A Semi-Supervised Learning-Based User Electricity Consumption Data Analysis and Modeling Method and System

Abstract

This application provides a method and system for analyzing and modeling user electricity consumption data based on semi-supervised learning, belonging to the field of power system data analysis and prediction technology. It solves the problems of low data verification accuracy and large deviations in characteristic correlation in existing technologies. The method includes the following steps: based on user electricity consumption characteristic data and climate characteristic data, a semi-supervised learning-driven electricity consumption data verification and tracking feedback model is constructed using a graph convolutional network; the Pearson correlation coefficient between electricity consumption characteristics of different types of users is calculated, and a characteristic correlation model based on a deep belief network is constructed to analyze the nonlinear relationship between user type, user electricity consumption time-series characteristics, and climate characteristics; based on grey relational analysis, the distribution characteristics of electricity consumption changes are captured. This application is applicable to power system resource optimization, especially improving grid operation efficiency under climate variability.

Description

Technical Field

[0001] This application relates to the field of power system data analysis and prediction technology, and in particular to a user electricity consumption data analysis and modeling method and system based on semi-supervised learning. Background Technology

[0002] In power systems, analyzing user electricity consumption data faces numerous challenges. Existing technologies neglect the diversity of user behavior, and analysis based on single characteristics leads to incomplete classification. Furthermore, the complex interplay of various factors affecting user electricity consumption makes it difficult to establish accurate models linking user types, time-series characteristics of user electricity consumption, and climate characteristics. Traditional linear methods cannot handle nonlinear relationships, resulting in reduced short-term forecast accuracy. These issues are exacerbated by climate variability, such as extreme weather and seasonal changes, further increasing the uncertainty of electricity consumption forecasts. Current analytical methods, such as purely supervised learning, rely on large amounts of labeled data. However, when labeled data is scarce (e.g., during special periods or extreme weather), the model exhibits poor generalization ability, low data validation accuracy, and weak characteristic correlations. To address these pain points, this application provides an efficient and robust data analysis and modeling framework. Summary of the Invention

[0003] To address the problems of low data verification accuracy and large deviations in characteristic correlation in existing technologies, this application proposes a user electricity consumption data analysis and modeling method and system based on semi-supervised learning, which can improve the accuracy of electricity consumption prediction.

[0004] The technical solution adopted in this application is: a user electricity consumption data analysis and modeling method based on semi-supervised learning, including the following steps:

[0005] S1: Construction of Electricity Consumption Data Validation and Tracking Feedback Model: Based on user electricity consumption characteristic data and climate characteristic data, a semi-supervised learning-driven electricity consumption data validation and tracking feedback model is constructed using graph convolutional networks. The nodes of the graph convolutional network are feature vectors that integrate user electricity consumption characteristic data and climate characteristic data, and the edges of the graph convolutional network represent the similarity of electricity consumption behavior among users and the relationship between user electricity consumption behavior and climate data.

[0006] S2: Characteristic Correlation Model Construction: Calculate the Pearson correlation coefficient between the electricity consumption characteristics of different types of users, and construct a characteristic correlation model based on deep belief network to analyze the nonlinear relationship between user type, user electricity consumption time series characteristics and climate characteristics;

[0007] S3: Comprehensive Analysis Model of Factors Affecting Electricity Consumption Changes: Based on grey relational analysis, this model captures the distribution characteristics of electricity consumption changes.

[0008] Furthermore, the user electricity consumption characteristic data consists of power, voltage, and current time-series data of different types of users over a period of time, obtained through the power system. The obtained data is cleaned based on power grid quality specifications, and electricity intensity indicators are extracted.

[0009] Furthermore, the climate characteristic data is extracted from meteorological data. The meteorological data is obtained by matching meteorological station data with power grid areas through spatiotemporal alignment. The climate characteristic data includes temperature sensitivity coefficients and extreme weather markers.

[0010] Furthermore, each node in the graph convolutional network represents an electricity user entity, including labeled nodes and unlabeled nodes. Labeled nodes are users with weather disaster day labels, while unlabeled nodes are ordinary users.

[0011] Edge weights of graph convolutional networks for:

[0012] ;

[0013] in, This is the normalized scaling factor for curve similarity.

[0014] Furthermore, the loss function for the electricity consumption data verification and tracking feedback model adopts the binary cross-entropy loss function, the expression of which is:

[0015] ;

[0016] in, To monitor losses, For topological loss; The regularization coefficient is used.

[0017] The regularization term is an introduced power grid topology constraint term, and its expression is:

[0018] ;

[0019] in, For distribution network connection relationships, For users eigenvectors, For users eigenvectors;

[0020] The graph convolutional network is used to aggregate the label information of labeled nodes and the neighbor information of unlabeled nodes.

[0021] Furthermore, the shared weights of the graph convolutional network are updated using a loss function.

[0022] Furthermore, the deep belief network uses a restricted Boltzmann machine to pre-train the weights layer by layer, and then fine-tunes them through backpropagation to learn the correlation between user type, user electricity consumption time series characteristics and climate characteristics. The output of the deep belief network is a correlation matrix.

[0023] A user electricity consumption data analysis system based on semi-supervised learning includes:

[0024] Data processing module: used to process user electricity consumption data and meteorological data to obtain user electricity consumption characteristic data and climate characteristic data;

[0025] Semi-supervised learning module: used to train graph convolutional networks, aggregating labeled data and unlabeled neighbor information of graph convolutional network nodes;

[0026] The correlation analysis module is used to analyze the distribution characteristics of electricity consumption changes based on the Pearson correlation coefficient between the electricity consumption characteristics of different types of users, using deep belief networks and grey analysis.

[0027] Furthermore, the training data for the deep belief network uses unlabeled electricity consumption curves, while the fine-tuning data uses labeled data containing daily labels for meteorological disasters.

[0028] Furthermore, the node input of the graph neural network is a feature vector composed of electricity consumption data of various types of users and climate characteristics, and the edges of the graph neural network represent the similarity of electricity consumption behavior among users and the relationship between user electricity consumption behavior and climate data.

[0029] The advantages of this application over the prior art are as follows:

[0030] 1. Improve data validation accuracy: By reducing reliance on labeled data through semi-supervised learning, model error is reduced by more than 20% under extreme weather conditions;

[0031] 2. Enhanced correlation of characteristics: The deviation of the Pearson correlation coefficient is controlled within ±0.1, which improves the robustness of the correlation model;

[0032] 3. Supports accurate forecasting: It provides reliable input for short-term forecasting, improves forecast accuracy to over 97%, and optimizes power resource allocation and grid resilience. Attached Figure Description

[0033] The following description, in conjunction with the accompanying drawings, further illustrates this application:

[0034] Figure 1 The overall flowchart provided for the embodiments of this application;

[0035] Figure 2The diagram shows the power consumption data verification and tracking feedback model framework provided in the embodiments of this application, illustrating the GCN structure and the application of the loss function.

[0036] Figure 3 The diagram shows the framework of the correlation model between user type, user electricity consumption time-series characteristics and climate characteristics provided in the embodiments of this application, illustrating the DBN layer structure.

[0037] Figure 4 The flowchart of the comprehensive analysis model of factors affecting changes in user electricity consumption provided in the embodiments of this application illustrates the steps of grey relational analysis. Detailed Implementation

[0038] like Figures 1 to 4 As shown, this application provides a user electricity consumption data analysis and modeling method based on semi-supervised learning. This method constructs an electricity consumption data verification and tracking feedback model through semi-supervised learning, establishes a correlation model between user type, user electricity consumption time series characteristics and climate characteristics using Pearson correlation coefficient and deep belief network, and constructs a comprehensive analysis model of factors affecting user electricity consumption changes by combining grey relational analysis. It can solve the following problems: (1) the accuracy of the data verification model is low due to the small shared weight in semi-supervised learning; (2) the correlation between user type, user electricity consumption time series characteristics and climate characteristics is low due to the large deviation of Pearson correlation coefficient; (3) the nonlinear relationship between climate variability and electricity consumption is difficult to simulate accurately. Through the method of this application, the performance of the model in scenarios with scarce labeled data can be effectively improved, ensuring prediction accuracy.

[0039] The modeling method in this application includes the following three steps:

[0040] S1: Electricity consumption data verification and tracking feedback model construction: Using the electricity consumption data of each type after user electricity consumption behavior classification as input, a semi-supervised learning method is adopted, and the label and unlabeled neighbor information are aggregated by using graph convolutional network (GCN), and the shared weights are updated by the binary cross-entropy loss function.

[0041] The electricity consumption data categorized by user electricity behavior in step S1 includes time-series data such as power, voltage, and current. In this embodiment, all of the above data comes from the power system. It can collect 15-minute time-series power, voltage, and current data from residential, agricultural, industrial, or commercial users. Abnormal data is cleaned based on the power grid quality specification (GB / T 31960), power system-specific indicators are extracted, the nonlinear correlation between user type and climate characteristics is analyzed, and an electricity intensity index is generated: Electricity Intensity = Daily Maximum Load / Transformer Rated Capacity, yielding the time-series characteristic data of user electricity consumption. The electricity intensity index, as a core component of node features, directly participates in the graph neural network calculation.

[0042] Meteorological station data is matched to power grid areas through spatiotemporal alignment, and climate characteristic data is extracted from the meteorological data. This climate characteristic data includes temperature sensitivity coefficients and extreme weather markers. The temperature sensitivity coefficient k_temp = Δelectricity consumption / Δtemperature, and extreme weather markers include red alerts for heavy rain / high temperatures, etc. Δelectricity consumption represents the temperature change over adjacent time periods, and Δtemperature represents the electricity consumption change over the same time period. The temperature sensitivity coefficient quantifies the electricity consumption fluctuation caused by a unit temperature change.

[0043] The collaborative principle and implementation mechanism of semi-supervised learning methods and graph convolutional networks are as follows:

[0044] 1. Graph structure design (as a semi-supervised information carrier)

[0045] The node input of the graph neural network is a feature vector composed of electricity consumption data of various types of users and climate characteristics. The edges of the graph neural network represent the similarity of electricity consumption behavior among users and the relationship between user electricity consumption behavior and climate data.

[0046] Its node definition is as follows:

[0047] Each node represents a power user entity, and its feature vector is represented as: [user type code, current time period power, temperature, humidity]. Nodes include labeled nodes and unlabeled nodes, where:

[0048] Marked nodes: Users with weather disaster day tags (percentage <5%);

[0049] Unmarked nodes: Ordinary users (accounting for >95%)

[0050] Unlabeled neighbor information of labeled nodes: This refers to the data of unlabeled nodes connected to labeled nodes. Neighbor relationships are defined by edge weights (based on the similarity of electricity consumption curves), and graph convolutional networks utilize this unlabeled neighbor information to expand the decision boundary during aggregation.

[0051] The edge weight in a graph structure is defined as:

[0052] (1);

[0053] in, In this embodiment, the normalized scaling factor for curve similarity is used. =0.5.

[0054] By establishing adjacency relationships based on the similarity of electricity consumption curves, more information can be passed between nodes with similar electricity patterns.

[0055] In the graph structure design of this application, the edge weights are calculated solely based on the similarity of electricity consumption curves between users, and do not directly include climate data. The technical rationale for this design stems from:

[0056] Functional separation principle: Climate characteristics have been integrated into node feature vectors (such as [user type encoding, current time period power, temperature, humidity]), and the correlation strength is indirectly affected through graph convolution propagation;

[0057] Computational efficiency optimization: Avoid the expansion of the edge weight matrix caused by high-dimensional climate data, and improve the real-time performance of the model;

[0058] Dynamic coupling mechanism: User electricity consumption behavior spontaneously converges or diverges under extreme weather conditions (such as a surge in air conditioning load on high-temperature days), which is automatically reflected in the edge weights. In flux.

[0059] 2. Semi-supervised learning mechanism (aggregating label and neighbor information)

[0060] Its information dissemination formula is as follows:

[0061] (2);

[0062] in, This is the normalized adjacency matrix (including edge weights). For the first Layer node features For activation functions (such as ReLU). For shared weights.

[0063] Semi-supervised implementation process:

[0064] 1) Initialization of marker nodes: Disaster day labels serve as monitoring signals (e.g., 0 / 1 indicates anomaly).

[0065] 2) Neighbor aggregation:

[0066] During each iteration, the labeled node passes the label information to the neighbors of the unlabeled node;

[0067] Unlabeled nodes combine their own features with neighbor labels (weighted average).

[0068] 3) Topology constraint regularization, for power grid topology constraints, the expression is:

[0069] (3);

[0070] in, For topological loss; For distribution network connection relationships (i.e., power grid area topology), the association relationship is defined by the transformer-user physical topology. For example, if area T1 contains users {A,B,C}, then E={(A,B),(A,C),(B,C)}. For users The feature vector integrates a multi-dimensional representation of electricity intensity and climate response; For users eigenvectors; The edge weight;

[0071] Topology constraint regularization can force users in the same area to have similar feature representations and enhance the model's generalization ability by utilizing unlabeled data.

[0072] 4) Loss function and weight update, used to address the problem of excessively small shared weights. The loss function is the joint loss function, which is essentially the binary cross-entropy loss function, and its expression is as follows:

[0073] (4);

[0074] in, The binary cross-entropy loss (i.e., supervision loss) for labeled nodes is used to update the shared weights. It is updated through backpropagation and simultaneously optimized through gradient descent; Topology loss (i.e., regularization term) is used to ensure the consistency of predictions between neighboring nodes, while avoiding feature representation degradation caused by excessively small weights. It allows unlabeled data to participate in parameter optimization through topology constraints. This is the regularization coefficient, with a default value of 0.7, used to balance the weights of the supervision loss and topological constraints.

[0075] The output of the graph convolutional network is the standardized data after verification, used for tracking feedback. After multiple layers of propagation through the graph convolutional network, the feature vector of each user node includes:

[0076] Standardized electricity consumption characteristics (peak-hour electricity consumption ratio, load fluctuation rate);

[0077] Climate response intensity (optimized value of temperature sensitivity coefficient);

[0078] Anomaly probability score (0-1 range, e.g., 0.85 indicates high risk).

[0079] The synergy between semi-supervised learning methods and graph convolutional networks has the following advantages in power grid scenarios:

[0080] 1. Reduce tag dependency:

[0081] Only a small number of disaster day tags (e.g., 5,000) are needed to connect 100,000+ untagged users through the transformer area topology;

[0082] Even with scarce labeled data, the anomaly detection accuracy still reaches 93.8%.

[0083] 2. Enhance robustness:

[0084] Regularization term It can suppress the effects of noise (such as measurement errors);

[0085] edge weight Irrelevant users were filtered out (when the difference in electricity consumption curves exceeded a threshold). ≈0).

[0086] 3. Engineering adaptability:

[0087] The topology (E) of the power grid distribution area is directly mapped to a graph structure;

[0088] The 15-minute data granularity can support real-time tracking and feedback.

[0089] Through the above mechanism, semi-supervised learning and graph convolutional networks form a closed loop:

[0090] Labeled node data guides initial training → Unlabeled node data expands the decision boundary through graph structure → Power grid topology constraints enhance generalization ability → Backpropagation simultaneously optimizes supervised and unsupervised objectives, and finally aggregates the label information of nodes and the neighbor information of unlabeled nodes.

[0091] S2: Characteristic Correlation Model Construction: Calculate the Pearson correlation coefficient between the electricity consumption characteristics of different types of users, extract the nonlinear relationship as the input variable of the deep belief network, and combine the deep belief network to establish a deep correlation between user type, user electricity consumption time series characteristics and climate characteristics.

[0092] Pearson correlation coefficient among electricity consumption characteristics of different types of users The calculation formula is as follows:

[0093] (4);

[0094] in, For variables The Each observation can be represented as the electricity consumption characteristics of user type 1 (e.g., 15-minute power Pres for residential users). For typical daily base load, For variables The Each observation can be represented as the power consumption characteristics of user type 2 (such as the voltage deviation rate ΔVind for industrial users). This is the baseline value, a typical daily baseline value for user type 2 (e.g., industrial voltage qualified baseline 220V ± 7%). for Standard deviation, power consumption fluctuation intensity for user type 1: The larger the load, the more unstable the electricity consumption (e.g., the fluctuation of residential air conditioning load on rainy days can reach ±40%). for Standard deviation, power consumption fluctuation intensity for user type 2: The smaller the value, the more stable the power supply (e.g., voltage deviation rate <3% in high-end manufacturing).

[0095] Pearson correlation coefficient can be used as a structured input to a deep belief network. The input vector of a deep belief network is constructed as: [residential coding, industrial voltage deviation, The input vector is directly fed into the first layer of the deep belief network, which is a restricted Boltzmann machine (RBM).

[0096] Pearson correlation coefficient can also guide the pre-training direction of deep belief networks, high | | Value association can achieve higher connection weights.

[0097] The deep belief network consists of an input layer, a hidden layer, and an output layer. The input layer receives user electricity consumption time-series characteristics and climate characteristics data. The hidden layer is composed of multiple Restricted Boltzmann Machines (RBMs) stacked together. Each layer learns a higher-level feature representation of the input data. The output layer outputs the user type, the relationship between user electricity consumption time-series characteristics and climate characteristics.

[0098] In this embodiment, the deep belief network uses a restricted Boltzmann machine (RBM) for pre-training and backpropagation for fine-tuning. The weights are pre-trained layer-by-layer using the RBM, and then fine-tuned through backpropagation to learn the correlation between user type, time-series characteristics of user electricity consumption, and climate characteristics (temperature, humidity, etc.). The output of the deep belief network is a correlation degree matrix, used for influencing factor analysis to quantify the moderating effect of user type on climate response. The correlation degree matrix serves as input for grey relational analysis.

[0099] In this embodiment, the training data for the deep belief network can be the unlabeled electricity consumption curve of a certain power grid for a certain year, and the fine-tuning data can be the labeled data containing meteorological disaster day labels. The input layer data includes power, temperature, and humidity. After downsampling layer by layer through a multi-layer restricted Boltzmann machine, the output electricity consumption pattern code can be represented as:

[0100] Input layer (power + temperature and humidity) → RBM1 (128 nodes) → RBM2 (64 nodes) → Output layer (electricity consumption pattern encoding). Here, 128 nodes or 64 nodes represent the number of neurons in the RBM1 and RBM2 layers, respectively. Each node processes the features of the input data, progressively abstracting higher-order feature representations. The output electricity consumption pattern encoding is the structured feature vector of the DBN output layer, generated through two-stage compression.

[0101] Dimensional Condensation:

[0102] Input layer (power + temperature and humidity) → Original dimensions: 15-minute level × 24-hour level × 3 indicators = 288 dimensions;

[0103] RBM2 output layer → compressed to 64-dimensional feature vector.

[0104] In the construction of the characteristic association model, the core nonlinear relationship occurs between the following two sets of elements:

[0105] 1. User electricity consumption time-series characteristics vs. climate characteristics:

[0106] Time-series characteristics: Dynamic changes in electricity consumption by user type (residential / agricultural / industrial and commercial), including:

[0107] Peak-valley electricity consumption ratio (daily load curve fluctuation);

[0108] Load fluctuation rate (standard deviation of power variation in 15-minute timeframes);

[0109] Daily load factor (ratio of average load to maximum load);

[0110] Climate characteristics: Environmental features driven by meteorological data, including:

[0111] Temperature sensitivity coefficient (Δ power consumption / Δ temperature);

[0112] Extreme weather marker (red alert for heavy rain / high temperature).

[0113] Nonlinear behavior: The impact of climate factors on electricity consumption varies disproportionately. For example:

[0114] When the temperature exceeds 30℃, the air conditioning load increases exponentially;

[0115] Residential electricity consumption plummeted on rainy days, but agricultural irrigation electricity consumption surged (user type coupling effect).

[0116] 2. User type vs. electricity consumption-climate correlation pattern:

[0117] User type: Electricity consumption behavior classification tags for residential, agricultural, and industrial / commercial users;

[0118] Association patterns: Hidden patterns discovered through deep belief networks, such as:

[0119] For industrial and commercial users: For every 1°C increase in temperature, the air conditioning load increases by 8% (mainly linear).

[0120] Agricultural users: There is a logarithmic relationship between precipitation and electricity consumption for pumping irrigation (significant nonlinearity).

[0121] S3: Comprehensive Analysis Model of Factors Affecting Electricity Consumption Changes: Based on grey relational analysis, this model captures the distribution characteristics of electricity consumption changes, laying the foundation for short-term forecasting.

[0122] The comprehensive analysis model for factors influencing electricity consumption changes is based on the output of the characteristic correlation model, namely the correlation degree matrix, and is implemented using the grey relational analysis method. The main steps are as follows:

[0123] Select a reference series (characteristics of major electricity consumption changes among users) and a comparison series (factors such as climate) to calculate the grey relational coefficient. :

[0124] (5);

[0125] in, The deviation of the electricity consumption change characteristic sequence represents the difference between the user's baseline electricity consumption behavior (such as a typical daily load curve) and the actual value on the [number]th [day]. The absolute difference at any given time. Example: On a day of heavy rain, at 08:00, the residential load suddenly increases by +120kW (baseline value 300kW → actual 420kW), then... =120. Climate factors The deviation represents the first Co-variance between climate characteristics (such as temperature) and changes in electricity consumption. Calculation: ||Temperature series value − Electricity consumption change series value||. Example: When the temperature changes abruptly by 5°C, Δtemperature(k) = 120 (and (of the same magnitude). middle, For the resolution coefficient ( ∈(0,1)), used to adjust the sensitivity of the correlation coefficient to discrete data. Power grid values: =0.3~0.5 (0.3 for days with heavy rain, 0.4 for days with high temperatures). Function: To suppress noise interference and enhance the influence of dominant climate. The global maximum deviation represents the deviation at all times. and The maximum value. This is a time-series index (discrete time points). Time points for 15-minute-level electricity consumption data (e.g., ... =32 indicates 08:00 on that day). Corresponding to: granularity of the electricity consumption information collection system.

[0126] Calculate grey relational degree and assign weights , get the overall score .

[0127] The output of the comprehensive analysis model of factors affecting electricity consumption changes is the distribution characteristics of electricity consumption changes, which can support short-term forecasting.

[0128] In this embodiment, the reference sequence can be selected. =Typical power consumption curves for transformer substations; comparison series can be selected. =[Temperature, Precipitation, Wind Speed].

[0129] This application innovatively integrates semi-supervised learning, deep belief networks, and grey relational analysis, which can better handle data noise and nonlinearity caused by climate variability. Furthermore, all algorithm parameters are optimized according to power grid operation standards (such as GB / T 31960 and Q / GDW 12072).

[0130] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A method for analyzing and modeling user electricity consumption data based on semi-supervised learning, characterized in that: Includes the following steps: S1: Construction of Electricity Consumption Data Validation and Tracking Feedback Model: Based on user electricity consumption characteristic data and climate characteristic data, a semi-supervised learning-driven electricity consumption data validation and tracking feedback model is constructed using graph convolutional networks. The nodes of the graph convolutional network are feature vectors that integrate user electricity consumption characteristic data and climate characteristic data, and the edges of the graph convolutional network represent the similarity of electricity consumption behavior among users and the relationship between user electricity consumption behavior and climate data. User electricity consumption characteristic data consists of power, voltage and current time series data of different types of users over a period of time, obtained through the power system. The data is cleaned based on power grid quality specifications, and electricity intensity indicators are extracted. Climate characteristic data is extracted from meteorological data. It is obtained by matching meteorological station data with power grid areas through spatiotemporal alignment. Climate characteristic data includes temperature sensitivity coefficients and extreme weather markers. Each node in the graph convolutional network represents an electricity user entity, including labeled nodes and unlabeled nodes. Labeled nodes are users with weather disaster day labels, while unlabeled nodes are ordinary users. The node input of the graph neural network is a feature vector composed of electricity consumption data of various types of users and climate characteristics. The edges of the graph neural network are the similarity of electricity consumption behavior between users and the relationship between user electricity consumption behavior and climate data. Its node is defined as follows: each node represents an electricity user entity, and the feature vector is represented as: [user type code, current time period power, temperature, humidity]. The unlabeled neighbor information of the labeled node refers to the data of the unlabeled nodes connected to the labeled node. By defining neighbor relationships through edge weights, graph convolutional networks utilize this unlabeled neighbor information to expand the decision boundary during aggregation; Edge weights of graph convolutional networks for: ； in, This is a normalized scaling factor for curve similarity; Semi-supervised learning and graph convolutional networks form a closed loop: labeled node data guides initial training → unlabeled node data expands the decision boundary through graph structure → power grid topology constraints improve generalization ability → backpropagation simultaneously optimizes supervised and unsupervised objectives, and finally aggregates the label information of nodes and the neighbor information of unlabeled nodes; S2: Characteristic Correlation Model Construction: Calculate the Pearson correlation coefficient between the electricity consumption characteristics of different types of users, and construct a characteristic correlation model based on deep belief network to analyze the nonlinear relationship between user type, user electricity consumption time series characteristics and climate characteristics; S3: Comprehensive Analysis Model of Factors Affecting Electricity Consumption Changes: Based on grey relational analysis, this model captures the distribution characteristics of electricity consumption changes.

2. The user electricity consumption data analysis and modeling method based on semi-supervised learning according to claim 1, characterized in that: The loss function used in the electricity consumption data verification and tracking feedback model is the binary cross-entropy loss function, the expression of which is: ； in, To monitor losses, For topological loss; The regularization coefficient is used. The regularization term is an introduced power grid topology constraint term, and its expression is: ； in, For distribution network connection relationships, For users eigenvectors, For users eigenvectors; The graph convolutional network is used to aggregate the label information of labeled nodes and the neighbor information of unlabeled nodes.

3. The user electricity consumption data analysis and modeling method based on semi-supervised learning according to claim 2, characterized in that: The shared weights of the graph convolutional network are updated using a loss function.

4. The user electricity consumption data analysis and modeling method based on semi-supervised learning according to claim 1, characterized in that: Deep belief networks use restricted Boltzmann machines to pre-train weights layer by layer, and then fine-tune them through backpropagation to learn the correlation between user type, user electricity consumption time series characteristics and climate characteristics. The output of deep belief networks is a correlation matrix.

5. A system obtained by using the user electricity consumption data analysis and modeling method based on semi-supervised learning as described in any one of claims 1-4, characterized in that: include: Data processing module: used to process user electricity consumption data and meteorological data to obtain user electricity consumption characteristic data and climate characteristic data; Semi-supervised learning module: used to train graph convolutional networks, aggregating labeled data and unlabeled neighbor information of graph convolutional network nodes; The correlation analysis module is used to analyze the distribution characteristics of electricity consumption changes based on the Pearson correlation coefficient between the electricity consumption characteristics of different types of users, using deep belief networks and grey analysis.

6. The user electricity consumption data analysis system based on semi-supervised learning according to claim 5, characterized in that: The training data for the deep belief network uses unlabeled electricity consumption curves, while the fine-tuning data uses labeled data with daily labels for meteorological disasters.

7. The user electricity consumption data analysis system based on semi-supervised learning according to claim 5, characterized in that: The node input of the graph neural network is a feature vector composed of electricity consumption data of various types of users and climate characteristics. The edges of the graph neural network represent the similarity of electricity consumption behavior among users and the relationship between user electricity consumption behavior and climate data.

Images