A power distribution load forecasting method and system

Abstract

The application discloses a power distribution load prediction method and system, and relates to the technical field of graph calculation, and comprises the following steps: obtaining load data and auxiliary features of a user, and constructing a load portrait; according to the load portrait, selecting a model subset from a pre-trained model library, and migrating the selected model to output an initial load prediction result; constructing a load correlation graph according to spatial adjacency and behavior similarity between users, and propagating load features in the load correlation graph; according to the propagated load features and the initial load prediction result, regulating and controlling the model of a target user; obtaining the distribution change of the regulated and controlled user load data, and identifying behavior mutation; through multi-model integration, individual fine-tuning, load correlation graph construction, incremental learning and drift detection, the application effectively solves the problems of low precision, poor dynamic adaptability and slow reaction to sudden changes in traditional power distribution load prediction.

Description

Technical Field

[0001] This invention relates to the field of graph calculation technology, and more specifically, to a method and system for predicting power distribution load. Background Technology

[0002] With increasingly complex and diversified electricity demand, load forecasting of distribution systems has gradually become one of the core issues in power dispatching and management. The accuracy of distribution load forecasting directly affects the operating efficiency, stability, and optimal allocation of resources of the power grid. Traditional load forecasting methods often rely on statistical models or basic machine learning models, but these methods exhibit low forecasting accuracy and adaptability when faced with highly complex and nonlinear load changes. Therefore, improving the accuracy, dynamic adaptability, and response to sudden load fluctuations in load forecasting has become a crucial technical challenge that urgently needs to be addressed in the power sector.

[0003] For example, the invention patent with announcement number CN111582548A discloses a method for predicting electricity load based on multi-user behavior profiles. This method predicts the electricity load of multiple users and includes the following steps: selecting historical electricity load data from several multi-user groups and preprocessing the historical electricity load data; training a model on the preprocessed historical electricity load data for different types of multi-users to obtain corresponding prediction models. These different types of multi-users include users with poor electricity consumption stability, users with good electricity consumption stability, and users with low temperature sensitivity; predicting the electricity load of the multi-users to obtain prediction results, and displaying the prediction results. This method can apply corresponding prediction methods to different types of multi-users, thereby achieving rapid and accurate prediction of power system load.

[0004] For example, the invention patent with announcement number CN118523291A discloses a low-voltage transformer area load forecasting method based on neural networks, which specifically includes the following steps: Step 1, collecting historical load data of the low-voltage transformer area; Step 2, constructing a joint neural network model and pre-training the joint neural network model; Step 3, inputting the historical load data into the trained joint neural network model to obtain preliminary prediction values; Step 4, performing error and pass rate analysis on the preliminary prediction values. If the pass rate is satisfactory, the preliminary prediction value is recognized as the final prediction value; if the pass rate is unsatisfactory, returning to Step 3, a new preliminary prediction value is obtained. By combining multiple algorithms, the problem of slow low-voltage transformer area load forecasting process and low prediction accuracy in existing methods is solved.

[0005] The above-disclosed technical solutions have at least the following technical problems:

[0006] Traditional power distribution load forecasting methods rely on a single global or fixed model, failing to consider individual differences among users. This results in inaccurate or unpersonalized forecasts when dealing with users with varying load demands. Furthermore, traditional methods often use static load correlation models, ignoring the dynamic characteristics of load changes over time. For users with large load fluctuations and unstable electricity consumption behavior, there is a lack of targeted adjustments, leading to decreased forecast accuracy.

[0007] To address the above problems, this invention proposes a solution. Summary of the Invention

[0008] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a power distribution load prediction method and system, which solves the problems of low accuracy, poor dynamic adaptability, and slow response to sudden changes in traditional methods by constructing load correlation diagrams and detecting drift.

[0009] To achieve the above objectives, the present invention provides the following technical solution:

[0010] A power distribution load forecasting method includes: acquiring user load data and auxiliary features to construct a load profile; selecting a subset of models from a pre-trained model library based on the load profile, performing transfer on the selected models, and outputting initial load forecasting results; constructing a load association graph based on spatial adjacency and behavioral similarity between users, and propagating load features in the load association graph; adjusting the model of the target user based on the propagated load features and the initial load forecasting results; acquiring changes in the distribution of user load data after adjustment, and identifying behavioral mutations.

[0011] In a preferred embodiment, the step of acquiring user load data and auxiliary features to construct a load profile is as follows: cleaning and processing the collected load data; acquiring user auxiliary features and extracting load features by combining them with the load data; constructing a load feature vector based on the load features and merging the load feature vectors to form a load profile.

[0012] In a preferred embodiment, the step of selecting a subset of models from the pre-trained model library based on the load profile is as follows: establishing a model library containing multiple pre-trained models; vectorizing the individual load profile of each user into features, and calculating the similarity between the user load profile features and the feature vectors of each model in the pre-trained model library using the Manhattan distance method, outputting the similarity between the user load profile features and the features of each model; and selecting the top K most similar models according to the similarity ranking to form an initial model subset.

[0013] In a preferred embodiment, the transfer of the selected model to output the initial load prediction result is specifically as follows: For each model in the initial model subset, the encoder part is used as a shared feature extraction module, and a target domain adaptation module is constructed. The target domain adaptation module consists of a domain discriminator, an output predictor, and a transfer layer, forming a transfer model; the load profile of the target user and historical load data are input into the encoder of the transfer model, and the intermediate feature expression vector of the input is extracted as the shared feature space representation of the transfer structure; a domain discriminator is constructed based on the shared feature representation to identify source domain features and target domain features; the feature representations of the source domain samples and the target domain features are extracted, and the mean difference in the shared feature space is output; the initial model subset in the overall transfer model is switched according to the mean difference, and weights are assigned to the selected models according to the score of each model in the similarity calculation; the prediction results of the model subset are weighted and averaged according to the weights, and the initial load prediction result is output.

[0014] In a preferred embodiment, the construction of the load association graph based on spatial adjacency and behavioral similarity between users is as follows: Based on power grid topology data, the distribution network access relationship between each user is extracted to establish an initial spatial adjacency matrix; within each time window, the weights of the initial spatial adjacency matrix are dynamically adjusted to form a time-varying spatial adjacency matrix; based on a dynamic time warping algorithm, user load behavior characteristics are extracted, and the behavioral similarity matrix within each time window is calculated; within each time window, the time-varying spatial adjacency matrix and the behavioral similarity matrix are fused using a weighted fusion method to form a joint adjacency matrix, i.e., the load association graph.

[0015] In a preferred embodiment, the weights of the initial spatial adjacency matrix are dynamically adjusted within each time window to form a time-varying spatial adjacency matrix, specifically as follows: For each pair of users, the load similarity of the two sets of load data is calculated within the current time window through dynamic time warping; the load data of each user is statistically analyzed using the standard deviation algorithm to calculate the load volatility of the user within the current time window; the weights of the initial spatial adjacency matrix are dynamically adjusted based on the load volatility and load similarity; if a user's load fluctuates drastically within a certain period, the correlation with surrounding users should be enhanced, i.e., the weight value should be increased; if the load behavior of a user is extremely similar to that of other users within a certain time window, the connection strength between them will be enhanced, i.e., the weight value should be increased.

[0016] In a preferred embodiment, the propagation of load features in the load association graph is specifically as follows: the load profile features of each user node are mapped to a high-dimensional space as initial node features; based on the initial node features, the connection strength of each pair of neighboring nodes is dynamically adjusted through a graph self-attention mechanism; the adjusted node features are input into a graph neural network for propagation, and after several layers of graph convolution in the graph neural network, an enhanced feature representation of each user node is obtained.

[0017] In a preferred embodiment, the adjustment of the target user's model based on the propagated load characteristics combined with the initial load prediction results is as follows: The enhanced feature representation and the initial load prediction results are concatenated to form an input feature representation; this input feature representation is input into a pre-trained transfer model, and the mean squared error between the current prediction result and the actual load is output as a loss function using an incremental learning method; based on the loss function, the parameters in the transfer model are progressively updated using a gradient descent method, and a preliminary load prediction result is output; based on the mean absolute error between the load prediction results before and after updating the transfer model, the learning rate in the gradient descent method is dynamically adjusted to obtain a corrected parameter update model; and the load prediction result is output based on the corrected parameter update model.

[0018] In a preferred embodiment, the acquisition of changes in the distribution of user load data after regulation and the identification of behavioral abrupt changes are specifically as follows: The acquired user load data is input into the ADWIN drift detection algorithm in time series order; the ADWIN algorithm divides the load data stream into multiple data windows through a sliding window mechanism and calculates the mean and variance of the load data; the mean and variance of the load data are mapped to a preset drift threshold, and a dynamic drift threshold is output based on the mapping relationship between the mean and variance of the load data and the preset drift threshold; when ADWIN detects that the difference in mean and variance between the distribution of the current window data and the previous window data exceeds the drift threshold, a drift alarm is issued; when distribution drift is detected, an incremental learning mechanism is triggered, and the algorithm returns to the parameter update model to update the parameters.

[0019] A power distribution load forecasting system includes a data acquisition module, an initial forecasting module, a propagation module, a control module, and an identification module, with connections between the modules. The data acquisition module acquires user load data and auxiliary features to construct a load profile. The initial forecasting module selects a subset of models from a pre-trained model library based on the load profile, performs transfer operations on the selected models, and outputs initial load forecast results. The propagation module constructs a load association graph based on spatial adjacency and behavioral similarity between users and propagates load features within the load association graph. The control module controls the model of the target user based on the propagated load features and the initial load forecast results. The identification module acquires changes in the distribution of user load data after control and identifies behavioral abrupt changes.

[0020] The technical effects and advantages of the power distribution load forecasting method and system of the present invention are as follows:

[0021] 1. This invention, through multi-model ensemble, combines the advantages of different models, effectively reducing the errors that may arise from a single model. When selecting a suitable subset of pre-trained models, calculating the similarity between the load profile and the model ensures that the model best suited to the current user load characteristics is chosen. This personalized model selection and weighted ensemble method significantly improves the accuracy of distribution load forecasting.

[0022] 2. This invention, by introducing a load correlation graph, leverages the spatial adjacency and behavioral similarity between users to not only better capture dynamic changes in user load but also utilize graph neural networks to propagate and enhance load features. This enables the system to adapt to changes in user behavior in real time and dynamically adjust the prediction model within different time windows, further improving the accuracy and flexibility of prediction. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the power distribution load prediction method of the present invention.

[0024] Figure 2 This is a schematic diagram of the structure of a power distribution load prediction system according to the present invention. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0026] Example 1, Figure 1 The present invention provides a method for predicting power distribution load, comprising the following steps:

[0027] S1, acquire user load data and auxiliary characteristics, and construct load profiles;

[0028] In this embodiment, user load data and auxiliary features are obtained to construct a load profile, as detailed below:

[0029] User load data is collected in real time through smart meters or high-precision measuring devices. The collected load data includes: instantaneous power (load curve), historical electricity consumption data (including the user's historical load data, with time granularity selectable at hourly, daily or higher frequency levels), and load peak and valley values.

[0030] The collected load data is cleaned and processed. Abnormal data is identified and corrected using methods such as the Isolation Forest algorithm or DBSCAN. Wavelet transform is used to denoise the load fluctuation data to ensure data smoothness.

[0031] Obtain the user's auxiliary features, which include:

[0032] User category features: Classified according to user type (residential, commercial, industrial, etc.), labeled as static features;

[0033] Geographic information features: The user's geographical location is obtained through a geographic information system, and combined with environmental factors such as regional temperature and humidity, to supplement the electricity consumption prediction model;

[0034] Environmental change characteristics: Real-time collection of weather information in the user's location to identify its impact on load, such as peak air conditioning load in summer;

[0035] User behavior pattern characteristics: Based on users' historical load data, extract users' periodic electricity consumption pattern characteristics, including peak hours, holiday fluctuations, etc.

[0036] Based on load data and auxiliary features, load features are extracted, including: time features, spatial features and behavioral features;

[0037] A load feature vector is constructed based on load characteristics, and the load feature vectors are merged to form a high-dimensional user profile vector, namely the load profile.

[0038] The process of extracting load features based on load data and auxiliary features is as follows:

[0039] The time characteristics include periodicity and peak and trough periods;

[0040] The spatial features include the influence of geographical location and regional similarity features;

[0041] The behavior includes load fluctuation characteristics and equipment usage pattern characteristics;

[0042] Based on historical load data, frequency domain analysis of the load data is performed using Fourier transform or wavelet transform to extract periodic features such as daily, weekly, and monthly characteristics.

[0043] Based on historical load data, identify peak and off-peak periods for user electricity consumption (e.g., morning or evening peaks each day);

[0044] Cluster analysis (such as the K-means algorithm) can be used to identify the geographical region characteristics of users, namely regional similarity characteristics, which reflect the similarity of electricity consumption patterns among users in the same geographical location.

[0045] Geographic location impact: Spatial features are extracted by combining user geographic location information (such as city, region, building level, etc.) with environmental data (such as temperature, humidity, etc.). For example, users in warm regions may consume more electricity for air conditioning, while users in cold regions may consume more electricity for electric heating devices;

[0046] By analyzing the volatility of user load curves, load fluctuation characteristics such as peak values, valley values, and fluctuation intensity can be extracted. For example, some users may have large load fluctuations, which may be related to electric vehicle charging behavior or seasonal use of home appliances.

[0047] Based on the user's load fluctuation characteristics, we can infer the home appliances they may be using. For example, if the load curve rises sharply at a certain time, it may indicate that the air conditioner has been turned on; while if the fluctuation is small, it may indicate that the user is using relatively stable appliances (such as electric water heaters), thus obtaining the characteristics of the equipment usage patterns.

[0048] S2, based on the load profile, select a subset of models from the pre-trained model library, perform transfer on the selected models, and output the initial load prediction results;

[0049] In this embodiment, a subset of models is selected from the pre-trained model library based on the load profile, as follows:

[0050] Establish a model library containing various pre-trained models, each of which is trained based on historical load data and user profile data, including the following types of models: LSTM (Long Short-Term Memory Network), XGBoost model, and Transformer model.

[0051] Each user's individual load profile is vectorized into feature vectors, and the similarity between the user load profile features and the feature vectors of each model in the pre-trained model library is calculated using the Manhattan distance method. The similarity between the user load profile features and the features of each model is then output.

[0052] Based on similarity ranking, the top K most similar models are selected to form the initial model subset. ,in It is a load forecasting model, i.e., a pre-trained model. It is a target user profile. Each model is accompanied by its training data feature vector. It is the similarity threshold.

[0053] Furthermore, by vectorizing user load profile features and calculating Manhattan distance similarity with model feature vectors in a pre-trained model library, accurate matching of load characteristics and model capabilities is achieved. Compared to random model selection, this mechanism effectively reduces prediction errors, and is particularly suitable for complex scenarios such as new energy users and highly fluctuating industrial loads.

[0054] In this embodiment, the selected model is transferred to output the initial load forecast results, as follows:

[0055] For each model in the initial model subset, its encoder part (such as the first few layers of a neural network) is used as a shared feature extraction module. At the same time, a target domain adaptation module is constructed. This module consists of a domain discriminator, an output predictor (used to predict the load value based on the extracted shared features), and a transfer layer (used to perform domain alignment and feature space matching operations), forming a transfer model.

[0056] The target user's load profile and historical load data are taken as input and fed into the encoder in the transfer model. The intermediate feature representation vector of the input is extracted and used as the shared feature space representation of the transfer structure.

[0057] A domain discriminator is built based on shared feature representations to determine whether features come from the source domain (model training domain) or the target domain;

[0058] The feature representations of the source domain samples are extracted and their mean differences with the target domain features are calculated in the shared feature space to measure the degree of latent space alignment.

[0059] The initial model subset in the overall transfer model is switched based on the mean difference, and weights are assigned to the selected models based on the score of each model in the similarity calculation.

[0060] The prediction results of the model subset are weighted and averaged according to the weights to output the initial load prediction results.

[0061] The migration loss term is as follows:

[0062]

[0063] In the formula: Let the mean difference of the k-th transfer model be denoted as . Let be the intermediate feature representation extracted from the i-th source domain sample by the encoder of the k-th model. This represents the intermediate features extracted from the j-th target user sample by the encoder of the k-th model. The number of samples in the source domain. This represents the number of samples in the target domain.

[0064] S3, construct a load association graph based on spatial adjacency and behavioral similarity between users, and propagate load features in the load association graph (through GNN);

[0065] In this embodiment, a load correlation graph is constructed based on the spatial adjacency and behavioral similarity between users, as detailed below:

[0066] Based on Geographic Information System (GIS) data or power grid topology data, extract the physical location relationships or distribution network access relationships between users and establish an initial spatial adjacency matrix;

[0067] Within each time window, the weights of the initial spatial adjacency matrix are dynamically adjusted to form a time-varying spatial adjacency matrix;

[0068] Based on algorithms such as Dynamic Time Warping (DTW), user load behavior features are extracted, and the behavior similarity matrix within each time window is calculated.

[0069] Within each time window, the time-varying spatial adjacency matrix and the behavioral similarity matrix are fused using a weighted fusion method to form a joint adjacency matrix, i.e., a load correlation graph.

[0070] The establishment of the initial spatial adjacency matrix is ​​as follows:

[0071] Adjacency relationships are constructed based on the distance between users (such as spatial Euclidean distance or grid topology distance);

[0072] If users are located in the same power distribution area or connected to the same substation, they are given a strong connection relationship;

[0073] If users are far apart, a weaker connection weight is assigned.

[0074] Furthermore, by fusing a time-varying spatial adjacency matrix (based on dynamic adjustment of the power grid topology) and a behavioral similarity matrix (calculated using the DTW algorithm), a joint adjacency matrix is ​​constructed, overcoming the limitations of traditional static association models. Moreover, a spatial adjacency matrix is ​​constructed with electrical coupling strength as the core, accurately quantifying real electrical dependencies and avoiding misjudgments of spatially close but electrically isolated entities.

[0075] Furthermore, by generating a load correlation diagram through weighted fusion, strongly coupled user clusters can be identified, and load groups with similar behaviors can be accurately located, effectively shortening the fault location time.

[0076] Within each time window, the weights of the initial spatial adjacency matrix are dynamically adjusted to form a time-varying spatial adjacency matrix, as detailed below:

[0077] For each pair of users, the time series similarity of the two sets of load data is calculated within the current time window through dynamic time warping, i.e., load similarity. As the time window progresses, the similarity between each pair of users is calculated and updated periodically to ensure that the similarity is continuously updated with changes in load fluctuations.

[0078] The load data of each user is statistically analyzed by the standard deviation algorithm to calculate the load volatility of the user in the current time window. The load volatility of each user is stored as a numerical value or vector as its load volatility characteristic.

[0079] The weights of the initial spatial adjacency matrix are dynamically adjusted based on load volatility and load similarity.

[0080] If a user's load fluctuates drastically within a certain period of time, its correlation with surrounding users should be strengthened, that is, its weight value should be increased.

[0081] If, within a certain time window, a user's load behavior is extremely similar to that of another user, the connection strength between them will be enhanced, i.e., the weight value will increase.

[0082] In this embodiment, load characteristics are propagated in the load correlation graph as follows:

[0083] The load profile features of each user node are mapped to a high-dimensional space as the initial node features. The feature mapping is performed through linear projection or convolution to ensure that the node features have sufficient expressive power.

[0084] Based on the initial node characteristics, the connection strength of each pair of neighboring nodes is dynamically adjusted through a graph self-attention mechanism;

[0085] The adjusted node features are input into the graph neural network for propagation. After several layers of graph convolution, the enhanced feature representation of each user node is obtained.

[0086] The specific formula for dynamically adjusting the connection strength of each pair of neighboring nodes using the graph self-attention mechanism is as follows:

[0087]

[0088] In the formula: It is the attention coefficient between node i and node j. This is the weight matrix. For the features of node i, Let be the features of node j. The features of each node are numerical representations of information such as the node's load behavior, geographical location, and historical data. For attention weight vectors, It is a ReLU function with a small slope correction. This represents the feature vector of neighbor node k. It is the index of the neighboring nodes, representing all the neighboring nodes of node i. It is the set of all neighboring nodes of node i. Let be the new feature of node i after propagation, i.e., the enhanced feature representation, and be the activation function.

[0089] S4. Adjust the model for the target user based on the load characteristics after propagation and the initial load prediction results;

[0090] In this embodiment, the model for the target user is adjusted based on the load characteristics after propagation and the initial load prediction results, as follows:

[0091] The enhanced feature representation and the initial load prediction results are concatenated to form the input feature representation;

[0092] The input feature representation is fed into a pre-trained transfer model. Through incremental learning, the mean squared error between the current prediction result and the actual load is output as a loss function to adjust the parameters in the target user model, including weights and attention coefficients.

[0093] Based on the loss function, the parameters in the transfer model are updated step by step using the gradient descent method, and preliminary load prediction results are output.

[0094] Based on the mean absolute error between the load prediction results before and after updating the transfer model, the learning rate in the gradient descent method is dynamically adjusted to obtain the corrected parameters and update the model.

[0095] The model outputs load prediction results based on the corrected parameters.

[0096] The parameter update model is as follows:

[0097]

[0098] In the formula: These are the updated values ​​of the model parameters. These are the old values ​​of the model parameters. It is a loss function. It is the loss function with respect to the model parameters. gradient, It is the model's prediction result. It is the target true value. This is the updated learning rate. This is the current learning rate. To control the coefficient of adjustment rate, It is the loss of the current step. It's the loss from the previous round. It is the learning rate updated in the previous step, that is, the learning rate actually used when training the current round.

[0099] S5 acquires changes in the distribution of user load data after regulation and identifies sudden behavioral changes;

[0100] In this embodiment, the changes in user load data distribution after regulation are obtained, and behavioral abrupt changes are identified, as follows:

[0101] The acquired user load data (such as hourly load data) is input into the ADWIN drift detection algorithm in chronological order, and the data within each time period is treated as a data window.

[0102] The ADWIN algorithm divides the load data stream into multiple data windows using a sliding window mechanism. Each window contains a certain number of historical data points, and calculates the mean and variance of the load data to obtain the distribution changes of the load data.

[0103] The mean and variance of the load data are mapped to a preset drift threshold. Based on the mapping relationship between the mean and variance of the load data and the preset drift threshold, a dynamic drift threshold is output.

[0104] When ADWIN detects that the difference in mean and variance between the distribution of the current window data and the previous window data exceeds the drift threshold, a drift alarm is issued.

[0105] When a distribution drift is detected, an incremental learning mechanism is triggered, and the system reverts to the parameter update model to update the parameters.

[0106] The specific dynamic drift threshold is as follows:

[0107]

[0108] In the formula: It is the dynamic drift threshold. and It is a hyperparameter for adjusting sensitivity. Let Variance be the variance of the k-th data window. This represents the total number of historical data windows. This is the average of all historical data windows.

[0109] Furthermore, by mapping the mean and variance statistics of load data to a preset drift threshold in real time, a dynamic drift threshold is generated, overcoming the deficiency of insufficient sensitivity of traditional fixed thresholds in complex power consumption scenarios. Compared with single-index detection, this design significantly improves the adaptability to non-stationary load data such as seasonal fluctuations and abnormal power consumption.

[0110] Furthermore, the sliding window mechanism using the ADWIN algorithm only needs to retain historical data points of the current window for mean and variance calculations, eliminating the need for full data storage and backtracking. This reduces computational complexity from O(N²) to O(N), making it possible to achieve second-level response times on edge computing devices (such as smart meters), thus meeting the real-time monitoring needs of the power Internet of Things.

[0111] Example 2, Figure 2 The present invention provides a power distribution load forecasting system, comprising a data acquisition module, an initial forecasting module, a propagation module, a control module, and an identification module, wherein the modules are interconnected.

[0112] The data acquisition module is used to acquire user load data and auxiliary characteristics to build load profiles;

[0113] The initial prediction module is used to select a subset of models from the pre-trained model library based on the load profile, perform transfer on the selected models, and output the initial load prediction results.

[0114] The propagation module is used to construct a load association graph based on the spatial adjacency and behavioral similarity between users, and to propagate load features in the load association graph.

[0115] The control module is used to control the model of the target user based on the load characteristics after propagation and the initial load prediction results;

[0116] The identification module is used to acquire changes in the distribution of user load data after regulation and to identify sudden behavioral changes.

[0117] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0118] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0119] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0120] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0121] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0122] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting power distribution load, characterized in that, include: Acquire user load data and auxiliary characteristics to construct load profiles; Based on the load profile, a subset of models is selected from the pre-trained model library, and the selected models are transferred to output the initial load prediction results. A load correlation graph is constructed based on the spatial adjacency and behavioral similarity between users, and load characteristics are propagated in the load correlation graph; Based on the load characteristics after propagation and the initial load prediction results, the model for the target user is adjusted. Acquire changes in user load data distribution after regulation to identify sudden behavioral changes.

2. The power distribution load forecasting method according to claim 1, characterized in that, The process of acquiring user load data and auxiliary features to construct a load profile is as follows: The collected load data is cleaned and processed; Obtain user auxiliary characteristics and combine them with load data to extract load characteristics; Load feature vectors are constructed based on load characteristics, and then these load feature vectors are merged to form a load profile.

3. The power distribution load forecasting method according to claim 2, characterized in that, The process of selecting a subset of models from the pre-trained model library based on the load profile is as follows: Establish a model library containing various pre-trained models; Each user's individual load profile is vectorized into feature vectors, and the similarity between the user load profile features and the feature vectors of each model in the pre-trained model library is calculated using the Manhattan distance method. The similarity between the user load profile features and the features of each model is then output. Based on the similarity ranking, the top K most similar models are selected to form the initial model subset.

4. The power distribution load forecasting method according to claim 3, characterized in that, The selected model is then transferred to output the initial load forecast results, as detailed below: For each model in the initial model subset, the encoder part is used as a shared feature extraction module, and a target domain adaptation module is constructed. The target domain adaptation module consists of a domain discriminator, an output predictor, and a transfer layer to form a transfer model. The target user's load profile and historical load data are input into the encoder of the migration model, and the intermediate feature expression vector of the input is extracted as the shared feature space representation of the migration structure. A domain discriminant is constructed based on shared feature representations to identify source domain features and target domain features; Extract the feature representations of the source domain samples and the features of the target domain, and output the mean difference in the shared feature space; The initial model subset in the overall transfer model is switched based on the mean difference, and weights are assigned to the selected models based on the scores of each model in the similarity calculation. The prediction results of the model subset are weighted and averaged according to the weights to output the initial load prediction results.

5. The power distribution load forecasting method according to claim 4, characterized in that, The load correlation graph is constructed based on the spatial adjacency and behavioral similarity between users, as detailed below: Based on power grid topology data, the distribution network access relationship between each user is extracted, and an initial spatial adjacency matrix is ​​established; Within each time window, the weights of the initial spatial adjacency matrix are dynamically adjusted to form a time-varying spatial adjacency matrix; Based on the dynamic time warping algorithm, user load behavior features are extracted, and the behavior similarity matrix within each time window is calculated. Within each time window, the time-varying spatial adjacency matrix and the behavioral similarity matrix are fused using a weighted fusion method to form a joint adjacency matrix, i.e., a load correlation graph.

6. The power distribution load forecasting method according to claim 5, characterized in that, Within each time window, the weights of the initial spatial adjacency matrix are dynamically adjusted to form a time-varying spatial adjacency matrix, as detailed below: For each pair of users, the load similarity between the two sets of load data is calculated within the current time window through dynamic time warping. The load data of each user is statistically analyzed using the standard deviation algorithm to calculate the load fluctuation of the user within the current time window; The weights of the initial spatial adjacency matrix are dynamically adjusted based on load volatility and load similarity. If a user's load fluctuates beyond a preset threshold within a certain period of time, the correlation with surrounding users should be enhanced, i.e., the weight value should be increased. If, within a certain time window, the load similarity between a user and another user exceeds a preset threshold, the connection strength between them will be enhanced, i.e., the weight value will increase.

7. The power distribution load forecasting method according to claim 6, characterized in that, The propagation of load characteristics in the load correlation graph is specifically as follows: The load profile features of each user node are mapped to a high-dimensional space as the initial node features; Based on the initial node characteristics, the connection strength of each pair of neighboring nodes is dynamically adjusted through a graph self-attention mechanism; The adjusted node features are input into the graph neural network for propagation. After several layers of graph convolution, the enhanced feature representation of each user node is obtained.

8. The power distribution load forecasting method according to claim 7, characterized in that, The model for the target user is adjusted based on the load characteristics after propagation and the initial load prediction results, as follows: The enhanced feature representation and the initial load prediction results are concatenated to form the input feature representation; The input feature representation is fed into a pre-trained transfer model, and the mean square error between the current prediction result and the actual load is output as the loss function through incremental learning. Based on the loss function, the parameters in the transfer model are updated step by step using the gradient descent method, and preliminary load prediction results are output. Based on the mean absolute error between the load prediction results before and after updating the transfer model, the learning rate in the gradient descent method is dynamically adjusted to obtain the corrected parameters and update the model. The model outputs load prediction results based on the corrected parameters.

9. The power distribution load forecasting method according to claim 8, characterized in that, The acquisition of changes in user load data distribution after regulation and the identification of behavioral abrupt changes are as follows: The acquired user load data is input into the ADWIN drift detection algorithm in time series order; The ADWIN algorithm divides the load data stream into multiple data windows using a sliding window mechanism and calculates the mean and variance of the load data. The mean and variance of the load data are mapped to a preset drift threshold. Based on the mapping relationship between the mean and variance of the load data and the preset drift threshold, a dynamic drift threshold is output. When ADWIN detects that the difference in mean and variance between the distribution of the current window data and the previous window data exceeds the drift threshold, a drift alarm is issued. When a distribution drift is detected, an incremental learning mechanism is triggered, and the system reverts to the parameter update model to update the parameters.

10. A power distribution load forecasting system using the power distribution load forecasting method as described in any one of claims 1-9, characterized in that, It includes a data acquisition module, an initial prediction module, a propagation module, a control module, and an identification module, and the modules are interconnected. The data acquisition module is used to acquire user load data and auxiliary characteristics to build load profiles; The initial prediction module is used to select a subset of models from the pre-trained model library based on the load profile, perform transfer on the selected models, and output the initial load prediction results. The propagation module is used to construct a load association graph based on the spatial adjacency and behavioral similarity between users, and to propagate load features in the load association graph. The control module is used to control the model of the target user based on the load characteristics after propagation and the initial load prediction results; The identification module is used to acquire changes in the distribution of user load data after regulation and to identify sudden behavioral changes.

Images