Flexible load identification method based on load area characteristic enhancement

Abstract

The invention discloses a flexible load identification method based on load area characteristic enhancement, and aims to solve the problems that the traditional flexible load identification method is seriously insufficient in robustness and generalization ability, poor in dynamic adaptability and over-coarse in classification granularity when facing complex and changeable power consumption behaviors, so that the flexible load identification accuracy is low, and the flexible load identification efficiency is low. And the actual requirements of a modern power grid on efficient and accurate development and scheduling of massive flexible load resources are difficult to meet. According to the method, a load area characteristic enhancement expression model and a time sequence characteristic description model with an area label are trained and constructed by using a load basic information set and an urban function area point information set, and the association between the load and the urban function interval is deeply described. And training a decision-making model by using the decision-making table, and judging whether the load is a flexible load or not by the decision-making model according to a splicing vector of an output result of the load region characteristic enhancement expression model and the time sequence characteristic description model with the region label. The invention belongs to the field of power system management and analysis.

Description

Technical Field

[0001] This invention relates to the field of power system management and analysis, and specifically to a flexible load identification method based on enhanced load area characteristics. Background Technology

[0002] In building new power systems, flexible loads, as key regulating resources, play a crucial role in enhancing the grid's capacity to absorb new energy sources and ensuring the safe and stable operation of the system. With the advancement of "dual-carbon" goals, the influx of massive flexible loads such as electric vehicles, smart homes, and energy storage devices has brought unprecedented opportunities and challenges to the power grid. However, how to accurately and efficiently identify and utilize these dispersed and diverse flexible resources has become a technical bottleneck restricting their large-scale application. Therefore, developing an advanced flexible load identification method is not only an urgent need for the development of power technology but also a major issue serving national energy transition and sustainable social development, possessing extremely high research value and broad application prospects.

[0003] Traditional flexible load identification methods are mainly divided into pattern recognition methods based on load curve features and manual classification methods based on prior knowledge.

[0004] The first category of methods is mostly based on deep learning, such as Long Short-Term Memory (LSTM) networks. It attempts to distinguish between flexible and rigid loads by learning the inherent patterns in load time-series data. The basic logic is that rigid loads (such as industrial production lines) typically exhibit strong periodicity and regularity, while flexible loads (such as residential air conditioning and electric vehicle charging) show greater randomness and volatility. However, this method has encountered insurmountable obstacles in practice. The core dilemma lies in the fact that a key criterion for judging flexible loads is precisely their "lack of regularity" or "weak regularity." For neural networks, learning "lack of regularity" is inherently paradoxical. Even with a diverse range of sample curves used for network model training, it's difficult to exhaust all true forms of "lack of regularity." This makes the network model highly susceptible to "overfitting," meaning it only memorizes specific curve shapes from the training samples, and its recognition accuracy drops significantly when encountering new, unseen load curves. In other words, the model learns the "irregularity in the samples" rather than the abstract concept of true "irregularity". This makes the method severely lacking in robustness and generalization ability when faced with the complex and ever-changing electricity consumption behavior in the real world, ultimately leading to recognition failure.

[0005] The second type of method relies on industry expert experience or established standards to classify loads. For example, it labels users based on their industry (e.g., commercial, industrial, residential) and main electrical equipment types (e.g., air conditioners, electric stoves, charging stations). While this method is simple and intuitive, its classification criteria are highly subjective in the context of modern power grid management and increasing integration with urban functions. Different experts or regions may have different classification bases, lacking a unified and objective standard. Secondly, the classification granularity is too coarse, categorizing users in a region or industry into the same type without considering significant internal differences. For example, even among commercial users, a shopping mall in a prime business district and a convenience store in a transportation hub have drastically different load characteristics and regulation potential. Finally, this method lacks dynamic adaptability and cannot keep up with changes in urban development and user behavior. The functional positioning of a region may change with urban planning, and users' electricity consumption habits are not static; this static label-based classification method clearly cannot capture these dynamic characteristics.

[0006] In summary, traditional methods have significant shortcomings in terms of refinement, dynamism, and intelligence, making it difficult to meet the practical needs of modern power grids for the efficient and precise development and scheduling of massive flexible load resources. Therefore, there is an urgent need for a flexible load identification method that is easy to implement and has high accuracy. Summary of the Invention

[0007] To address the problem that traditional flexible load identification methods suffer from insufficient robustness and generalization ability, poor dynamic adaptability, and excessively coarse classification granularity when facing complex and ever-changing electricity consumption behaviors, resulting in low accuracy in flexible load identification and failing to meet the actual needs of modern power grids for efficient and precise development and scheduling of massive flexible load resources, this invention proposes a flexible load identification method based on enhanced load area characteristics.

[0008] The technical solution adopted in this invention is:

[0009] It includes the following steps:

[0010] S1. Input the set of basic load information of multiple electricity accounts in a city and the set of information of urban functional areas;

[0011] S2. Construct a load area characteristic enhancement expression model. Input the load basic information set and the urban functional area point information set into the load area characteristic enhancement expression model for training, and output the feature vector to obtain the trained load area characteristic enhancement expression model.

[0012] S3. Construct a time series feature description model with regional labels. Input the basic load information set and the urban functional area point information set into the time series feature description model with regional labels for training. Output feature vectors to obtain the trained time series feature description model with regional labels.

[0013] S4. Based on S1-S3, construct a decision table;

[0014] S5. Construct a decision model, train the decision model using a decision table, and obtain a trained decision model.

[0015] S6. Input load information to be identified, which includes load time series information and load location. Input the load location into the trained load region characteristic enhancement expression model and output the feature vector UnknownLACEEMVector. Input the load time series information and load location into the trained time series feature description model with region labels and output the feature vector UnknownRLTSFDVector. Concatenate the feature vectors UnknownLACEEMVector and UnknownRLTSFDVector to obtain the input vector UnknownInputVector. Input the input vector UnknownInputVector into the trained decision model and output the probability that the load is a flexible load. If the probability is greater than 0.5, the load is determined to be a flexible load; otherwise, the load is determined to be a non-flexible load.

[0016] The beneficial effects of this invention are as follows:

[0017] This invention introduces a load area characteristic enhancement expression model and a time series feature description model with area labels to encode load time series data and spatial location data, deeply describe the relationship between load and urban functional areas, and then construct a decision model to identify flexible loads.

[0018] One advantage of this invention is that it decouples specific electricity load time series data from specific sequence features, avoiding the predicament of learning from "irregular" patterns. Load flexibility essentially stems from human work and lifestyle habits, which are closely linked to people's activity patterns in urban spaces. For example, peak loads in commercial areas coincide with shopping and entertainment activities, while load patterns in residential areas reflect residents' daily routines. Therefore, this invention introduces the urban functional environment in which the load is located, establishing a more stable and interpretable flexible load identification model with dual data perspectives: load curves and urban areas. This invention improves the dynamic adaptability and identification accuracy of the flexible load method based on correlation when facing complex and ever-changing electricity consumption behaviors. Attached Figure Description

[0019] Figure 1 This is a flowchart of the present invention; Detailed Implementation

[0020] Specific implementation method one: Combining Figure 1 This embodiment describes a flexible load identification method based on enhanced load area characteristics, which includes the following steps:

[0021] S1. Input a set of basic load information (FHBaseSet) for multiple electricity accounts in a city. The set includes N FHBase entries with flexible labels. Each entry includes load time series information (LoadTimeSeries), load location coordinates (Position), and a flexible load label (FlexibilityLabel). The load time series information records the load changes of the electricity account over a period of time. The load location coordinates include the longitude and latitude coordinates of the electricity account. The flexible load label uses a Boolean value or an integer, where 1 represents a flexible load and 0 represents a non-flexible load.

[0022] Input the City Functional Point Set (CFPSet). This set includes multiple CFPs (City Functional Point Sets), each containing a CFP location (CFPPosition) and a CFP type (CFPType). The location is the representative latitude and longitude coordinates of that functional area. Area types are categorized as Residential, Commercial, Administrative, and Leisure.

[0023] The main purpose of this step is to prepare the foundational dataset for model training and decision-making, including a large amount of load sample information with flexible labels and geospatial information of urban functional areas.

[0024] S2. Construct a load area characteristic enhancement expression model. Input the load basic information set FHBaseSet and the urban functional area point information set CFPSet into the load area characteristic enhancement expression model for training, and output the feature vector to obtain the trained load area characteristic enhancement expression model. The specific process is as follows:

[0025] S21. Construct the Load Area Characteristic Enhancement Expression Module (LACEEM). The input to this model is a latitude and longitude coordinate position, and the output is a 40-dimensional feature vector (LACEEMVector). The LACEEM includes sub-modules for quantifying spatial relationships in residential, commercial, administrative, and leisure areas. It quantifies the spatial relationship between each load location point and the four core functional areas of the city, encoding it as a fixed-dimensional feature vector for subsequent calculation and processing.

[0026] S22. Select all area points of residential type from the urban functional area point information set CFPSet and form a residential point set ResidentialSet. Input the residential point set into the residential spatial relationship quantification submodule, calculate the distance from each load position to the 10 nearest area points in the residential point set, sort them according to the distance, and encode the spatial relationship between each input load position and the residential point set into a 10-dimensional feature vector ResidentialVector.

[0027] S23. Select all area points of the area type as commercial area from the urban functional area point information set CFPSet, and form a commercial area point set CommercialSet. Input the commercial area point set into the commercial area spatial relationship quantification submodule, calculate the mean, variance, skewness and kurtosis of the distance between each load position Position and all area points in the commercial area point set, and then reduce the dimension to 10 dimensions through principal component analysis (PCA) to generate a 10-dimensional feature vector CommercialVector.

[0028] S24. From the urban functional area point information set CFPSet, select all area points whose area type is administrative region and form an administrative region point set AdministrativeSet. Input the administrative region point set into the administrative region spatial relationship quantification submodule and calculate the geospatial distance between each load location Position and each area point in the administrative region point set. According to the law of gravity The idea is to measure the gravitational force of each region point on the Position. Defined as After calculating the gravitational values ​​between Position and all regional points in the administrative district point set, the 10 regional points with the highest gravitational values ​​(i.e., the 10 closest regional points) are selected. These 10 gravitational values ​​are then sorted from largest to smallest. It is output directly as a 10-dimensional feature vector, AdministrativeVector.

[0029] S25. Select all area points of the leisure area type from the urban functional area point information set CFPSet, and form a leisure area point set LeisureSet. Input the leisure area point set into the leisure area spatial relationship quantification submodule. In order to measure the cluster density of leisure areas at multiple spatial scales, this invention presets 10 bandwidth values ​​that can cover the range from "neighborhood level" to "city level" and grow exponentially. Let represent 100 meters, 250 meters, 500 meters, 1 kilometer, 2 kilometers, 4 kilometers, 8 kilometers, 16 kilometers, 32 kilometers, and 64 kilometers, respectively. Then, based on all area points in the recreation area point set, a standard Gaussian kernel function is selected, and the 10 bandwidth values ​​are successively substituted into the Gaussian kernel function to calculate the kernel density estimate for each load location, resulting in 10 kernel density estimates at different spatial scales. These 10 kernel density estimates are then... Arranging them in chronological order yields a 10-dimensional feature vector, LeisureVector. This feature vector comprehensively describes the concentration of a location's "leisure attributes" from micro to macro levels.

[0030] S26. Concatenate the eigenvectors ResidentialVector, CommercialVector, AdministrativeVector, and LeisureVector in sequence to output a 40-dimensional eigenvector LACEEMVector.

[0031] S3. Construct a time-series feature description model with regional labels. Input the load basic information set FHBaseSet and the urban functional area point information set CFPSet into the time-series feature description model with regional labels for training, and output feature vectors to obtain the trained time-series feature description model with regional labels. The specific process is as follows:

[0032] S31. Construct a region-labeled time series feature description model (RL-TSFD). The input to this model is the load time series information (LoadTimeSeries) and its corresponding latitude and longitude information (Position). The output is a 40-dimensional feature vector (RLTSFDVector). The region-labeled time series feature description model includes sub-modules for extracting labeled features from residential areas, commercial areas, administrative areas, and leisure areas. This model not only analyzes the characteristics of the load time series itself but also incorporates the regional functional tendencies of its location, thus generating a more distinctive feature vector.

[0033] S32. Calculate the average distance from each load location (Position) to all area points in the residential point set. .

[0034] Calculate the average distance from each load position to all area points in the business district point set. .

[0035] Calculate the average distance from each load position to all area points in the administrative district point set. .

[0036] Calculate the average distance from each load position to all area points in the recreation area point set. .

[0037] S33. Calculate the average distance from each load location (Position) to all area points in the Urban Functional Area Point Information Set (CFPSet). .

[0038] S34. Calculate the residential area index based on S32 and S33 respectively. Commercial District Index Administrative District Index and leisure area index :

[0039] (1)

[0040] (2)

[0041] (3)

[0042] (4)

[0043] This step aims to quantify the correlation between load location and the four functional zones: residential, commercial, administrative, and leisure.

[0044] S35. Input the load time series information and residential area index of each load into the residential area labeled feature extraction submodule, and use Fourier transform to extract the main frequency components of the load time series information and the residential area index. As an additional channel or weight, it is fed along with the main frequency components into a small autoencoder SelfCode, outputting a 10-dimensional feature vector. .

[0045] SelfCode is a neural network consisting of three fully connected layers. The input layer receives the fused feature vector (a 21-dimensional vector formed by fusing 20 Fourier principal frequency components and one residential (functional) index). The middle layer is a fully connected layer with 10 neurons; its output is the final 10-dimensional feature vector. To maintain information integrity, this layer uses a linear activation function. The output layer is a fully connected layer with 21 neurons, whose goal is to reconstruct the original input vector as accurately as possible. The goal of SelfCode training is to minimize the error (e.g., mean squared error) between the vector reconstructed by the output layer and the original vector of the input layer. After training, this invention only uses its "encoder" part (i.e., from the input layer to the intermediate layers) to generate a 10-dimensional feature vector.

[0046] S36. Input the load time series information and commercial area index of each load into the commercial area label feature extraction submodule, and extract features based on the commercial area index. The load time series is encoded to generate feature vectors related to the business district. The specific method is as follows:

[0047] First, a Fourier transform is applied to the input load time series to extract its main frequency components. Then, the business district index is... This is used as an independent additional channel or weighting coefficient to fuse with the extracted main frequency components. Finally, the fused data is input into a pre-defined small autoencoder, SelfCode, where the network's encoding process extracts and generates a 10-dimensional feature vector. .

[0048] S37. Input the load time series information and administrative region index of each load into the administrative region label feature extraction submodule, and extract the feature based on the administrative region index. The load time series is encoded to generate feature vectors related to administrative regions. The specific method is as follows:

[0049] First, a Fourier transform is applied to the input load time series to extract its main frequency components. Then, the administrative region index is... This is used as an independent additional channel or weighting coefficient to fuse with the extracted main frequency components. Finally, the fused data is input into a pre-defined small autoencoder, SelfCode, where the network's encoding process extracts and generates a 10-dimensional feature vector. .

[0050] S38. Input the load time series information and leisure area index of each load into the leisure area labeled feature extraction submodule, and extract the feature based on the leisure area index. The load time series is encoded to generate feature vectors related to the leisure area. The specific method is as follows:

[0051] First, a Fourier transform is applied to the input load time series to extract its main frequency components. Then, the leisure area index is... This is used as an independent additional channel or weighting coefficient to fuse with the extracted main frequency components. Finally, the fused data is input into a pre-defined small autoencoder, SelfCode, where the network's encoding process extracts and generates a 10-dimensional feature vector. .

[0052] S39. The feature vector eigenvectors eigenvectors and eigenvectors By concatenating the features in sequence, a 40-dimensional feature vector RLTSFDVector is output, resulting in a trained time series feature description model with region labels.

[0053] S4. Based on S1-S3, construct the decision table FHDecisionTable. This step iteratively processes all known basic load information, utilizes the features generated by the model constructed in S2 and S3, and combines them with flexible labels containing known basic load information to construct a structured decision table for training a machine learning model. The specific process is as follows:

[0054] S41. Initialize an empty decision table FHDecisionTable.

[0055] S42. Initialize a loop variable FHCounter to 0.

[0056] S43. Retrieve the FHCounter-th load basic information record from the load basic information set FHBaseSet. .

[0057] S44. Extract basic load information Load location in and load time series information .

[0058] S45, Position the load The input is fed into the pre-trained load region feature enhancement expression model in S2, and the output is a 40-dimensional feature vector. .

[0059] S46. Load time series information and load location The input is fed into the pre-trained time-series feature description model with region labels in S3, and the output is a 40-dimensional feature vector. .

[0060] S47. The feature vector and eigenvectors Concatenate them into an 80-dimensional attribute vector. .

[0061] S48. Extract basic load information Flexible load markings in As a decision label .

[0062] S49. Add a new row to the empty decision table FHDecisionTable, containing the attribute vectors mentioned above. and decision labels .

[0063] S410, the loop variable FHCounter is incremented by 1.

[0064] S411. If the loop variable FHCounter is less than the total number N of load basic information in the load basic information set FHBaseSet, then execute S43; otherwise, the decision table FHDecisionTable is completed.

[0065] S5. This step uses the decision table generated in S4 to train a deep learning model, enabling it to learn the complex mapping relationship between 80-dimensional composite features and flexible load labels. The specific process is as follows:

[0066] S51. Construct a decision model FHDecisionModel (Flexible-load Host Decision Model) containing a 5-layer convolutional neural network. The specific structure is as follows:

[0067] The first layer is the input layer, which receives a feature vector with a dimension of 80. The decision table FHDecisionTable obtained from S411 is used as input.

[0068] The second layer is a one-dimensional convolutional layer, using 16 convolutional kernels of size 3 to perform convolution operations on the input feature vector to extract local features. The ReLU activation function is used, which increases the non-linear expressive power of the network.

[0069] The third layer is a 1D Max Pooling Layer, used to downsample the output of the 1D convolutional layer. The pooling window size is 2, which aims to reduce data dimensionality, retain the most salient features, and improve the generalization ability of the decision model.

[0070] The fourth layer is a fully connected layer (or dense layer). Before this layer, the output of the one-dimensional max-pooling layer is flattened into a one-dimensional vector. Then, a fully connected layer with 32 neurons is used to integrate the previously extracted features. The ReLU activation function is also used.

[0071] The fifth layer is the output layer, containing one neuron that outputs the final decision result. The activation function uses the sigmoid function, mapping the output value to between 0 and 1, representing the probability that the load is identified as a flexible load.

[0072] S52. Using the attribute vectors in the decision table FHDecisionTable as training input and the decision labels corresponding to the attribute vectors as training targets, train the decision model FHDecisionModel. Output the probability that the load corresponding to the attribute vector is a flexible load. If the probability is greater than 0.5, the load is determined to be a flexible load; otherwise, it is determined to be a non-flexible load. The trained decision model is then obtained. After training, save the network structure and weight parameters of FHDecisionModel.

[0073] S6. This step describes how to use a trained decision model to flexibly identify new, unlabeled load information. The specific process is as follows:

[0074] S61. Input a load information to be identified, UnknownLoad. The load information includes the load time series information UnknownLoadTimeSeries and the latitude and longitude information of the load location UnknownPosition.

[0075] S62. Input the latitude and longitude information of the load location (UnknownPosition) into the trained load area characteristic enhancement expression model, and output a 40-dimensional feature vector (UnknownLACEEMVector).

[0076] S63. Input the load time series information UnknownLoadTimeSeries and the latitude and longitude information UnknownPosition of the load location into the trained time series feature description model with regional labels, and output a 40-dimensional feature vector UnknownRLTSFDVector.

[0077] S64. Concatenate the feature vectors UnknownLACEEMVector and UnknownRLTSFDVector into an 80-dimensional input vector UnknownInputVector.

[0078] S65. Input the input vector UnknownInputVector into the decision model trained in S5 and output the value OutputScore. If OutputScore is greater than 0.5, UnknownLoad is determined to be a flexible load; otherwise, it is determined to be a non-flexible load.

[0079] The above examples of the present invention are merely illustrative of the computational model and process of the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is impossible to exhaustively list all possible implementations here. Any obvious variations or modifications derived from the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A flexible load identification method based on load region characteristic enhancement, characterized in that: It comprises the following steps: S1, input a plurality of load basic information sets of a plurality of electricity accounts in a city and a city functional area point information set; S2, construct a load area characteristic enhanced expression model, input the load basic information set and the city functional area point information set into the load area characteristic enhanced expression model for training, output a feature vector, and obtain the trained load area characteristic enhanced expression model; S3, construct a time series feature description model with a region label, input the load basic information set and the city functional area point information set into the time series feature description model with a region label for training, output a feature vector, and obtain the trained time series feature description model with a region label; S4, based on S1-S3, construct a decision table; S5, construct a decision model, train the decision model using the decision table, and obtain the trained decision model; S6, input a load information to be identified, the load information including load time series information and a load position, input the load position into the trained load area characteristic enhanced expression model, output a feature vector UnknownLACEEMVector, input the load time series information and the load position into the trained time series feature description model with a region label, output a feature vector UnknownRLTSFDVector, splice the feature vector UnknownLACEEMVector and the feature vector UnknownRLTSFDVector to obtain an input vector UnknownInputVector, input the input vector UnknownInputVector into the trained decision model, output a probability that the load is a flexible load, if the probability is greater than 0.5, determine that the load is a flexible load, otherwise, determine that the load is a non-flexible load.

2. The flexible load identification method based on load region characteristic enhancement according to claim 1, characterized in that: The load basic information set in S1 comprises N pieces of load basic information, each piece of load basic information comprising load time series information, a load position, and a flexible load label; The city functional area point information set comprises a plurality of region point information, each region point information comprising a region position and a region type, and the region type is divided into a residential area, a commercial area, an administrative area, and a leisure area.

3. The flexible load identification method based on load region characteristic enhancement according to claim 2, characterized in that: The specific process of S2 is as follows: S21, construct a load area characteristic enhanced expression model, the load area characteristic enhanced expression model comprising a residential area spatial relationship quantification submodule, a commercial area spatial relationship quantification submodule, an administrative area spatial relationship quantification submodule, and a leisure area spatial relationship quantification submodule; S22, select all region points with a residential area type from the city functional area point information set, and form a residential area point set, input the residential area point set into the residential area spatial relationship quantification submodule, calculate the distance from each load position to the nearest 10 region points in the residential area point set, and sort the distances according to the distance size, encode all the sorted distances into a 10-dimensional feature vector ResidentialVector, and take the feature vector ResidentialVector as the output; S23, all region points of which region type is commercial district are filtered out from the city function region point information set, and a commercial district point set is formed, the commercial district point set is input into a commercial district spatial relationship quantization submodule, the mean, variance, skewness and kurtosis indexes of the distance between each load position and all region points in the commercial district point set are calculated, and a 10-dimensional feature vector CommercialVector is output through principal component analysis; S24, filtering out all the region points with the region type of administrative region from the urban functional region point information set, and composing an administrative region point set, inputting the administrative region point set into an administrative region spatial relationship quantification submodule, and defining the attraction of each region point in the administrative region point set to each load position , the geographic spatial distance between each load position and each region point, according to the attraction value of each load position to each region point in the administrative region point set is calculated, the largest 10 attraction values are sorted from large to small, and encoded into a 10-dimensional feature vector AdministrativeVector, and the feature vector AdministrativeVector is taken as the output; S25, all region points of which region type is leisure district are filtered out from the city function region point information set, and a leisure district point set is formed, the leisure district point set is input into a leisure district spatial relationship quantization submodule, 10 bandwidth values are preset, a standard Gaussian kernel function is selected according to all region points in the leisure district point set, the 10 bandwidth values are sequentially substituted into the Gaussian kernel function to calculate the kernel density estimation of each load position, 10 kernel density estimation values under different spatial scales are obtained, the 10 kernel density estimation values are arranged in order, and a 10-dimensional feature vector LeisureVector is output; S26, the feature vector ResidentialVector, the feature vector CommercialVector, the feature vector AdministrativeVector and the feature vector LeisureVector are spliced in order, a 40-dimensional feature vector LACEEMVector is output, and a trained load region characteristic enhancement expression model is obtained.

4. The flexible load identification method based on load region characteristic enhancement according to claim 3, characterized in that: The specific process of S3 is: S31, a time series feature description model with region labels is constructed, the time series feature description model with region labels includes a residential area label feature extraction submodule, a commercial district label feature extraction submodule, an administrative district label feature extraction submodule and a leisure district label feature extraction submodule; S32, calculate the average distance from each load location to all zone points in the set of residential zone points ; calculating an average distance of each load location to all zone points in the set of commercial zone points ; calculating an average distance of each load location to all zone points in the administrative zone point set ; calculating an average distance of each load location to all region points in the set of region points ; S33, calculate the average distance of each load location to all regional points in the city regional point information set ; S34, based on S32 and S33, residential district indexes, commercial district indexes, administrative district indexes and leisure district indexes are calculated respectively; S35, input the load time series information and the residential area index of each load into the residential area band label feature extraction submodule, extract the main frequency component of the load time series information using Fourier transform, input the residential area index as an additional channel or weight together with the main frequency component into the small self-encoder SelfCode, and output a 10-dimensional feature vector ; S36, input the load time series information of each load and the business district index into a business district band label feature extraction submodule, extract the main frequency components of the load time series information by using Fourier transform, fuse the main frequency components with the business district index as an independent additional channel or a weighting coefficient, input the fusion result into a small self-encoder SelfCode, and output a 10-dimensional feature vector ; S37, input the load time series information and the administrative district index of each load into the administrative district band label feature extraction submodule, extract the main frequency component of the load time series information by using Fourier transform, fuse the administrative district index as an independent additional channel or weighting coefficient with the main frequency component, input the fusion result into a small self-encoder SelfCode, and output a 10-dimensional feature vector ; S38, input the load time series information and the leisure area index of each load into the leisure area band label feature extraction submodule, extract the main frequency component of the load time series information by using Fourier transform, fuse the main frequency component with the leisure area index as an independent additional channel or weighting coefficient, input the fusion result into a small self-encoder SelfCode, and output a 10-dimensional feature vector ; S39、concatenate the feature vectors , the feature vectors , the feature vectors , and the feature vectors in the order of precedence, output a 40-dimensional feature vector RLTSFDVector, and obtain the trained time series feature description model with region labels.

5. The flexible load identification method based on load region characteristic enhancement according to claim 4, characterized in that: The expression of the residential district index in S34 is: （1） wherein is the residential index.

6. The flexible load identification method based on load region characteristic enhancement according to claim 5, characterized in that: The expression of the commercial district index in S34 is: （2） wherein is the business district index.

7. The flexible load identification method based on load region characteristic enhancement according to claim 6, characterized in that: The expression of the administrative district index in S34 is: （3） wherein is the administrative district index.

8. The flexible load identification method based on load region characteristic enhancement according to claim 7, characterized in that: The expression of the leisure district index in S34 is: （4） wherein, is the leisure area index.

9. The flexible load identification method based on load region characteristic enhancement according to claim 8, characterized in that: The specific process of S4 is: S41, an empty decision table FHDecisionTable is initialized, S42, a loop variable FHCounter is initialized to 0; S43, take out the FHCounter pieces of load basic information from the load basic information set ; S44, extracting load basic information in the load position and load time series information ; S45, the load position The 40-dimensional feature vector is output in the load region characteristic enhancement expression model trained in S2 ; S46, load time series information and load position In the time series feature description model with region label trained in S3, output a 40-dimensional feature vector ; S47, the feature vector and the feature vector spliced into an 80-dimensional attribute vector ; S48, extract load basic information flexible load mark in as decision tag ; S49, a row is added to the empty decision table FHDecisionTable, which contains the attribute vector and the decision tag ; S410, the loop variable FHCounter is automatically increased by 1; S411, if the loop variable FHCounter is less than the total number N of load basic information in the load basic information set, S42 is executed, otherwise, the decision table FHDecisionTable is constructed.

10. The flexible load identification method based on load region characteristic enhancement according to claim 9, characterized in that: The specific process of S5 is: S51, a decision model is constructed, the decision model includes an input layer, a one-dimensional convolution layer, a one-dimensional maximum pooling layer, a full connection layer and an output layer; S52, input the attribute vector in the decision table into the decision model for training, take the decision label corresponding to the attribute vector as the training target, output the probability that the load corresponding to the attribute vector is a flexible load, if the probability is greater than 0.5, determine that the load is a flexible load, otherwise, determine that the load is a non-flexible load, and obtain the trained decision model.

Images