Distribution network load prediction method and system based on transboundary heterogeneous big data

Through feature extraction and deep learning model of cross-border heterogeneous big data, combined with multi-scale analysis and adaptive algorithms, a hybrid prediction model is constructed, which solves the multi-scale and interpretability problems of distribution network load prediction in the existing technology, and realizes the High precision and stable load prediction.

CN119070270APending Publication Date: 2024-12-03ELECTRIC POWER RES INST OF GUANGXI POWER GRID CO LTD
View PDF 0 Cites 3 Cited by

Patent Information

Application Number
CN202410959989.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-17
Publication Date
2024-12-03

AI Technical Summary

Technical Problem

The existing distribution network load prediction technology has shortcomings in processing cross-border heterogeneous big data, adaptive feature extraction, multi-scale prediction and result application, and it is difficult to make full use of multi-source data information, and the prediction model is difficult to take into account both short-term fluctuations and long-term Trends, lack of interpretability and practical application value.

Method used

Using the distribution network load prediction method based on cross-border heterogeneous big data, feature extraction and deep feature extraction are performed by obtaining historical meteorological data, mobile base station population data, station area historical load data and equipment operation data, combined with fast Fourier transform and discrete wavelet transformation, form the initial feature set and the deep feature set, and select the optimal feature subset through the adaptive weight optimization particle optimization algorithm to build a hybrid prediction model, including extreme gradient enhancement algorithm, variational autoencoder and time series Model.

Benefits of technology

It significantly improves the accuracy and reliability of distribution network load prediction, makes full use of multi-source data information, enhances the generalization ability and prediction effect of the model, takes into account short-term fluctuations and long-term trends, improves the accuracy and stability of the prediction, and provides Explainability and practical application value.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119070270A_ABST
    Figure CN119070270A_ABST
Patent Text Reader

Abstract

The invention discloses a cross-boundary heterogeneous big data-based distribution network load prediction method and system, and relates to the technical field of distribution network load prediction, and the method comprises the steps: obtaining the cross-boundary heterogeneous big data of a target distribution area, and carrying out the feature extraction of the cross-boundary heterogeneous big data, and forming an initial feature set; carrying out transfer learning by using a pre-trained convolutional neural network model, carrying out depth feature extraction on the initial feature set to obtain a depth feature set, and selecting an optimal feature subset of the initial feature set and the depth feature set through an adaptive weight optimization particle optimization algorithm; and constructing a hybrid prediction model, and performing distribution network load prediction on the optimal feature subset. According to the method, more efficient feature extraction and selection are realized by collecting the transboundary heterogeneous big data and multiple algorithms, the precision and stability of distribution network load prediction are improved through the hybrid prediction model, the interpretability and practical application value of a prediction result are remarkably enhanced, and intelligent management and optimization of a power system are facilitated.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of distribution network load prediction, and specifically to a distribution network load prediction method and system based on cross-border heterogeneous big data. Background Art

[0002] As a key link in the planning and operation of power systems, distribution network load forecasting has always been a hot topic of research in academia and industry. With the rapid development of smart grid and big data technologies, traditional load forecasting methods based on a single data source have been unable to meet the increasingly complex needs of power systems. In recent years, researchers have begun to explore the use of multi-source heterogeneous data to improve forecasting accuracy. For example, some scholars have proposed a forecasting model that combines meteorological data and historical load data, and other research teams have tried to incorporate socioeconomic indicators into the forecasting framework. In terms of algorithms, from early time series analysis and regression models to recent machine learning and deep learning methods such as support vector machines, random forests, and long short-term memory networks, forecasting technology has been continuously innovated. However, these methods still face many challenges when dealing with large-scale heterogeneous data, such as low feature extraction efficiency, insufficient model generalization ability, and difficulty in capturing complex spatiotemporal correlations.

[0003] Although the existing technology has made progress in some aspects, it still has significant shortcomings. First, most methods only focus on a single or a few data sources, and fail to fully utilize the rich information contained in cross-border heterogeneous big data. Secondly, in terms of feature extraction and selection, existing methods often adopt a fixed feature engineering process and lack the ability to adapt to different scenarios. Furthermore, in the construction of prediction models, it is difficult for a single model to take into account both short-term fluctuations and long-term trends at the same time, and simple model integration is difficult to give full play to the advantages of each model. In addition, most of the existing methods focus on improving prediction accuracy, while ignoring the interpretability and practical application value of the prediction results, such as the lack of an automatic early warning mechanism. These shortcomings limit the application effect of load forecasting technology in complex distribution network systems. Summary of the invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the technical problem solved by the present invention is: how to provide a distribution network load forecasting method based on cross-border heterogeneous big data to solve the shortcomings of existing distribution network load forecasting technology in processing cross-border heterogeneous big data, adaptive feature extraction, multi-scale prediction and result application.

[0006] To solve the above technical problems, the present invention provides the following technical solutions: a distribution network load prediction method based on cross-border heterogeneous big data, comprising: obtaining cross-border heterogeneous big data of a target distribution station area, and forming an initial feature set by performing feature extraction on the cross-border heterogeneous big data; performing transfer learning using a pre-trained convolutional neural network model, performing deep feature extraction on the initial feature set to obtain a deep feature set, and selecting the optimal feature subset of the initial feature set and the deep feature set through an adaptive weighted optimization particle optimization algorithm; constructing a hybrid prediction model to perform distribution network load prediction on the optimal feature subset.

[0007] As a preferred solution of the distribution network load prediction method based on cross-border heterogeneous big data described in the present invention, the cross-border heterogeneous big data includes historical meteorological data, mobile base station population data, historical load data of the substation area, and historical equipment operation data.

[0008] As a preferred solution of the distribution network load prediction method based on cross-border heterogeneous big data described in the present invention, wherein: feature extraction is performed on the historical meteorological data, mobile base station population data, station area historical load data, and historical equipment operation data, and the extracted feature sets are integrated into an initial feature set; the feature extraction includes the following steps:

[0009] Preprocessing and aligning the cross-border heterogeneous big data; performing a fast Fourier transform operation on the preprocessed and aligned cross-border heterogeneous big data to obtain frequency domain features after fast Fourier transform:

[0010] F(D i )=FFT(D i );

[0011] Among them, D i is the original data of the ith data source, specifically historical meteorological data D1, mobile base station population data D2, station area historical load data D3, and historical equipment operation data D4; FFT (D i ) is the frequency domain feature after fast Fourier transform of the i-th data source.

[0012] The preprocessed and aligned cross-border heterogeneous big data is subjected to multi-level wavelet decomposition to obtain multi-scale features:

[0013]

[0014] Among them, DWT (D i ) is the i-th data source D i Multi-scale features after discrete wavelet transform; ψ i is the i-th wavelet basis function used for discrete wavelet transform DWT; is the i-th data source D i The coefficients obtained after discrete wavelet transform.

[0015] The multi-scale features are harmonically averaged to obtain the multi-scale features after harmonization:

[0016]

[0017] The frequency domain features and the multi-scale features after harmonic averaging are weightedly fused to obtain the initial feature set F:

[0018]

[0019] Among them, α i The weights used for weighted fusion are set based on feature importance or experience; is the multi-scale feature obtained by harmonizing and averaging the discrete wavelet transform results of the i-th data source; F is the multi-scale feature obtained by averaging the discrete wavelet transform results of the i-th data source; i The comprehensive extracted feature set is also called the initial feature set.

[0020] As a preferred solution of the distribution network load prediction method based on cross-border heterogeneous big data described in the present invention, wherein: the use of a pre-trained convolutional neural network model for transfer learning, deep feature extraction of the initial feature set, and the deep feature set obtained include selecting multiple pre-trained convolutional neural network models, freezing the bottom convolution layer of each pre-trained convolutional neural network model, and adding a fully connected layer adapted to the current distribution network load on the top; taking the initial feature set F as input, extracting deep features through the convolution layer of the pre-trained convolutional neural network model, specifically assuming that the deep features extracted by the kth pre-trained convolutional neural network model are Extract all pre-trained convolutional neural network models Fusion is performed to obtain the deep feature set F deep .

[0021] As a preferred solution of the distribution network load prediction method based on cross-border heterogeneous big data described in the present invention, wherein: selecting the optimal feature subset of the initial feature set and the deep feature set by the adaptive weighted optimization particle optimization algorithm includes converting the deep feature set F deep and the initial feature set F to form an extended feature set F e ; In the adaptive weighted optimization particle optimization algorithm, the inertia weight is dynamically adjusted according to the convergence of the algorithm:

[0022]

[0023] Among them, w max and w minare the maximum and minimum values ​​of the inertia weight respectively, t is the current iteration number; τ is the decay time constant, which controls the decay rate of the inertia weight.

[0024] Dynamically adjust the acceleration constants c1 and c2 according to the fitness value of the particle:

[0025]

[0026]

[0027] Among them, c 1,max 、c 1,min 、c 2,max and c 2,min are the maximum and minimum values ​​of the acceleration constant respectively; f(x i (t)) is the fitness value of particle i at the tth iteration; is the average value of the fitness of all particles at the tth iteration; σ(t) is the standard deviation of the fitness of all particles at the tth iteration.

[0028] Define the fitness function and update the particle position; the formula for updating the particle position is:

[0029]

[0030] in, is a random threshold that follows uniform distribution U(0,1), that is, a random number uniformly distributed between 0 and 1.

[0031] Update the individual best position and the global best position. Specifically, for each particle i, if the current fitness value is better than its historical best fitness value, update the individual best position p i , update the global optimal position g to the particle position corresponding to the current optimal fitness value.

[0032] Check the convergence conditions and select the optimal feature subset. Specifically, if the maximum number of iterations T is reached or the global optimal fitness value has not improved in multiple iterations, the algorithm converges. After convergence, the feature subset corresponding to the global optimal position g is selected as the final feature selection result, recorded as the optimal feature subset F o .

[0033] As a preferred solution of the distribution network load forecasting method based on cross-border heterogeneous big data described in the present invention, the hybrid prediction model is constructed based on the tree model of the extreme gradient boosting algorithm, the variational autoencoder and the time series model.

[0034] As a preferred solution of the distribution network load forecasting method based on cross-border heterogeneous big data described in the present invention, the construction process of the hybrid forecasting model includes:

[0035] Set up extreme gradient boosting model for:

[0036]

[0037] Suppose variational autoencoder for:

[0038]

[0039] Assume time series model for:

[0040]

[0041] Based on the extreme gradient boosting model Variational Autoencoder and time series models The hybrid prediction model is determined as:

[0042]

[0043] To further solve the above technical problems, the present invention provides the following technical solutions: a distribution network load prediction system based on cross-border heterogeneous big data, comprising: a first data processing module, used to obtain cross-border heterogeneous big data of a target distribution station area, and form an initial feature set by performing feature extraction on the cross-border heterogeneous big data; a second data processing module, used to perform transfer learning using a pre-trained convolutional neural network model, perform deep feature extraction on the initial feature set to obtain a deep feature set, and select the optimal feature subset of the initial feature set and the deep feature set through an adaptive weighted optimal particle optimization algorithm; a hybrid prediction module, used to construct a hybrid prediction model to perform distribution network load prediction on the optimal feature subset.

[0044] A computer device includes a memory and a processor, wherein the memory stores a computer program, and is characterized in that when the processor executes the computer program, the steps of the distribution network load prediction method based on cross-border heterogeneous big data as described above are implemented.

[0045] A computer-readable storage medium having a computer program stored thereon, characterized in that when the computer program is executed by a processor, the steps of the distribution network load prediction method based on cross-border heterogeneous big data as described above are implemented.

[0046] Beneficial effects of the present invention: The present invention significantly improves the accuracy and reliability of distribution network load prediction by integrating multi-source heterogeneous data and advanced algorithm models; integrates historical meteorological data, mobile base station population data, historical load data of substations and equipment operation data, and makes full use of the rich information contained in cross-border data; adopts methods such as fast Fourier transform and discrete wavelet transform for multi-scale feature extraction, and selects the optimal feature subset through adaptive weighted optimization particle optimization algorithm, thereby improving the generalization ability and prediction effect of the model; combines extreme gradient boosting algorithm, variational autoencoder and time series model, takes into account both short-term fluctuations and long-term trends, and enhances the accuracy and stability of prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0048] Figure 1 An overall flow chart of a distribution network load prediction method based on cross-border heterogeneous big data provided for one embodiment of the present invention.

[0049] Figure 2 This is a computer equipment diagram of the distribution network load prediction method based on cross-border heterogeneous big data of the present invention. DETAILED DESCRIPTION

[0050] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without creative work should fall within the scope of protection of the present invention.

[0051] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0052] Example 1

[0053] Reference Figure 1 , which is the first embodiment of the present invention, and provides a distribution network load prediction method based on cross-border heterogeneous big data, comprising:

[0054] In related technologies, the distribution network load forecasting in the substation often only focuses on a single or a few data sources, and fails to fully utilize the rich information contained in cross-border heterogeneous big data. In terms of feature extraction and selection of collected data, a fixed feature engineering process is usually adopted, which lacks the ability to adapt to different scenarios. In addition, when building a distribution network load forecasting model, a single model is difficult to take into account both short-term fluctuations and long-term trends at the same time, and simple model integration is difficult to give full play to the advantages of each model.

[0055] This application provides a method that can effectively solve the above-mentioned problems. Next, we will combine multiple embodiments to explain in detail how to implement the distribution network load prediction method based on cross-border heterogeneous big data;

[0056] Figure 1 The overall flow chart of a distribution network load forecasting method based on cross-border heterogeneous big data is shown, including:

[0057] S1: Obtain cross-border heterogeneous big data of the target distribution station area, and form an initial feature set by extracting features from the cross-border heterogeneous big data.

[0058] Specifically, cross-border heterogeneous big data includes historical meteorological data, mobile base station population data, historical load data of substations, and historical equipment operation data.

[0059] It should be noted that historical meteorological data can be meteorological observation records of the target distribution substation area in the past unit time, including temperature, humidity, precipitation, wind speed, wind direction, atmospheric pressure, etc. For distribution network load forecasting, weather conditions directly affect electricity consumption, so meteorological data is extremely important. For example, extremely high temperature weather will cause a surge in air conditioning power consumption, while rainy weather may reduce lighting power consumption. By analyzing the relationship between historical meteorological data and load changes, the sensitivity of the forecast model to weather factors can be improved.

[0060] Mobile base station population data can be data that estimates the population distribution and mobility in the target distribution network area by analyzing the signal flow of the mobile communication base station. The present invention can reflect the population density, mobility pattern and activity pattern in the area through the mobile base station population data. Mobile base station population data provides population dynamic information for distribution network load prediction. For example, the population density in the commercial area is high on weekdays, but it may be lower on weekends, which will directly affect the electricity demand. By integrating mobile base station population data, the present invention can more accurately capture the load changes caused by population mobility.

[0061] The historical load data of the substation area can refer to the electricity consumption record of the target distribution substation area in the past period of time, usually in the form of time series, recording the power load at different time points (such as every 15 minutes or every hour). The historical load data of the substation area is the core basis of the distribution network load forecast. It reflects the power consumption pattern, periodic changes and long-term trends in the area. By analyzing the historical load data, it is possible to identify characteristics such as intraday load curves, differences between weekdays and weekends, and seasonal changes, providing key historical information.

[0062] Historical equipment operation data can refer to the operating status and performance records of various equipment (such as transformers, switches, lines, etc.) in the distribution network. These data can specifically be parameters such as equipment load rate, temperature, voltage, current, as well as maintenance records, fault events and other information. Equipment operation data reflects the health status and capacity limitations of the distribution network in the distribution network load forecast. For example, some equipment operating close to full load may limit load growth, while regular maintenance or failure of equipment may cause temporary changes in load distribution. By integrating equipment operation data, network constraints can be considered to improve the practicality and reliability of distribution network load forecasting.

[0063] Furthermore, the overall idea of ​​feature extraction for the above four types of cross-border heterogeneous big data is: first, data preprocessing and alignment are carried out, specifically to unify the data format: convert all data sources into the same timestamp format and sampling frequency; missing value processing: use interpolation to fill in missing data and smooth outliers; data standardization: standardize all data sources to ensure that each data source has the same dimension.

[0064] Secondly, perform a fast Fourier transform (FFT). In distribution network load forecasting, load data and related data usually have periodic characteristics, and these frequency components can be quickly extracted through FFT:

[0065] F(D i )=FFT(D i );

[0066] Among them, D i is the original data of the ith data source, which in the present invention is the historical meteorological data D1, the mobile base station population data D2, the historical load data of the area D3, and the historical equipment operation data D4. i ) is the frequency domain feature after fast Fourier transform of the i-th data source.

[0067] Thirdly, multi-level wavelet decomposition (multi-scale analysis). Multi-scale analysis of cross-border heterogeneous big data can identify various periodicity and mutation characteristics and improve the accuracy of prediction. Therefore, this embodiment uses multi-level wavelet decomposition to extract the multi-scale characteristics of the signal and capture important information on different time scales. Since only the most important scale is selected for calculation, the complexity can be reduced. The specific formula of multi-level wavelet decomposition is:

[0068]

[0069] Among them, DWT (D i ) is the i-th data source D i Multi-scale features after discrete wavelet transform; ψ i is the i-th wavelet basis function used for discrete wavelet transform DWT; is the i-th data source D i The coefficients obtained after discrete wavelet transform.

[0070] Then, harmonic mean and weighted fusion are performed. In the distribution network load forecasting, the prediction ability of the model can be improved by comprehensively considering the multiple features of meteorological data, population data, load data and equipment data. This embodiment reduces the impact of outliers by harmonic mean, and combines frequency domain and time domain features to make feature extraction more comprehensive and robust. Weighted fusion ensures that the contribution of different features can be adjusted according to actual conditions. The specific formula of harmonic mean and weighted fusion is:

[0071]

[0072]

[0073] Among them, α i The weights used for weighted fusion are set based on feature importance or experience; is the multi-scale feature obtained by harmonizing and averaging the discrete wavelet transform results of the i-th data source; F is the multi-scale feature obtained by averaging the discrete wavelet transform results of the i-th data source; i The feature set extracted comprehensively from the historical meteorological data, mobile base station population data, station area historical load data, and historical equipment operation data is also the initial feature set.

[0074] Finally, combining the above steps, the complete expression of the initial feature set F is obtained as follows:

[0075]

[0076] In an optional implementation, since each cross-border heterogeneous big data has different characteristics in actual applications, it is necessary to adaptively adjust the feature extraction refinement method according to the specific characteristics. Since the historical load data of the substation has obvious periodicity and trend (such as daily cycle, weekly cycle), and is affected by many factors (such as weather, holidays, etc.), the above-mentioned feature extraction process can be directly applied to extract periodicity and multi-scale features. Since the meteorological data has obvious seasonality and periodicity and contains multiple variables, this embodiment independently applies FFT and DWT to each meteorological variable in the historical meteorological data, and then performs feature fusion.

[0077] Since the mobile base station population data has obvious spatiotemporal distribution characteristics, there may be high noise and outliers. Therefore, this embodiment uses harmonic mean to reduce the influence of noise and outliers, and captures the spatiotemporal variation characteristics of the mobile base station population data through multi-level wavelet decomposition. The specific process of feature extraction of mobile base station population data is as follows:

[0078] For example, it is assumed that the mobile base station population data of the target distribution station area has been obtained, there are n mobile base station population data variables, and the data set of each variable is recorded as D i (i=1,2,...,n). i Perform a time-space alignment operation (the existing time-space alignment scheme is used here, so no further details are given), and obtain D' i , for D' i Perform harmonic averaging to reduce noise:

[0079]

[0080] After harmonic and average H p (D i ') Perform Fast Fourier Transform FFT:

[0081] F p (H p (D i ′))=FFT(H p (D i ′));

[0082] After harmonic and average H p (D i ') Perform multi-level wavelet decomposition to obtain multi-scale features:

[0083]

[0084] The fast Fourier transform FFT and multi-scale features are combined to extract the feature F of the mobile base station population data. p :

[0085]

[0086] Finally, the feature F of the mobile base station population data p The complete expression is:

[0087]

[0088] Preferably, in the feature extraction process of mobile base station population data, the noise is reduced by harmonic averaging, and the periodicity and local features in the data are extracted by combining FFT and multi-level wavelet decomposition. Then, through feature fusion, abnormal population change events, such as sudden large-scale migration or major events, can be efficiently identified and warned in advance. This multi-level and multi-angle feature extraction and fusion significantly improves the accuracy and timeliness of anomaly detection. In addition, FFT and DWT not only extract key features, but also effectively compress the data. Through feature fusion, the compressed data still retains key information, thereby achieving efficient storage and transmission, while reducing the consumption of computing resources.

[0089] Furthermore, since the historical equipment operation data includes a variety of equipment states and operating parameters, which may contain a large number of outliers and noise, features can be extracted independently for the operating parameters of each device, and the harmonic mean can be used to process the outliers and noise.

[0090] S2: Use the pre-trained convolutional neural network model for transfer learning, perform deep feature extraction on the initial feature set to obtain the deep feature set, and select the optimal feature subset of the initial feature set and the deep feature set through the adaptive weighted optimization particle optimization algorithm.

[0091] Specifically, multiple pre-trained convolutional neural network models are selected, the bottom convolutional layer of each pre-trained convolutional neural network model is frozen, and a fully connected layer adapted to the current distribution network load is added on the top.

[0092] In a preferred embodiment, the pre-trained convolutional neural network model can be ResNet, ResNet, Inception, EfficientNet, etc. The purpose of selecting multiple pre-trained convolutional neural network models is to enhance the diversity and robustness of feature extraction. The current distribution network load specifically refers to the distribution network load prediction problem to be solved by the present invention. This embodiment uses a pre-trained convolutional neural network model as a feature extractor, and adds a fully connected layer on top to adapt to the load prediction task. The fully connected layer here can be designed according to the characteristics and requirements of the load prediction problem. For example: if it is a short-term load forecast (such as the load in the next few hours or days), fewer fully connected layers and neurons can be used to avoid overfitting; if it is a long-term load forecast (such as the load in the next few weeks or months), more fully connected layers and neurons may be required to capture long-term trends and complex patterns; if load forecasting is a regression problem, a linear activation function can be used in the last layer, and a regression loss function such as mean square error can be used; if load forecasting is converted into a classification problem (such as high, medium, and low categories of load levels), a softmax activation function can be used in the last layer, and a classification loss function such as cross entropy can be used. Therefore, the "fully connected layer adapted to the current distribution network load" is designed according to the characteristics and requirements of the distribution network load forecasting problem, and its purpose is to map the deep features extracted by the pre-trained model to the output space of the load forecasting to complete the load forecasting task. Through transfer learning, the present invention can utilize the feature extraction capability of the pre-trained model and solve the specific load forecasting problem by adding an adaptive fully connected layer.

[0093] Specifically, the initial feature set F extracted from S1 is used as input, and the deep features are extracted through the convolutional layer of the pre-trained convolutional neural network model. Specifically, the deep features extracted by the k-th pre-trained convolutional neural network model are set to

[0094]

[0095] Among them, P k is the kth pre-trained convolutional neural network model.

[0096] Extract all pre-trained convolutional neural network models Fusion is performed to obtain a more comprehensive and robust feature representation. The fusion method can be splicing or weighted averaging. Let the fused deep feature set be F deep :

[0097]

[0098] Among them, Fusion is the feature fusion function, and K is the number of pre-trained convolutional neural network models.

[0099] Specifically, the fused deep feature F deep and the initial feature set F to form an extended feature set F e :

[0100] F e =[F,F deep ].

[0101] In the adaptive weighted optimization particle optimization algorithm, an adaptive weight adjustment strategy is introduced to dynamically adjust the inertia weight w(t), acceleration constants c1 and c2 according to the convergence of the algorithm to balance the global exploration and local development capabilities. In this embodiment, a linearly decreasing inertia weight adjustment strategy can be adopted:

[0102]

[0103] Among them, w max and w min are the maximum and minimum values ​​of the inertia weight respectively, t is the current iteration number; τ is the decay time constant, which controls the decay rate of the inertia weight.

[0104] In addition, an adaptive acceleration constant adjustment strategy is introduced to dynamically adjust the acceleration constants c1 and c2 according to the fitness value of the particle. For example, a sigmoidal function based on the fitness value can be used to adjust the acceleration constant:

[0105]

[0106]

[0107] Among them, c 1,max 、c 1,min 、c 2,max and c 2,min are the maximum and minimum values ​​of the acceleration constant respectively; f(x i (t)) is the fitness value of particle i at the tth iteration; is the average value of the fitness of all particles at the tth iteration; σ(t) is the standard deviation of the fitness of all particles at the tth iteration.

[0108] Preferably, the present invention introduces a nonlinear inertia weight adjustment strategy and an adaptive acceleration constant adjustment strategy based on fitness value, which can more accurately balance global exploration and local development capabilities. This not only enhances the adaptability of the algorithm at different iteration stages, but also improves the convergence speed and optimization performance of the algorithm, thereby more effectively selecting the optimal feature subset in the distribution network load forecasting task, and improving the accuracy and robustness of the forecast.

[0109] Furthermore, in the fitness function, a regularization term is introduced to control the sparsity of the feature subset, while considering the prediction performance of the feature subset and the redundancy between features. The fitness function can be designed as:

[0110]

[0111] Among them, Redundancy(S) represents the redundancy between features in feature subset S, which can be measured by indicators such as correlation coefficient or mutual information, and γ is the weight coefficient for controlling redundancy. Feature subset S comes from the extended feature set F e Redundancy(S) represents the redundancy between features in feature subset S, which can be measured by indicators such as correlation coefficient or mutual information; α is the weight coefficient for controlling performance; β is the weight coefficient for controlling sparsity.

[0112] Furthermore, when the particle position is updated, a binary position update mechanism is used to make the particle position take the value of 0 or 1, indicating the selection state of the feature. The continuous speed value can be mapped to the [0,1] interval through the Sigmoid function, and then binarized according to the random threshold. The position update formula is:

[0113]

[0114] in, is a random threshold that follows a uniform distribution U(0,1), that is, a random number uniformly distributed between 0 and 1. This random threshold is used to determine whether a particle selects a specific feature; is the position of particle i in the jth dimension (corresponding to the jth feature) at the t+1th generation; position x i,j The value is binary. In the feature selection problem, 1 means that the feature is selected and 0 means that the feature is not selected. is the velocity of particle i in the jth dimension at generation t+1; velocity v i,j It is a continuous value, which affects the particle position update through speed; To speed The value is mapped to the [0,1] interval through the Sigmoid function. The Sigmoid function is defined as:

[0115]

[0116] Update the individual best position and the global best position. Specifically, for each particle i, if the current fitness value is better than its historical best fitness value, update the individual best position p i , update the global optimal position g to the particle position corresponding to the current optimal fitness value.

[0117] Check convergence conditions and select the optimal feature subset: If the maximum number of iterations T is reached or the global optimal fitness value has not been significantly improved in multiple iterations, the algorithm converges. After the algorithm converges, select the feature subset corresponding to the global optimal position g as the final feature selection result, recorded as the optimal feature subset F o .

[0118] S3: Build a hybrid prediction model to predict the distribution network load based on the optimal feature subset.

[0119] Specifically, the hybrid prediction model is built based on the tree model of the extreme gradient boosting algorithm, the variational autoencoder and the time series model.

[0120] Set up extreme gradient boosting model for:

[0121]

[0122] Suppose variational autoencoder for:

[0123]

[0124] Assume time series model for:

[0125]

[0126] Based on extreme gradient boosting model Variational Autoencoder and time series models The hybrid prediction model is determined as:

[0127]

[0128] Furthermore, the distribution network load forecasting is carried out on the optimal feature subset through the hybrid prediction model, and finally the distribution network load forecasting based on cross-border heterogeneous big data is completed.

[0129] In summary, the present invention significantly improves the accuracy and reliability of distribution network load forecasting by integrating multi-source heterogeneous data and advanced algorithm models; integrates historical meteorological data, mobile base station population data, historical load data of substations and equipment operation data, and makes full use of the rich information contained in cross-border data; uses methods such as fast Fourier transform and discrete wavelet transform for multi-scale feature extraction, and selects the optimal feature subset through an adaptive weighted particle optimization algorithm, thereby improving the generalization ability and prediction effect of the model; combines the extreme gradient boosting algorithm, variational autoencoder and time series model, taking into account both short-term fluctuations and long-term trends, and enhancing the accuracy and stability of the prediction.

[0130] Example 2

[0131] As an embodiment of the present invention, a distribution network load prediction system based on cross-border heterogeneous big data is provided, including: a first data processing module, used to obtain cross-border heterogeneous big data of a target distribution station area, and form an initial feature set by extracting features from the cross-border heterogeneous big data; a second data processing module, used to perform transfer learning using a pre-trained convolutional neural network model, perform deep feature extraction on the initial feature set to obtain a deep feature set, and select the optimal feature subset of the initial feature set and the deep feature set through an adaptive weighted optimization particle optimization algorithm; a hybrid prediction module, used to construct a hybrid prediction model, and perform distribution network load prediction on the optimal feature subset.

[0132] Example 3

[0133] Reference Figure 2 , is an embodiment of the present invention, which is different from the previous embodiment in that:

[0134] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc., which can store program codes.

[0135] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by an instruction execution system, device or apparatus (such as a computer-based system, a system including a processor, or other system that can fetch instructions from an instruction execution system, device or apparatus and execute instructions), or in conjunction with such instruction execution systems, devices or apparatuses. For the purposes of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate or transmit a program for use by an instruction execution system, device or apparatus, or in conjunction with such instruction execution systems, devices or apparatuses.

[0136] More specific examples of computer-readable media (a non-exhaustive list) include the following: an electrical connection with one or more wires (electronic device), a portable computer disk case (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disk read-only memory (CDROM). In addition, the computer-readable medium may even be a paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering or, if necessary, processing in another suitable manner, and then stored in a computer memory.

[0137] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, a plurality of steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0138] Example 4

[0139] As an embodiment of the present invention, a distribution transformer load forecasting method considering the population flow of communication base stations is provided. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.

[0140] The experimental design is as follows:

[0141] Data sources: historical meteorological data (temperature, humidity, rainfall, etc.), mobile base station population data (population density, activity patterns, etc.), historical load data of substations (daily load curves, etc.), and historical equipment operation data (equipment failure records, maintenance records, etc.).

[0142] The experimental methods are the present invention, prior art solution 1: traditional time series prediction method (such as ARIMA model) and prior art solution 2: machine learning method of single data source (such as random forest model based on historical load data of substation).

[0143] The experimental indicators are mean absolute percentage error (MAPE), root mean square error (RMSE), R2 value, data processing time (seconds), training time (seconds), memory usage (MB), prediction time (seconds), and response time to emergencies (seconds).

[0144] Table 1 Comparison with existing technology

[0145] index The present invention Prior art solution 1 Prior art solution 2 Mean Absolute Percentage Error (MAPE) 3.2% 7.8% 6.5% Root mean square error (RMSE) 4.5 10.2 8.9 <![CDATA[R 2 Value]]> 0.92 0.78 0.81 Data processing time (seconds) 120 180 150 Training time (seconds) 300 250 270 Memory usage (MB) 500 400 450 Prediction time (seconds) 10 15 12 Response time to emergencies (seconds) 30 45 40

[0146] As can be seen from Table 1, the present invention performs well in all indicators. Specifically, the mean absolute percentage error and root mean square error of the present invention are significantly lower than those of the prior art solution, indicating that its prediction accuracy is higher; R 2 The closer the value is to 1, the better the model's ability to fit the data. The present invention also shows higher efficiency in terms of data processing time and prediction time, and its memory usage is slightly higher but completely acceptable on modern computing devices. In addition, the present invention can adjust the prediction results more quickly when responding to sudden load changes, and the response time is significantly better than the existing technical solutions, reflecting its superiority in emergencies. Overall, the present invention is superior to existing technical solutions in terms of prediction accuracy, model stability, data processing efficiency, and emergency response capabilities, and has extremely high practical value and promotion prospects.

[0147] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A distribution network load prediction method based on cross-border heterogeneous big data, characterized in that: include: Obtaining cross-border heterogeneous big data of a target distribution station area, and forming an initial feature set by extracting features from the cross-border heterogeneous big data; Using a pre-trained convolutional neural network model for transfer learning, performing deep feature extraction on the initial feature set to obtain a deep feature set, and selecting the optimal feature subsets of the initial feature set and the deep feature set through an adaptive weighted optimization particle optimization algorithm; A hybrid prediction model is constructed to perform distribution network load prediction on the optimal feature subset.

2. The distribution network load prediction method based on cross-border heterogeneous big data according to claim 1 is characterized in that: The cross-border heterogeneous big data includes historical meteorological data, mobile base station population data, historical load data of substations, and historical equipment operation data.

3. The distribution network load prediction method based on cross-border heterogeneous big data according to claim 2 is characterized in that: Feature extraction is performed on the historical meteorological data, mobile base station population data, station area historical load data, and historical equipment operation data, and the extracted feature sets are integrated into an initial feature set; the feature extraction includes the following steps: Preprocessing and aligning the cross-border heterogeneous big data; The preprocessed and aligned cross-border heterogeneous big data is subjected to a fast Fourier transform operation to obtain the frequency domain features after the fast Fourier transform: F (D i )=FFT(D i ); Among them, D i is the original data of the ith data source, specifically historical meteorological data D1, mobile base station population data D2, station area historical load data D3, and historical equipment operation data D4; FFT (D i ) is the frequency domain feature after fast Fourier transform of the i-th data source; The preprocessed and aligned cross-border heterogeneous big data is subjected to multi-level wavelet decomposition to obtain multi-scale features: Among them, DWT (D i ) is the i-th data source D i Multi-scale features after discrete wavelet transform; ψ i is the i-th wavelet basis function used for discrete wavelet transform DWT; is the i-th data source D i The coefficients obtained after discrete wavelet transform; The multi-scale features are harmonically averaged to obtain the multi-scale features after harmonization: The frequency domain features and the multi-scale features after harmonic averaging are weightedly fused to obtain the initial feature set F: Among them, α i The weights used for weighted fusion are set based on feature importance or experience; is the multi-scale feature obtained by harmonizing and averaging the discrete wavelet transform results of the i-th data source; F is the multi-scale feature obtained by averaging the discrete wavelet transform results of the i-th data source; i The comprehensive extracted feature set is also called the initial feature set.

4. The distribution network load prediction method based on cross-border heterogeneous big data according to claim 3 is characterized by: The pre-trained convolutional neural network model is used for transfer learning, and deep feature extraction is performed on the initial feature set to obtain a deep feature set including: Select multiple pre-trained convolutional neural network models, freeze the bottom convolutional layer of each pre-trained convolutional neural network model, and add a fully connected layer on top that adapts to the current distribution network load; The initial feature set F is used as input, and the deep features are extracted through the convolution layer of the pre-trained convolutional neural network model. Specifically, the deep features extracted by the k-th pre-trained convolutional neural network model are set to Extract all pre-trained convolutional neural network models Fusion is performed to obtain the deep feature set F deep .

5. The distribution network load prediction method based on cross-border heterogeneous big data according to claim 4 is characterized in that: Selecting the optimal feature subsets of the initial feature set and the deep feature set by using an adaptive weighted optimization particle optimization algorithm includes: The deep feature set F deep and the initial feature set F to form an extended feature set F e ; In the adaptive weighted optimization particle optimization algorithm, the inertia weight is dynamically adjusted according to the convergence of the algorithm: Among them, w max and w min are the maximum and minimum values ​​of the inertia weight, respectively; t is the current iteration number; τ is the decay time constant, which controls the decay rate of the inertia weight; Dynamically adjust the acceleration constants c1 and c2 according to the fitness value of the particle: Among them, c 1,max 、c 1,min 、c 2,max and c 2,min are the maximum and minimum values ​​of the acceleration constant respectively; f(x i (t)) is the fitness value of particle i at the tth iteration; is the average value of the fitness values ​​of all particles at the tth iteration; σ(t) is the standard deviation of the fitness values ​​of all particles at the tth iteration; Define the fitness function and update the particle position; the formula for updating the particle position is: in, is a random threshold, which obeys the uniform distribution u(0,1), that is, a random number uniformly distributed between 0 and 1; Update the individual best position and the global best position. Specifically, for each particle i, if the current fitness value is better than its historical best fitness value, update the individual best position p i , update the global best position g to the particle position corresponding to the current optimal fitness value; Check the convergence conditions and select the optimal feature subset. Specifically, if the maximum number of iterations T is reached or the global optimal fitness value has not improved in multiple iterations, the algorithm converges. After convergence, the feature subset corresponding to the global optimal position g is selected as the final feature selection result, recorded as the optimal feature subset F o .

6. The distribution network load prediction method based on cross-border heterogeneous big data according to claim 5 is characterized in that: The hybrid prediction model is constructed based on a tree model of an extreme gradient boosting algorithm, a variational autoencoder and a time series model.

7. The distribution network load prediction method based on cross-border heterogeneous big data according to claim 6 is characterized in that: The construction process of the hybrid prediction model includes: Set up extreme gradient boosting model for: Suppose variational autoencoder for: Assume time series model for: Based on the extreme gradient boosting model Variational Autoencoder and time series models The hybrid prediction model is determined as:

8. A system using the distribution network load prediction method based on cross-border heterogeneous big data as described in any one of claims 1 to 7, characterized in that: include: A first data processing module is used to obtain cross-border heterogeneous big data of a target distribution station area, and to form an initial feature set by extracting features from the cross-border heterogeneous big data; A second data processing module is used to perform transfer learning using a pre-trained convolutional neural network model, perform deep feature extraction on the initial feature set to obtain a deep feature set, and select an optimal feature subset of the initial feature set and the deep feature set through an adaptive weighted optimization particle optimization algorithm; The hybrid prediction module is used to construct a hybrid prediction model to perform distribution network load prediction on the optimal feature subset.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the distribution network load prediction method based on cross-border heterogeneous big data described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the distribution network load prediction method based on cross-border heterogeneous big data described in any one of claims 1 to 7 are implemented.

Citation Information

Cited By

  • Cold storage load prediction method and system based on Internet of Things and artificial intelligence

    CN119903365A

  • Heat exchange system optimization method and system based on big data analysis

    CN120258203A

  • Medium and long term power load prediction method and system based on ensemble learning

    CN121939345A