3 need to undergo special extraction operations: including data union, feature screening, feature weighting and feature fusion; Joint feature extraction: For the pre-processed power data in the input layer, convolutional neural network and gated recurrent unit are used to extract spatial features and time series features respectively; Feature screening: Calculate the mutual information between features, construct a mutual information matrix, and set a mutual information threshold. When the correlation coefficient of two features exceeds the threshold, retain the feature that has a more critical impact on the operating status of the power system and remove redundant features with too high correlation. Feature weighting: Use the attention mechanism to assign weights to the selected features. The weight assignment is dynamically adjusted according to different operating scenarios and the importance of data features. Feature Fusion: The feature splicing method is used to fuse different types of features to form a comprehensive feature vector, and then the dimension is reduced to retain the key information of the data.

6. The adaptive pooling method for dynamic perception of power data characteristics according to claim 1 is characterized in that: In the dynamic perception layer, the specific steps of dynamically adjusting the perception window size according to the characteristics of the input data include data feature quantification, fluctuation level division, window mapping size and dynamic adjustment; Data feature quantification: Quantitatively analyze the comprehensive feature vector obtained through feature extraction, calculate the variance of the statistical index of each feature, and evaluate the degree of data fluctuation; Volatility level classification: set three levels of low volatility, medium volatility and high volatility, corresponding to different volatility ranges, set thresholds and classify according to the degree of data volatility; Window size mapping: pre-set the corresponding perception window size for each fluctuation level; Dynamic adjustment: When the volatility level of the data changes, the perception window size is adjusted accordingly.

7. The adaptive pooling method for dynamic perception of power data characteristics according to claim 1 is characterized in that: The violation degree calculation of the physical constraint pooling includes reconstruction error, calculation residual, and calculation violation degree; Reconstruction error: Build an autoencoder model and calculate the Euclidean distance between the original data and the reconstructed data as the reconstruction error; Calculate the residual: Based on the power node admittance matrix, calculate the absolute difference between the actual value and the theoretical value of the active power and reactive power as the residual; Calculate the violation degree: Based on the weighted fusion of the reconstruction error and the calculation residual, a violation degree indicator is formed.

Images