Distributed flexible load modeling method and system for reliability evaluation of power distribution network

Abstract

The invention discloses a distributed flexible load modeling method and system for reliability evaluation of a power distribution network, and the method comprises the steps: carrying out the classification of power consumption behaviors of residential users based on a parallel calculation and merging classification method, and converting a classification result into a gray image; constructing a demand curve function set based on multi-information fusion, and generating an image training sample; taking the image training sample as input, and training a nonlinear demand price elasticity evaluation model based on a deep learning method; inputting a grayscale image into the trained model, extracting polynomial representation of a demand curve, and outputting a demand price elastic curve polynomial set; a distributed flexible load model is constructed based on historical data and a demand price elastic curve polynomial set so as to evaluate the electricity consumption of a single user under a given electricity price, the method can be applied to personalized dynamic pricing and power distribution network reliability evaluation models and the like, and powerful support is provided for optimized operation and sustainable development of a power system.

Description

Technical Field

[0001] The present invention relates to the technical field of power distribution network systems, and in particular to a distributed flexible load modeling method and system for power distribution network reliability assessment. Background Art

[0002] With the advancement of the global energy structure transformation, demand response (DR) technology in smart distribution networks is regarded as a key tool to improve the efficiency of power systems, ensure the security of power supply, and enhance the accuracy of distribution network reliability assessment. Demand response, as a way to regulate power demand through user participation, has become an indispensable part of the modern power market. By encouraging users to adjust their electricity consumption patterns according to changes in power supply and demand, demand response can not only effectively reduce the operating costs of the power system, reduce peak-to-valley differences, and improve the flexibility of the power grid, but also reduce the load on power equipment and extend the service life of power infrastructure, thereby promoting more sustainable energy development.

[0003] Driven by power market reform and green energy policies, more and more power consumers, especially residential users, have begun to participate in demand response programs. Residential users account for a considerable proportion of the power system, and household electricity consumption accounts for about 20% of global electricity consumption. Therefore, studying and optimizing residential demand response has become an important direction in power system and market research. Especially in the context of the increasing proportion of renewable energy (such as wind power and solar energy), residential demand response can not only help regulate power load, but also connect with the green power market and further promote the consumption of renewable energy.

[0004] Demand response programs can be roughly divided into two types: incentive-based and price-based. Incentive-based demand response programs usually regulate electricity consumption by encouraging users to reduce electricity consumption during high-demand periods or increase electricity consumption during low-demand periods through direct load control or reward mechanisms. For residential users, incentive-based demand response is often the more common form, mainly because household electricity consumption is decentralized and the power is small, and it is more effective to use direct incentives or control equipment loads. However, with the deepening reform of the power market and the increasing sensitivity of users to electricity prices, price-based demand response has gradually become an effective means of regulation. In price-based demand response, retailers set dynamic electricity prices based on real-time electricity supply and demand conditions, and encourage users to adjust their electricity consumption behavior when electricity prices fluctuate. This model can not only help optimize grid operation, but also effectively stimulate consumers' flexible response.

[0005] However, while price-driven demand response (Price-based DR) has significant potential in theory, it also faces many challenges.

[0006] First, the relationship between electricity demand and price is often nonlinear, and most existing studies are based on the linear Price Elasticity of Demand (PED) model, which limits the understanding of complex electricity consumption behavior. The traditional PED model assumes that the relationship between price and demand is linear, but in reality, changes in household electricity consumption are often affected by many factors, such as weather changes, equipment types, and electricity consumption time. Therefore, linear models often cannot fully capture these nonlinear characteristics, resulting in poor prediction results.

[0007] Secondly, deep learning-based methods, as a data-driven approach, have gradually been applied in demand response research. Deep learning models can extract implicit rules from a large amount of historical data and then predict the nonlinear relationship between electricity consumption patterns and prices. The advantages of these methods are their powerful modeling capabilities and flexibility, and their ability to handle complex and dynamic electricity demand. However, the effectiveness of deep learning models usually relies on a large amount of user data, especially detailed usage information of various electrical appliances in the home. But in practical applications, it is often not feasible to collect such large-scale data and conduct effective modeling, especially when it comes to the privacy information of individual users. Home appliance information and user preferences are usually private data, and these data may be incomplete or inaccurate. Therefore, the application of traditional deep learning methods in practical scenarios faces great challenges.

[0008] In order to solve these problems, research in recent years has begun to explore new methods. For example, some studies have proposed the use of multi-information fusion technology and cluster analysis methods to avoid over-reliance on accurate household appliance data. By classifying the electricity consumption patterns of residential users, researchers can propose more simplified and efficient prediction methods by comprehensively analyzing the common characteristics of user groups. These methods can not only improve the adaptability of the model, but also reduce the reliance on data integrity, making them more feasible in actual operations. In addition, as the complexity of demand response systems increases, traditional demand response models are often unable to handle large-scale user data or cope with the uncertainty and high dynamics in the power market. Summary of the invention

[0009] In order to solve the above problems, the present invention proposes a distributed flexible load modeling method and system for distribution network reliability assessment, which classifies user power consumption patterns without assumptions through a parallel computing and merging (PC&M) classification method, thereby improving the adaptability of the classification; multi-information fusion technology is used to generate training samples, and a nonlinear PED model is established in combination with a deep learning method and used for distributed flexible load modeling, which can be applied to personalized dynamic pricing, distribution network reliability assessment and other models.

[0010] The technical solution adopted by the present invention is as follows:

[0011] A distributed flexible load modeling method for distribution network reliability assessment includes:

[0012] Based on parallel computing and combined classification methods, residential users' electricity consumption behaviors are classified and the classification results are converted into grayscale images.

[0013] Construct a demand curve function set based on multi-information fusion and generate image training samples;

[0014] Taking the image training sample as input, training a nonlinear demand price elasticity evaluation model based on a deep learning method; inputting the grayscale image into the trained model, extracting the polynomial representation of the demand curve, and outputting a polynomial set of demand price elasticity curves;

[0015] Based on historical data and the polynomial set of demand price elasticity curves, a distributed flexible load model is constructed to evaluate the electricity consumption of a single user at a given electricity price, which serves as an input parameter for distribution network reliability evaluation.

[0016] Furthermore, the method of classifying the electricity consumption behavior of residential users based on parallel computing and combined classification, and converting the classification results into grayscale images, includes:

[0017] Data preprocessing: normalize the collected residential electricity consumption behavior data, use the principal component analysis algorithm to reduce the dimension, and evenly distribute the processed data into each data block;

[0018] Parallel calculation: Calculate the electricity consumption behavior samples in each data block, use the three indicators of cumulative proximity, eccentricity and standardized eccentricity for recursive calculation, and use Chebyshev inequality to determine the category of the new electricity consumption behavior samples;

[0019] Category merging: Merge the categories collected from each data block by finding the nearest neighbor category and merging when the merging conditions are met, and update the parameters of the merged category;

[0020] Data integration: Pair residential electricity consumption in each category with the corresponding 24-hour electricity price data to form data points;

[0021] Image Conversion: Convert the data points for each hour into grayscale images, generating 24 images for each category.

[0022] Furthermore, the method of constructing a demand curve function set based on multi-information fusion and generating image training samples includes:

[0023] Construction of demand curve function set: Logarithmic demand function and exponential demand function are selected as the basis, and parameters are adjusted through experiments to balance stability and diversity;

[0024] Data point generation: Randomly select demand curve functions to generate data points to form image training samples;

[0025] Image training sample generation: Based on the demand curve function set and data generation function, a large number of data points are synthesized for model training.

[0026] Furthermore, the image training sample is used as input to train a nonlinear demand price elasticity evaluation model based on a deep learning method; the grayscale image is input into the trained model, a polynomial representation of the demand curve is extracted, and a polynomial set of demand price elasticity curves is output, including:

[0027] Model training: Use the generated image training samples to train the nonlinear demand price elasticity evaluation model based on the attention U-Net network architecture to learn the intrinsic relationship between electricity price and electricity consumption;

[0028] Pixel mapping and polynomial fitting: extract the polynomial representation of the demand curve from the input grayscale image through the trained model;

[0029] Price elasticity evaluation: The polynomial set of demand price elasticity curve is obtained through pixel mapping and polynomial fitting.

[0030] Furthermore, the distributed flexible load model is constructed based on the historical data and the polynomial set of the demand price elasticity curve, including:

[0031] Data division: divide the test data into working days and rest days, reflect the weather factors in the residents' electricity consumption behavior, and collect historical data of preset working days, which includes electricity prices and corresponding user electricity consumption;

[0032] Model construction: Using the obtained electricity consumption behavior classification results and nonlinear demand price elasticity evaluation model, a distributed flexible load model is established;

[0033] Recalculation and update: If the assessment date changes, the classification of residential electricity consumption behavior, the assessment of electricity demand price elasticity and the modeling of distributed flexible loads are re-performed.

[0034] A distributed flexible load modeling system for distribution network reliability assessment, comprising:

[0035] An electricity consumption behavior classification module is configured to classify the electricity consumption behavior of residential users based on parallel computing and a combined classification method, and convert the classification results into a grayscale image;

[0036] A training sample generation module is configured to construct a demand curve function set based on multi-information fusion and generate image training samples;

[0037] An evaluation model building module is configured to take the image training sample as input, train a nonlinear demand price elasticity evaluation model based on a deep learning method; input the grayscale image into the trained model, extract the polynomial representation of the demand curve, and output a polynomial set of demand price elasticity curves;

[0038] The load model building module is configured to build a distributed flexible load model based on historical data and the demand price elasticity curve polynomial set to evaluate the electricity consumption of a single user at a given electricity price as an input parameter for distribution network reliability evaluation.

[0039] Furthermore, the electricity usage behavior classification module includes:

[0040] A data preprocessing unit is configured to normalize the collected residential electricity consumption behavior data, perform dimension reduction using a principal component analysis algorithm, and evenly distribute the processed data into each data block;

[0041] The parallel computing unit is configured to calculate the electricity consumption behavior samples in each data block, recursively calculate using three indicators, namely, cumulative proximity, eccentricity and normalized eccentricity, and determine the category of the new electricity consumption behavior samples by using Chebyshev inequality;

[0042] A category merging unit is configured to merge the categories collected from each data block by finding the nearest neighbor category and merging when the merging condition is met, and updating the parameters of the merged category;

[0043] a data integration unit configured to pair residential electricity consumption in each category with corresponding 24-hour electricity price data to form a data point;

[0044] The image conversion unit is configured to convert each hour’s data points into grayscale images, generating 24 images for each class.

[0045] Furthermore, the training sample generation module includes:

[0046] a demand curve function set building unit, configured to select a logarithmic demand function and an exponential demand function as a basis, and adjust parameters through experiments to balance stability and diversity;

[0047] A data point generating unit is configured to randomly select a demand curve function to generate data points to form an image training sample;

[0048] The image training sample generating unit is configured to synthesize a large number of data points for model training based on the demand curve function set and the data generating function.

[0049] Furthermore, the evaluation model building module includes:

[0050] A model training unit is configured to use the generated image training samples to train a nonlinear demand price elasticity evaluation model based on an attention U-Net network architecture to learn the intrinsic relationship between electricity price and electricity consumption;

[0051] a pixel mapping and polynomial fitting unit, configured to extract a polynomial representation of a demand curve from an input grayscale image through a trained model;

[0052] The price elasticity evaluation unit is configured to obtain a polynomial set of demand price elasticity curves through pixel mapping and polynomial fitting.

[0053] Furthermore, the load model building module includes:

[0054] A data division unit is configured to divide the test data into working days and rest days, and reflect the weather factor in the electricity consumption behavior of residents, and collect historical data of preset working days, wherein the historical data includes electricity prices and corresponding user electricity consumption;

[0055] A model building unit is configured to build a distributed flexible load model using the obtained electricity consumption behavior classification results and the nonlinear demand price elasticity evaluation model;

[0056] The recalculation and updating unit is configured to reclassify the electricity consumption behavior of residential users, evaluate the price elasticity of electricity demand, and model distributed flexible loads when the evaluation date changes.

[0057] The beneficial effects of the present invention are:

[0058] The present invention effectively improves the classification adaptability of residential users' electricity consumption patterns by introducing the parallel computing and merging (PC&M) classification method, without over-reliance on accurate household appliance data, reducing the difficulty of data collection and the risk of privacy leakage. At the same time, the multi-information fusion technology is used to generate training samples, and the nonlinear PED model established by the deep learning method can more accurately capture the nonlinear relationship between electricity demand and price, thereby improving the prediction accuracy. In addition, the established distributed flexible load model can be applied to models such as personalized dynamic pricing and distribution network reliability assessment, providing strong support for the optimized operation and sustainable development of the power system. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 This is a flow chart of a distributed flexible load modeling method for distribution network reliability assessment according to Example 1 of the present invention.

[0060] Figure 2Schematic diagram of the PC&M classification method in Example 2 of the present invention;

[0061] Figure 3 A visualization process diagram of the classification results in Example 2 of the present invention;

[0062] Figure 4 A flow chart for generating training samples in Embodiment 2 of the present invention;

[0063] Figure 5 This is a structural diagram of the Attention U-net in Example 2 of the present invention;

[0064] Figure 6 This is one of the schematic diagrams of clustering results of the PC&M classification method in Example 2 of the present invention;

[0065] Figure 7 This is the second schematic diagram of clustering results of the PC&M classification method in Example 2 of the present invention;

[0066] Figure 8 This is the third schematic diagram of clustering results of the PC&M classification method in Example 2 of the present invention;

[0067] Fig. 9 This is a schematic diagram of the distributed flexible load evaluation results in Example 2 of the present invention. DETAILED DESCRIPTION

[0068] In order to have a clearer understanding of the technical features, purposes and effects of the present invention, the specific implementation methods of the present invention are now described. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention, that is, the embodiments described are only part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making creative work are within the scope of protection of the present invention.

[0069] Example 1

[0070] This embodiment provides a distributed flexible load modeling method for distribution network reliability assessment. Figure 1 As shown, including:

[0071] Based on parallel computing and combined classification methods, residential users' electricity consumption behaviors are classified and the classification results are converted into grayscale images.

[0072] Construct a demand curve function set based on multi-information fusion and generate image training samples;

[0073] Taking the image training sample as input, training a nonlinear demand price elasticity evaluation model based on a deep learning method; inputting the grayscale image into the trained model, extracting the polynomial representation of the demand curve, and outputting a polynomial set of demand price elasticity curves;

[0074] Based on historical data and the polynomial set of demand price elasticity curves, a distributed flexible load model is constructed to evaluate the electricity consumption of a single user at a given electricity price, which serves as an input parameter for distribution network reliability evaluation.

[0075] In this embodiment, parallel computing and combined classification methods are used to classify user electricity consumption patterns without assumptions, thereby improving the adaptability of classification; multi-information fusion technology is used to generate training samples, and a nonlinear PED model is established in combination with a deep learning method and used for distributed flexible load modeling, which can be applied to personalized dynamic pricing, distribution network reliability assessment and other models.

[0076] Specifically, the distributed flexible load modeling method of this embodiment can be implemented by the following steps:

[0077] Step 1: Classification method of residential electricity consumption behavior based on PC&M

[0078] Step 1.1: Data preprocessing module. The collected residential electricity consumption behavior data is first normalized and then dimensionally reduced using the principal component analysis algorithm. The processed data is evenly distributed into each data block.

[0079] Step 1.2: Parallel calculation module. The processed electricity consumption behavior samples are calculated separately in each data block. Through the PC&M framework, the three indicators of cumulative proximity, eccentricity and standardized eccentricity are used for recursive calculation, the cumulative proximity and eccentricity of each sample are updated, and the parameters are recursively updated through the Euclidean distance. The Chebyshev inequality is used to determine whether the new electricity consumption behavior sample belongs to a certain category. If the conditions are met, the sample is classified into the corresponding category, otherwise a new category is created.

[0080] Step 1.3: Category merging module. The categories collected from each processor are merged, starting from the smallest category, and finding its nearest neighbor category. When the merging condition is met, the two categories are merged and the merging operation continues. The merged category updates its parameters.

[0081] Step 2: Visualization process of electricity consumption behavior classification results

[0082] Step 2.1: Data integration. In each category, residential electricity consumption is paired with the corresponding 24-hour electricity price data to form a large number of (price, electricity consumption) data points.

[0083] Step 2.2: Image conversion. Convert each hour’s (price, electricity usage) data point into a grayscale image, generating 24 images for each category.

[0084] Step 2.3: Image input model. These grayscale images are used as input to the neural network trained in step 4 to obtain the complex relationship between user price and electricity consumption.

[0085] Step 3: Generation of training samples based on multi-information fusion

[0086] Step 3.1: Construct a set of demand curve functions. Based on prior knowledge, construct a demand curve function with a monotonically decreasing shape. Select logarithmic demand function and exponential demand function as the basis, and adjust their parameters through experiments to balance stability and diversity.

[0087] Step 3.2: Data point generation. A demand curve function is randomly selected and data points are generated based on the curve. After visualization, these data points form a pair of image samples.

[0088] Step 3.3: Image training sample generation. Based on the demand curve function set and the data generation function, a large number of data points are synthesized to overcome the defect of relying on historical data in traditional methods. The synthesized training samples can cover a variety of scenarios.

[0089] Step 4: PED evaluation model based on Attention U-Net

[0090] Step 4.1: Model training. Using the generated image samples, a deep learning model based on the Attention U-Net mechanism is trained to learn the intrinsic relationship between electricity price and electricity consumption. The model extracts feature information from the image through a convolutional neural network architecture and uses an attention gate to focus on the most important areas.

[0091] Step 4.2: Pixel mapping and polynomial fitting. The trained model can extract the polynomial representation of the demand curve from the input grayscale image. First, the pixel coordinates in the image are converted to historical data coordinates through pixel mapping, and then the polynomial fitting method is used to obtain the 24-hour demand curve.

[0092] Step 4.3: Price elasticity evaluation model. The images of each category and time obtained in step 2 are used as input to the trained model. The output RGB image will be converted into a grayscale image, and then the final 24-hour demand curve is obtained through pixel mapping and polynomial fitting steps. This process includes step-by-step mapping and fitting, and finally a polynomial set of demand price elasticity curves for each electricity consumption mode category is obtained.

[0093] Step 5: Distributed flexible load modeling.

[0094] Step 5.1: Divide the test data into working days and rest days, and reflect the weather factors in the residents' electricity consumption behavior. Collect historical data from the day d-1 to dD before the specific working day d, including electricity prices and corresponding user electricity consumption. Using the user electricity consumption behavior classification method and electricity demand price elasticity evaluation model mentioned in the previous steps, obtain the 24-hour demand price elasticity curve polynomial set for each electricity consumption mode category, so as to establish a distributed flexible load model and use it to evaluate the electricity consumption of a single user at a given electricity price as an input parameter for distribution network reliability evaluation.

[0095] Step 5.2: Recalculation and update. If the evaluation date changes, it is necessary to reclassify the residential user's electricity consumption behavior, evaluate the price elasticity of electricity demand for each category, and model the distributed flexible load. Since the trained deep learning model does not need to be retrained, it will not cause additional computational burden.

[0096] Example 2

[0097] This embodiment is based on embodiment 1:

[0098] This embodiment provides a distributed flexible load modeling method for distribution network reliability assessment, including the following steps:

[0099] Step 1: Parallel Computation and Merging (PC&M) classification method based on Typicality and Centrifugality Data Analysis (TEDA). As a data analysis framework, TEDA only relies on empirical data samples and does not contain any prior assumptions.

[0100] Step 1.1: Under this framework, three key indicators are proposed as follows.

[0101] Cumulative proximity π n (x i ) represents the sum of the distances between a specific data sample and other data samples in the group. i The cumulative proximity of is expressed as formula (1):

[0102]

[0103] Among them, d 2 (x i ,x j ) represents x i and x j The distance measure between .

[0104] The eccentricity and normalized eccentricity are obtained by normalizing the cumulative proximity, and are expressed as formulas (2) and (3) respectively:

[0105]

[0106] in, is the average cumulative proximity.

[0107] Step 1.2: For the classification task of residential users' electricity consumption behavior, traditional methods usually require pre-defining the number and center of categories, which is subjective. However, the method in this embodiment can avoid this problem, making the number and center of categories completely dependent on the original characteristics of the historical electricity consumption pattern data. The framework of the PC&M classification method is as follows Figure 2 As shown in Figure 2, this method consists of three independent modules, which will be described in detail below.

[0108] Step 1.3: In the data processing module, all the collected residential electricity consumption behavior data are first standardized, then the principal component analysis algorithm is used for dimensionality reduction, and finally the processed data are evenly placed in each data block.

[0109] Step 1.4: In the parallel computing module, the processed electricity consumption behavior samples are calculated on the corresponding processors in each data block. According to the single processing characteristics of this method, it can process a large amount of residential electricity consumption behavior data and perform recursive calculations.

[0110] Step 1.4.1: The main indicators in TEDA: cumulative proximity, eccentricity and normalized eccentricity are recursively updated by using the Euclidean distance. New data sample x n The cumulative proximity of can be recursively calculated using formulas (4)-(6):

[0111] π n (x n )=n(||x n -μ n || 2 +X n -||μ n || 2 )

[0112]

[0113] Where μ represents the local mean of the recursive update, i.e., the center of each category; X represents the scalar product of the recursive update.

[0114] Step 1.4.2: The sum of the accumulated proximity can be updated by formula (7):

[0115]

[0116] Step 1.4.3: Substituting formulas (6) and (7) into formula (1), we can obtain the expression of standardized eccentricity, as shown in formula (8):

[0117]

[0118] Step 1.4.4: Chebyshev inequality is used as a criterion for determining whether a new user's electricity consumption behavior sample belongs to a specific category. This criterion is based on the standardized eccentricity and is restated as formula (9):

[0119]

[0120] Here, κ means that the distance between the new data sample and the local mean of a specific category is greater than κ times the standard deviation.

[0121] Step 1.4.5: The sum of the eccentricities of all samples in each category is equal to 2β, where β represents the total number of electricity consumption behaviors in the category. Therefore, the average standardized eccentricity of each category is 2. In order to balance the sensitivity to abnormal data and the tolerance to internal variance, the value of κ in this study is set to 2. According to formula (9), the association condition between a new electricity consumption behavior sample and a category can be expressed as formula (10):

[0122]

[0123] Step 1.4.6: If the new electricity consumption behavior sample meets the conditions of multiple categories, its category will be determined by formula (11):

[0124]

[0125] Among them, k * represents the selected category, and K represents the number of existing categories in the processor. n After the category to which it should belong is determined, the relevant parameters of the category are updated according to formula (12).

[0126]

[0127] Step 1.4.7: If a sample does not meet the criteria of any category in formula (12), the formation of a new category is initiated.

[0128] Step 1.5: In the category merging module, the categories from each processor are gathered together. Then, starting from the smallest category, the nearest neighbor category is searched. Once the conditions of formula (13) are met, the two categories are merged into a new category and a new merge operation is started.

[0129]

[0130] where μ i and μ j They represent the means of the i-th and j-th categories respectively. Calculated by formula (14):

[0131]

[0132] After the merging process, the parameters of the new category are updated according to formula (15):

[0133]

[0134] Step 2: Visualization of electricity consumption behavior classification results and corresponding electricity prices. In this example, direct analysis of the relationship between electricity consumption and individual resident prices shows significant variability, which makes it challenging to accurately assess the price elasticity of specific users in different time intervals. This situation prompted us to classify electricity consumption patterns (ECPs) in a wide range of end-user groups by adopting the designed PC&M classification method, and then conduct in-depth research on each category using machine vision technology.

[0135] Step 2.1: After obtaining the classification results, first merge the residential ECP in each category with the corresponding 24-hour price data and discretize it into data points consisting of price and consumption. Then, convert the data points corresponding to each hour into grayscale images, and finally obtain 24 images. As a demonstration of the detailed visualization process, category 1 is selected as a representative example, such as Figure 3 shown.

[0136] Step 2.2: Assume that category 1 contains β residential ECPs, each ECP is paired with its corresponding 24-hour dynamic pricing data, resulting in a total of 24β data points (p,q). These data points are then depicted as hourly grayscale images. Thus, for each category, a set of 24-hour images is obtained as input data for the neural network model in step 4.

[0137] Step 3: Generate training samples based on multi-information fusion. In order to develop a well-trained model, a large number of training samples must be used. However, it is difficult to obtain diverse training samples from historical data, i.e., image pairs (I x ,I DC ), is challenging. Moreover, extracting real data from the collected data is not feasible.

[0138] Therefore, this embodiment proposes a data generation function based on multi-information fusion, which is used to generate the data on the selected demand curve f DC A large number of data points (p, q) are generated around the demand curve, which is regarded as real data. Then, these data are visualized and combined into an image pair as a training sample. The process of generating training samples is as follows Figure 4 shown.

[0139] Step 3.1: From Figure 4 It can be seen that the first stage is to establish a set of demand curve functions that exhibit monotonically decreasing curvature based on prior knowledge. In order to ensure simplicity and wide applicability, logarithmic demand functions and exponential demand functions are selected to create real data sets, which are expressed as formula (16):

[0140] q log (p) = a log +b log ln(p)

[0141] q exp (p) = a exp exp(b exp p)

[0142] Among them, parameter a log ,b log ,a exp ,b exp The setting is determined through preliminary experiments to achieve the generation of f DC A balance between stability and diversity.

[0143] Step 3.2: Data point generation. The second stage is to randomly select a demand curve function and generate data points around the curve, expressed as formula (17):

[0144] (p,q)=g ζ (f DC )

[0145] Among them, f DC represents the real data function describing the demand curve, g ζ represents the data generating function and (p,q) represents the pairing of synthetic electricity consumption and price.

[0146] Step 3.3: Generate function g based on data ζ , can be based on various predefined f DC Synthesizing sufficiently rich training samples overcomes the shortcomings of traditional methods that rely on historical data. In order to observe and analyze the inherent graphical features of real collected data points, g ζ Defined as formula (18):

[0147] Normal Data=(x,δ*derivative(f DC (x)+f DC (x)

[0148] Sparse Outliers=(a,b)

[0149]

[0150] Where δ represents the deviation of normal data from f DC is proportional to the first derivative of (x).

[0151] Step 3.4: By repeatedly executing the above procedure, a large number of training samples can be obtained. Considering the time cost of model training, a total of 12,000 pairs of images were generated, which is considered sufficient to cover various scenes within the scope of this study.

[0152] Step 4: PED evaluation model based on Attention U-Net. This embodiment designs a data-driven deep learning model that performs the inverse process of data generation. The model is able to extract the true demand curve f from the simulated data points DC , and can be extended to real data consisting of electricity demand and price. The simplified form of the model is as follows:

[0153]

[0154] Step 4.1: In order to combine the local spatial relationship between data points (p, q) in image visualization, a modeling method based on machine vision is selected. The objective function of this method is shown in formula (20):

[0155]

[0156] in, is the inverse function, with a weight coefficient of ω; L is the loss function; I x Represents the image visualization of the data point (p,q); I DC is the demand curve f DC Image visualization.

[0157] Step 4.2: This embodiment proposes a framework based on the Attention U-net mechanism to solve the above modeling task. Attention U-net is a convolutional neural network widely used in the field of medical image analysis. It selectively focuses on the areas containing the most information in the input image by using attention gates. Attention U-net has a significant ability to retain low-level features and capture high-level features, making it a suitable choice for modeling the inverse function. Figure 5 The Attention U-net architecture used in this study is shown, which contains four basic layers: convolutional layer, maximum pooling layer, upsampling layer, and output layer. The numbers shown above the figure indicate the configuration of the feature map.

[0158] Can be obtained from Figure 5As can be seen in Figure 1, the Attention U-net architecture consists of two parts. In the encoder part, the feature size is reduced from 256×256 to 32×32 through four downsampling operations. In the decoder part, the original image resolution is restored through four upsampling operations. During each training process, each convolutional layer repeats three consecutive operations (3×3 convolution, normalization, and ReLU activation). The maximum pooling layer uses a 2×2 pooling operation for downsampling, while the upper convolution layer is upsampled through a 2×2 transposed convolution operation. The output layer maps the feature map to the RGB image space through a single convolution operation. Skip Connection allows the low-level feature map to be concatenated with the upsampled feature map, thereby recovering the spatial information lost during the downsampling operation. The gating signal extracted from the encoder feature map selectively emphasizes areas in the feature map based on its relevance to the task.

[0159] Step 4.3: The mean square error (MSE) is used as the loss function, as shown in formula (21):

[0160]

[0161] After model training, the Attention U-net-based model can accurately x Extract I DC The grayscale image obtained in the third part is input into the trained model to generate the RGB image I DC In order to DC Derive the demand curve f DC A pixel mapping and polynomial fitting module is proposed.

[0162] To clarify, the coordinates of the normalized historical data consist of two floating point numbers, representing the normalized price and the normalized electricity consumption, respectively, while the coordinates of the pixels are represented by two integers, representing the relative position of the pixel in the image, which are subsequently mapped to the coordinates of the normalized data.

[0163] The output RGB image is first converted to a grayscale image of size 256×256. Considering that the pixel coordinates of an image in Python range from 32 to 229, there are 198 coordinate pairs (x pixel ,y pixel ). It is worth noting that the grayscale value of each pixel ranges from 0 to 255, where 0 represents black and 255 represents white.

[0164] Step 4.4: For each column of fixed x-coordinates, identify the pixel with the smallest grayscale value and assign it as the y-coordinate. Subsequently, map this selected pixel to the coordinate of the historical data. The complete process of pixel mapping is represented by formula (22):

[0165]

[0166] Among them, l represents the integer of the position of the pixel in the column, (x datal ,y datal ) represents the mapped data coordinate pair.

[0167] Step 4.5: After completing the pixel mapping process, a total of 198 normalized data points are obtained for deriving f DC This process is completed through the polynomial fitting formula (23):

[0168]

[0169] Among them, λ represents the penalty term; a i represents the weight coefficient. After this process, the polynomial representation of the demand curve is obtained.

[0170] Step 4.6: By repeating the above operations, the 24-hour grayscale image is processed and a set of polynomials containing the hourly demand curve fDC for each category is constructed, as shown in (24):

[0171] F k ={f 1,k ,f 2,k ,...,f 24,k},k=1,2,...,K

[0172] Step 5: Distributed flexible load modeling. In order to mitigate the impact of seasonal and date changes on residential ECPs, the historical data used for testing in this embodiment will first be divided into working days and rest days according to each season. It is assumed that weather factors are reflected in consumers' electricity consumption behavior. In order to evaluate the electricity consumption of a specific consumer on a certain working day d, the daily ECPs and corresponding retail prices from working days d-1 to dD will be collected as test data.

[0173] Step 5.1: The principle of the dedicated packaging model and how to evaluate the daily electricity consumption of a single end user are shown below. Within the jurisdiction of the retailer, numerous residents are able to adjust their electricity demand under a given dynamic electricity price. By utilizing the proposed ECP classification method and the PED evaluation model based on multi-information fusion, a set of polynomials that delineate the 24-hour demand curve for each category is obtained. Combined with the ECP distribution percentage of a specific consumer in each category, a packaging model is designed to evaluate the electricity consumption on day d, as shown in formula (25), where the input is the 24-hour price vector:

[0174]

[0175] Where K represents the number of ECP categories to which a specific consumer belongs; D k Indicates category C k The number of ECPs in F k Category C k The set of polynomials .

[0176] Step 5.2: Recalculation and update. If the evaluation date changes, it is necessary to reclassify the residential user's electricity consumption behavior, evaluate the price elasticity of electricity demand for each category, and model the distributed flexible load. Since the trained deep learning model does not need to be retrained, it will not cause additional computational burden.

[0177] like Figure 6 to Figure 8 The figure shows the clustering results of the PC&M classification method, where the vertical axis represents the user's power consumption at 24 hours, that is, the user's power consumption pattern / behavior.

[0178] like Fig. 9 The result display of the distributed flexible load assessment model for five consecutive working days is shown, where the vertical axis represents the actual value of the user's power consumption at 24 hours; GRU, LSTM, and U-net represent comparison models, and Proposed represents the model proposed in this embodiment.

[0179] Example 3

[0180] This embodiment provides a distributed flexible load modeling system for distribution network reliability assessment, including:

[0181] An electricity consumption behavior classification module is configured to classify the electricity consumption behavior of residential users based on parallel computing and a combined classification method, and convert the classification results into a grayscale image;

[0182] A training sample generation module is configured to construct a demand curve function set based on multi-information fusion and generate image training samples;

[0183] An evaluation model building module is configured to take the image training sample as input, train a nonlinear demand price elasticity evaluation model based on a deep learning method; input the grayscale image into the trained model, extract the polynomial representation of the demand curve, and output a polynomial set of demand price elasticity curves;

[0184] The load model building module is configured to build a distributed flexible load model based on historical data and a polynomial set of demand price elasticity curves to evaluate the power consumption of a single user at a given electricity price, which serves as an input parameter for distribution network reliability assessment.

[0185] Preferably, the electricity usage behavior classification module includes:

[0186] A data preprocessing unit is configured to normalize the collected residential electricity consumption behavior data, perform dimension reduction using a principal component analysis algorithm, and evenly distribute the processed data into each data block;

[0187] The parallel computing unit is configured to calculate the electricity consumption behavior samples in each data block, recursively calculate using three indicators, namely, cumulative proximity, eccentricity and normalized eccentricity, and determine the category of the new electricity consumption behavior samples by using Chebyshev inequality;

[0188] A category merging unit is configured to merge the categories collected from each data block by finding the nearest neighbor category and merging when the merging condition is met, and updating the parameters of the merged category;

[0189] a data integration unit configured to pair residential electricity consumption in each category with corresponding 24-hour electricity price data to form a data point;

[0190] The image conversion unit is configured to convert each hour’s data points into grayscale images, generating 24 images for each class.

[0191] Preferably, the training sample generation module includes:

[0192] a demand curve function set building unit, configured to select a logarithmic demand function and an exponential demand function as a basis, and adjust parameters through experiments to balance stability and diversity;

[0193] A data point generating unit is configured to randomly select a demand curve function to generate data points to form an image training sample;

[0194] The image training sample generating unit is configured to synthesize a large number of data points for model training based on the demand curve function set and the data generating function.

[0195] Preferably, the evaluation model building module includes:

[0196] A model training unit is configured to use the generated image training samples to train a nonlinear demand price elasticity evaluation model based on an attention U-Net network architecture to learn the intrinsic relationship between electricity price and electricity consumption;

[0197] a pixel mapping and polynomial fitting unit, configured to extract a polynomial representation of a demand curve from an input grayscale image through a trained model;

[0198] The price elasticity evaluation unit is configured to obtain a polynomial set of demand price elasticity curves through pixel mapping and polynomial fitting.

[0199] Preferably, the load model building module includes:

[0200] A data division unit is configured to divide the test data into working days and rest days, and reflect the weather factor in the electricity consumption behavior of residents, and collect historical data of preset working days, the historical data including electricity prices and corresponding user electricity consumption;

[0201] A model building unit is configured to build a distributed flexible load model using the obtained electricity consumption behavior classification results and the nonlinear demand price elasticity evaluation model;

[0202] The recalculation and updating unit is configured to reclassify the electricity consumption behavior of residential users, evaluate the price elasticity of electricity demand, and model distributed flexible loads when the evaluation date changes.

[0203] Example 4

[0204] This embodiment is based on embodiment 1:

[0205] This embodiment provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the distributed flexible load modeling method for distribution network reliability assessment of Embodiment 1 is implemented. The computer program may be in source code form, object code form, executable file, or some intermediate form.

[0206] Example 5

[0207] This embodiment is based on embodiment 1:

[0208] This embodiment provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the distributed flexible load modeling method for distribution network reliability assessment of Embodiment 1. The computer program may be in source code form, object code form, executable file, or some intermediate form, etc. The storage medium includes: any entity or device capable of carrying computer program code, recording medium, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal, and software distribution medium, etc. It should be noted that the content contained in the storage medium may be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, the storage medium does not include electric carrier signals and telecommunication signals.

[0209] It should be noted that, for the aforementioned method embodiments, for the sake of simplicity of description, they are expressed as a series of action combinations, but those skilled in the art should be aware that the present application is not limited by the order of the actions described, because according to the present application, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the present application.

Claims

1. A distributed flexible load modeling method for distribution network reliability assessment, characterized in that: include: Based on parallel computing and combined classification methods, residential users' electricity consumption behaviors are classified and the classification results are converted into grayscale images. Construct a demand curve function set based on multi-information fusion and generate image training samples; Taking the image training sample as input, training a nonlinear demand price elasticity evaluation model based on a deep learning method; inputting the grayscale image into the trained model, extracting the polynomial representation of the demand curve, and outputting a polynomial set of demand price elasticity curves; Based on historical data and the polynomial set of demand price elasticity curves, a distributed flexible load model is constructed to evaluate the electricity consumption of a single user at a given electricity price, which serves as an input parameter for distribution network reliability evaluation.

2. The distributed flexible load modeling method for distribution network reliability assessment according to claim 1 is characterized in that: The method of classifying the electricity consumption behavior of residential users based on parallel computing and combined classification, and converting the classification results into grayscale images, includes: Data preprocessing: normalize the collected residential electricity consumption behavior data, use the principal component analysis algorithm to reduce the dimension, and evenly distribute the processed data into each data block; Parallel calculation: Calculate the electricity consumption behavior samples in each data block, use the three indicators of cumulative proximity, eccentricity and standardized eccentricity for recursive calculation, and use Chebyshev inequality to determine the category of the new electricity consumption behavior samples; Category merging: Merge the categories collected from each data block by finding the nearest neighbor category and merging when the merging conditions are met, and update the parameters of the merged category; Data integration: Pair residential electricity consumption in each category with the corresponding 24-hour electricity price data to form data points; Image Conversion: Convert the data points for each hour into grayscale images, generating 24 images for each category.

3. The distributed flexible load modeling method for distribution network reliability assessment according to claim 1 is characterized in that: The method of constructing a demand curve function set based on multi-information fusion and generating image training samples includes: Construction of demand curve function set: Logarithmic demand function and exponential demand function are selected as the basis, and parameters are adjusted through experiments to balance stability and diversity; Data point generation: Randomly select demand curve functions to generate data points to form image training samples; Image training sample generation: Based on the demand curve function set and data generation function, a large number of data points are synthesized for model training.

4. The distributed flexible load modeling method for distribution network reliability assessment according to claim 1 is characterized in that: The image training sample is used as input to train a nonlinear demand price elasticity evaluation model based on a deep learning method; the grayscale image is input into the trained model, a polynomial representation of the demand curve is extracted, and a polynomial set of demand price elasticity curves is output, including: Model training: Use the generated image training samples to train the nonlinear demand price elasticity evaluation model based on the attention U-Net network architecture to learn the intrinsic relationship between electricity price and electricity consumption; Pixel mapping and polynomial fitting: extract the polynomial representation of the demand curve from the input grayscale image through the trained model; Price elasticity evaluation: The polynomial set of demand price elasticity curve is obtained through pixel mapping and polynomial fitting.

5. The distributed flexible load modeling method for distribution network reliability assessment according to claim 1 is characterized in that: The constructing of a distributed flexible load model based on historical data and the demand price elasticity curve polynomial set includes: Data division: divide the test data into working days and rest days, reflect the weather factors in the residents' electricity consumption behavior, and collect historical data of preset working days, which includes electricity prices and corresponding user electricity consumption; Model construction: Using the obtained electricity consumption behavior classification results and nonlinear demand price elasticity evaluation model, a distributed flexible load model is established; Recalculation and update: If the assessment date changes, the classification of residential electricity consumption behavior, the assessment of electricity demand price elasticity and the modeling of distributed flexible loads are re-performed.

6. A distributed flexible load modeling system for distribution network reliability assessment, characterized in that: include: An electricity consumption behavior classification module is configured to classify the electricity consumption behavior of residential users based on parallel computing and a combined classification method, and convert the classification results into a grayscale image; A training sample generation module is configured to construct a demand curve function set based on multi-information fusion and generate image training samples; An evaluation model building module is configured to take the image training sample as input, train a nonlinear demand price elasticity evaluation model based on a deep learning method; input the grayscale image into the trained model, extract the polynomial representation of the demand curve, and output a polynomial set of demand price elasticity curves; The load model building module is configured to build a distributed flexible load model based on historical data and the demand price elasticity curve polynomial set to evaluate the electricity consumption of a single user at a given electricity price as an input parameter for distribution network reliability evaluation.

7. A distributed flexible load modeling system for distribution network reliability assessment according to claim 6, characterized in that: The electricity usage behavior classification module includes: A data preprocessing unit is configured to normalize the collected residential electricity consumption behavior data, perform dimension reduction using a principal component analysis algorithm, and evenly distribute the processed data into each data block; The parallel computing unit is configured to calculate the electricity consumption behavior samples in each data block, recursively calculate using three indicators, namely, cumulative proximity, eccentricity and normalized eccentricity, and determine the category of the new electricity consumption behavior samples by using Chebyshev inequality; A category merging unit is configured to merge the categories collected from each data block by finding the nearest neighbor category and merging when the merging condition is met, and updating the parameters of the merged category; a data integration unit configured to pair residential electricity consumption in each category with corresponding 24-hour electricity price data to form a data point; The image conversion unit is configured to convert each hour’s data points into grayscale images, generating 24 images for each class.

8. A distributed flexible load modeling system for distribution network reliability assessment according to claim 6, characterized in that: The training sample generation module comprises: a demand curve function set building unit, configured to select a logarithmic demand function and an exponential demand function as a basis, and adjust parameters through experiments to balance stability and diversity; A data point generating unit is configured to randomly select a demand curve function to generate data points to form an image training sample; The image training sample generating unit is configured to synthesize a large number of data points for model training based on the demand curve function set and the data generating function.

9. A distributed flexible load modeling system for distribution network reliability assessment according to claim 6, characterized in that: The evaluation model building module includes: A model training unit is configured to use the generated image training samples to train a nonlinear demand price elasticity evaluation model based on an attention U-Net network architecture to learn the intrinsic relationship between electricity price and electricity consumption; a pixel mapping and polynomial fitting unit, configured to extract a polynomial representation of a demand curve from an input grayscale image through a trained model; The price elasticity evaluation unit is configured to obtain a polynomial set of demand price elasticity curves through pixel mapping and polynomial fitting.

10. A distributed flexible load modeling system for distribution network reliability assessment according to claim 6, characterized in that: The load model building module includes: A data division unit is configured to divide the test data into working days and rest days, and reflect the weather factor in the electricity consumption behavior of residents, and collect historical data of preset working days, wherein the historical data includes electricity prices and corresponding user electricity consumption; A model building unit is configured to build a distributed flexible load model using the obtained electricity consumption behavior classification results and the nonlinear demand price elasticity evaluation model; The recalculation and updating unit is configured to reclassify the electricity consumption behavior of residential users, evaluate the price elasticity of electricity demand, and model distributed flexible loads when the evaluation date changes.

Images