Power distribution load prediction method and system
By constructing a load association graph and transfer learning, combining the spatial adjacency and behavioral similarity of users, and dynamically adjusting the prediction model, the problems of inaccurate prediction and poor adaptability in traditional methods are solved, and high-precision distribution load prediction is achieved.
Patent Information
- Application Number
- CN202511247473.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-03
- Publication Date
- 2025-10-10
- Estimated Expiration
- 2045-09-03
AI Technical Summary
Traditional distribution load forecasting methods rely on a single global model or a fixed model, which fails to consider the individual differences of different users, resulting in inaccurate prediction results and a lack of dynamic adaptability to load fluctuations and unstable electricity consumption behavior.
By constructing a load association graph, combining the spatial adjacency and behavioral similarity between users, using graph neural networks to propagate load features, and selecting a suitable pre-trained model subset through transfer learning, the prediction model is dynamically adjusted to adapt to changes in user behavior.
It significantly improves the accuracy and flexibility of distribution load forecasting, can adapt to changes in user behavior in real time, reduces single model errors, and improves the accuracy and personalized characteristics of forecasting.
Smart Images

Figure CN120763590A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of graph computing technology, and more specifically, to a distribution load forecasting method and system. Background Art
[0002] As electricity demand becomes increasingly complex and diverse, load forecasting for distribution systems has become a core issue in power dispatch and management. The accuracy of distribution load forecasts directly impacts the grid's operational efficiency, stability, and optimal resource allocation. Traditional load forecasting methods often rely on statistical models or basic machine learning models, but these methods exhibit low accuracy and adaptability when faced with highly complex and nonlinear load fluctuations. Therefore, improving load forecasting accuracy, dynamic adaptability, and responsiveness to sudden load fluctuations has become a critical technical challenge that 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 the behavior portrait of multiple users. The method is used to predict the electricity load of multiple users, including the following steps: selecting historical electricity load data of several multiple users and preprocessing the historical electricity load data of the multiple users; training the preprocessed historical electricity load data for different types of multiple users to obtain corresponding prediction models, wherein the different types of multiple users include users with poor electricity stability, users with good electricity stability, and users with low temperature sensitivity; predicting the electricity load of multiple users to obtain prediction results, and displaying the prediction results. It is possible to adopt corresponding prediction methods for different types of multiple users, thereby achieving rapid and accurate prediction of the power system load.
[0004] For example, the invention patent with announcement number CN118523291A discloses a neural network-based low-voltage substation load forecasting method, which specifically includes the following steps: Step 1, collecting historical load data of the low-voltage substation; 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 a preliminary prediction value; Step 4, performing error and pass rate analysis on the preliminary prediction value. If the pass rate meets the requirements, the preliminary prediction value is identified as the final prediction value. If the pass rate does not meet the requirements, return to step 3 and obtain a new preliminary prediction value. By combining multiple algorithms, the problems of slow low-voltage substation load forecasting process and low prediction accuracy are solved.
[0005] The above disclosed technical solutions have at least the following technical problems: Traditional distribution load forecasting methods rely on a single global or fixed model, failing to account for individual differences among users. This often results in inaccurate or uncustomized forecasts for users with varying load demands. Furthermore, traditional methods often use static load correlation models, ignoring the dynamic nature of load changes over time. For users with large load fluctuations and erratic electricity usage, the lack of targeted adjustments leads to reduced forecast accuracy.
[0006] In view of the above problems, the present invention proposes a solution. Summary of the Invention
[0007] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides a distribution load forecasting method and system, which solves the problems of low accuracy, poor dynamic adaptability and slow response to sudden changes in traditional methods through load correlation graph construction and drift detection.
[0008] To achieve the above object, the present invention provides the following technical solutions: A distribution load forecasting method includes: obtaining user load data and auxiliary features to construct a load profile; selecting a model subset from a pre-trained model library based on the load profile, migrating the selected model, and outputting an initial load forecast result; constructing a load association graph based on spatial adjacency and behavioral similarity between users, and propagating load features in the load association graph; regulating the model of the target user based on the propagated load features combined with the initial load forecast result; obtaining changes in the distribution of user load data after regulation and identifying behavioral mutations.
[0009] In a preferred embodiment, the method of obtaining the user's load data and auxiliary features and constructing a load portrait is as follows: cleaning and processing the collected load data; obtaining the user's auxiliary features and extracting the load features in combination with the load data; constructing a load feature vector based on the load features, and merging the load feature vectors to form a load portrait.
[0010] In a preferred embodiment, the method of selecting a model subset from a pre-trained model library based on the load profile is as follows: establishing a model library containing multiple pre-trained models; quantizing the features of each user's individual load profile, and calculating the similarity with each model feature vector in the pre-trained model library using the Manhattan distance method, and outputting the similarity between the user load profile features and the features of each model; and selecting the top K most similar models based on the similarity ranking to form an initial model subset.
[0011] In a preferred embodiment, the selected model is migrated and the initial load prediction result is output, 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 at the same time. The target domain adaptation module includes a domain discriminator, an output predictor and a migration layer to form a migration model; the load profile of the target user and the historical load data are input into the encoder in the migration model, and the input intermediate feature expression vector is extracted as the shared feature space representation of the migration structure; a domain discriminator is constructed based on the shared feature representation to identify source domain features and target domain features; the feature representation of the source domain sample and the target domain feature are extracted, and the mean difference in the shared feature space is output; the initial model subset in the overall migration model is switched according to the mean difference, and the selected model is assigned a weight according to the score of each model in the similarity calculation; the prediction results of the model subset are weighted averaged according to the weight, and the initial load prediction result is output.
[0012] In a preferred embodiment, the load association graph is constructed based on the spatial adjacency and behavioral similarity between users, specifically as follows: based on the 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 weight of the initial spatial adjacency matrix is dynamically adjusted to form a time-varying spatial adjacency matrix; based on the dynamic time warping algorithm, the 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 by weighted fusion to form a joint adjacency matrix, i.e., the load association graph.
[0013] 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, 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 through a 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 according to the load volatility and load similarity; if the user load fluctuates violently within a certain period of time, the correlation with the surrounding users should be enhanced, that is, the weight value increases; if within a certain time window, the load behavior of the user and other users is extremely similar, the connection strength between them will be enhanced, that is, the weight value increases.
[0014] In a preferred embodiment, the load feature propagation in the load association graph is specifically as follows: the load portrait feature of each user node is mapped to a high-dimensional space as the initial node feature; based on the initial node feature, the connection strength of each pair of neighbor nodes is dynamically adjusted through the graph self-attention mechanism; the adjusted node feature is input into the graph neural network for propagation, and after the graph neural network passes through several layers of graph convolution, an enhanced feature representation of each user node is obtained.
[0015] In a preferred embodiment, the model of the target user is regulated based on the load characteristics after propagation combined with the initial load forecast results, specifically as follows: the enhanced feature representation and the initial load forecast results are spliced to form an input feature representation; the input feature representation is input into a pre-trained migration model, and the mean square error between the current prediction result and the actual load is output as a loss function through an incremental learning method; according to the loss function, the parameters in the migration model are gradually updated based on the gradient descent method, and a preliminary load forecast result is output; according to the mean absolute error between the load forecast results before and after the migration model is updated, the learning rate in the gradient descent method is dynamically adjusted to obtain a revised parameter update model; and the load forecast result is output according to the revised parameter update model.
[0016] In a preferred embodiment, the changes in the distribution of user load data after acquisition and regulation and the identification of behavioral mutations are specifically as follows: the acquired user load data is input into the ADWIN drift detection algorithm in time series; 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 mean and variance differences between the distribution of the current window data and the previous window data exceed the drift threshold, a drift alarm is issued; when distribution drift is detected, the incremental learning mechanism is triggered, and the parameter update model is returned to update the parameters.
[0017] A distribution load forecasting system includes a data acquisition module, an initial forecasting module, a propagation module, a regulation module and an identification module, and there are connections between the modules; the data acquisition module is used to obtain user load data and auxiliary features and construct a load portrait; the initial forecasting module is used to select a model subset from a pre-trained model library based on the load portrait, migrate the selected model, and output an initial load forecast result; the propagation module is used to construct a load association graph based on spatial adjacency and behavioral similarity between users, and propagate load characteristics in the load association graph; the regulation module is used to regulate the model of the target user based on the propagated load characteristics combined with the initial load forecast result; the identification module is used to obtain the distribution changes of user load data after regulation and identify behavioral mutations.
[0018] The technical effects and advantages of the power distribution load forecasting method and system of the present invention are as follows: 1. This invention utilizes multi-model integration, combining the advantages of different models to effectively reduce the potential errors introduced by a single model. When selecting an appropriate pre-trained model subset, the similarity between the load profile and the model is calculated to ensure that the model most appropriate for the current user's load characteristics is selected. This personalized model selection and weighted integration approach significantly improves the accuracy of distribution load forecasting.
[0019] 2. By introducing a load association graph, this invention not only better captures dynamic changes in user load based on spatial adjacency and behavioral similarity between users, but also propagates and enhances load characteristics using graph neural networks. 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 prediction accuracy and flexibility. BRIEF DESCRIPTION OF THE DRAWINGS
[0020] Figure 1 The figure is a flow chart of a method for predicting power distribution load according to the present invention.
[0021] Figure 2 The figure is a schematic diagram of the structure of a power distribution load forecasting system of the present invention. DETAILED DESCRIPTION
[0022] The following will provide a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.
[0023] Example 1, Figure 1 The present invention provides a distribution load forecasting method, which includes the following steps: S1, obtains the user's load data and auxiliary features and builds a load profile; In this embodiment, the user's load data and auxiliary features are obtained to construct a load profile, as follows: Collect user load data in real time through smart meters or high-precision measurement equipment. The collected load data includes: instantaneous power (load curve), historical electricity consumption data (including user's historical load data, with time granularity selectable at hourly, daily, or higher frequency levels), and load peak and valley values. Clean and process the collected load data, identify and correct abnormal data using methods such as the isolation forest algorithm or DBSCAN, and denoise the load fluctuation data using wavelet transform to ensure data smoothness; Obtain auxiliary features of the user, the auxiliary features including: User category features: Classified according to user type (residential, commercial, industrial, etc.), marked as static features; Geographic information features: The user's geographic location is obtained through the geographic information system, combined with environmental factors such as regional temperature and humidity to supplement the electricity consumption prediction model; Environmental change characteristics: Real-time collection of weather information in the user's area to identify its impact on load, such as the peak air conditioning load in summer; User behavior pattern characteristics: Based on the user's historical load data, extract the user's periodic power consumption pattern characteristics, including peak hours, holiday fluctuations, etc.; Extracting load features based on load data and auxiliary features, wherein the load features include: time features, spatial features, and behavioral features; A load feature vector is constructed based on the load characteristics, and the load feature vectors are merged to form a high-dimensional user portrait vector, namely the load portrait.
[0024] The load characteristics are extracted based on the load data and the auxiliary characteristics, as follows: The time characteristics include periodic characteristics and peak and valley periods; The spatial characteristics include geographical location influence and regional similarity characteristics; The behaviors include load fluctuation characteristics and equipment usage pattern characteristics; Based on historical load data, frequency domain analysis is performed on the load data through Fourier transform or wavelet transform to extract daily, weekly, monthly and other periodic features; Based on historical load data, identify peak hours (such as morning or evening peak hours) and off-peak hours for users' electricity consumption; Use cluster analysis (such as the K-means algorithm) to identify the geographical characteristics of users, namely regional similarity characteristics, which reflect the similarity in electricity usage patterns of users in the same geographical location; Geographic influence: By combining user location information (e.g., city, region, building level, etc.) with environmental data (e.g., temperature, humidity, etc.), spatial features are extracted. For example, users in warm regions may use more electricity for air conditioning, while users in cold regions may use more electricity for electric heating equipment. By analyzing the volatility of user load curves, we can extract load fluctuation characteristics such as peak values, valley values, and fluctuation intensity. For example, large load fluctuations for some users may be related to electric vehicle charging behavior or seasonal appliance usage. Based on the user's load fluctuation characteristics, we can infer the home appliances they may use. For example, if the load curve rises sharply in a certain period of time, it may indicate that the air conditioner is turned on; if the fluctuation is small, it may mean that the user is using a more stable appliance (such as an electric water heater), thus obtaining the device usage pattern characteristics.
[0025] S2, based on the load profile, selects a model subset from the pre-trained model library, migrates the selected model, and outputs the initial load forecast result; In this embodiment, a model subset is selected from the pre-trained model library based on the load profile, as follows: Establish a model library containing various pre-trained models, each of which is trained based on historical load data and user profile data. The library includes the following types of models: LSTM (Long Short-Term Memory Network), XGBoost model, and Transformer model. Each user's individual load profile is feature-vectorized and similarity is calculated with each model feature vector in the pre-trained model library using the Manhattan distance method. The similarity between the user's load profile features and each model feature is output; According to the similarity ranking, select the top K most similar models to form the initial model subset ,in It is a load forecasting model, i.e. a pre-trained model. It is the target user portrait, is the feature vector of each model accompanied by its training data, is the similarity threshold.
[0026] Furthermore, by quantizing user load profile features and calculating Manhattan distance similarity with model feature vectors in a pre-trained model library, we achieve a precise match between load characteristics and model capabilities. Compared to random model selection, this mechanism effectively reduces prediction error and is particularly suitable for complex scenarios such as new energy users and highly fluctuating industrial loads.
[0027] In this embodiment, the selected model is migrated and the initial load forecast results are output as follows: For each model in the initial model subset, its encoder part (such as the first few layers of the 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 (for load value prediction based on the extracted shared features), and a transfer layer (for performing domain alignment and feature space matching operations) to form a transfer model. The target user's load profile and historical load data are taken as input and passed to the encoder in the transfer model. The intermediate feature expression vector of the input is extracted as the shared feature space representation of the transfer structure. Build a domain discriminator based on the shared feature representation to determine whether the feature comes from the source domain (model training domain) or the target domain; Extract the feature representation of the source domain sample and the target domain feature and calculate their mean difference in the shared feature space to measure the degree of latent space alignment; Switch the initial model subset in the overall migration model based on the mean difference, and assign weights to the selected models based on the scores of each model in the similarity calculation; The prediction results of the model subsets are weighted averaged according to the weights and the initial load prediction results are output.
[0028] The migration loss items are as follows:
[0029] Where: is the mean difference of the kth migration model, is the intermediate feature representation extracted by the encoder of the k-th model for the i-th source domain sample, is the intermediate feature representation extracted by the encoder of the k-th model for the j-th target user sample, is the number of source domain samples, is the number of samples in the target domain.
[0030] S3, builds a load association graph based on the spatial adjacency and behavioral similarity between users, and propagates load features in the load association graph (via GNN); In this embodiment, a load association graph is constructed based on the spatial adjacency and behavioral similarity between users, as follows: Based on geographic information system (GIS) data or power grid topology data, the physical location relationship between users or the distribution network access relationship is extracted to establish an initial spatial adjacency matrix; In each time window, the weights of the initial spatial adjacency matrix are dynamically adjusted to form a time-varying spatial adjacency matrix; Based on algorithms such as Dynamic Time Warping (DTW), user load behavior characteristics are extracted and the behavior similarity matrix within each time window is calculated; In each time window, the time-varying spatial adjacency matrix and the behavioral similarity matrix are fused by weighted fusion to form a joint adjacency matrix, namely the load association graph.
[0031] The initial spatial adjacency matrix is established as follows: Constructing adjacency relationships through the distance between users (e.g., spatial Euclidean distance or grid topological distance); If users are located in the same distribution area or connected to the same substation, give them a stronger connection relationship; If the distance between users is far, a weaker connection weight is assigned.
[0032] Furthermore, by fusing a time-varying spatial adjacency matrix (based on dynamic adjustments to the grid topology) with a behavioral similarity matrix (calculated using a rolling DTW algorithm), a joint adjacency matrix is constructed, breaking through the limitations of traditional static association models. Furthermore, the spatial adjacency matrix, centered around electrical coupling strength, accurately quantifies true electrical dependencies and avoids misjudgments of spatially proximal but electrically isolated components.
[0033] Furthermore, load correlation diagrams are generated through weighted fusion, strongly coupled user clusters are identified, and load groups with similar behaviors are accurately located, effectively shortening the fault location time.
[0034] In each time window, the weights of the initial spatial adjacency matrix are dynamically adjusted to form a time-varying spatial adjacency matrix, as follows: For each pair of users, dynamic time warping is used to calculate the time series similarity of the two sets of load data within the current time window, namely the load similarity. As the time window progresses, the similarity between each pair of users is regularly calculated and updated to ensure that the similarity is continuously updated as the load fluctuations change. The load data of each user is statistically analyzed using the standard deviation algorithm to calculate the load fluctuation of the user in the current time window. The load fluctuation of each user is stored as a numerical value or vector as its load fluctuation feature. Dynamically adjust the weights of the initial spatial adjacency matrix according to load volatility and load similarity; If the user load fluctuates violently within a certain period, its correlation with surrounding users should be enhanced, that is, the weight value should be increased; If within a certain time window, the load behavior of a user is very similar to that of other users, the connection strength between them will be enhanced, that is, the weight value will increase.
[0035] In this embodiment, load characteristics are propagated in the load correlation graph as follows: Map the load profile features of each user node to a high-dimensional space as the initial node features. Feature mapping is performed through linear projection or convolution to ensure that the node features have sufficient expressive power. According to the initial node features, the connection strength of each pair of neighbor nodes is dynamically adjusted through the graph self-attention mechanism; The adjusted node features are input into the graph neural network for propagation. After the graph neural network passes through several layers of graph convolution, the enhanced feature representation of each user node is obtained.
[0036] The formula for dynamically adjusting the connection strength of each pair of neighbor nodes through the graph self-attention mechanism is as follows:
[0037] Where: is the attention coefficient between node i and node j, is the weight matrix, is the feature of node i, is the feature of node j. The feature of each node is the numerical representation of the node's load behavior, geographical location, historical data and other information. is the attention weight vector, is a ReLU function with a small correction of the slope, represents the feature vector of neighbor node k, is the index of the neighbor node, representing all neighbor nodes of node i, is the set of all neighbor nodes of node i, is the new feature of node i after propagation, that is, the enhanced feature representation, and is the activation function.
[0038] S4, adjusts the model of the target user based on the load characteristics after propagation and the initial load forecast results; In this embodiment, the model of the target user is regulated based on the propagated load characteristics combined with the initial load forecast results, as follows: The enhanced feature representation and the initial load forecast result are concatenated to form the input feature representation; Input feature representations are fed into a pre-trained transfer model. Using an incremental learning method, the mean squared error between the current prediction result and the actual load is output as a loss function to adjust parameters in the target user model, including weights and attention coefficients. According to the loss function, the parameters in the migration model are gradually updated based on the gradient descent method, and the preliminary load forecast results are output; According to the mean absolute error between the load forecast results before and after updating the migration model, the learning rate in the gradient descent method is dynamically adjusted to obtain a revised parameter update model; The model is updated based on the revised parameters and the load forecast results are output.
[0039] The parameter update model is specifically as follows:
[0040] Where: is the updated value of the model parameters, are the old values of the model parameters, is the loss function, is the loss function with respect to the model parameters The gradient, is the prediction result of the model, is the target true value, is the updated learning rate, is the current learning rate, To control the coefficient of the adjustment rate, is the loss of the current step, It's the loss from the previous round. It is the learning rate after the previous step update, that is, the learning rate actually used in training the current round.
[0041] S5, obtain the distribution change of user load data after regulation and identify behavioral mutations; In this embodiment, the distribution change of user load data after regulation is obtained and behavioral mutations are identified, specifically as follows: The acquired user load data (such as hourly load data) is input into the ADWIN drift detection algorithm in chronological order, and the data in each time period is used as a data window; The ADWIN algorithm divides the load data stream into multiple data windows through a sliding window mechanism. Each window contains a certain number of historical data points. The mean and variance of the load data are calculated to obtain the distribution change of the load data. Mapping the mean and variance of the load data to a preset drift threshold, and outputting a dynamic drift threshold based on the mapping relationship between the mean and variance of the load data and the preset drift threshold; When ADWIN detects that the mean and variance differences between the distribution of the current window data and the previous window data exceed the drift threshold, a drift alarm is issued; When distribution drift is detected, the incremental learning mechanism is triggered and the parameters are updated in the parameter update model.
[0042] The dynamic drift threshold is specifically as follows:
[0043] Where: is the dynamic drift threshold, and is a hyperparameter that adjusts the sensitivity, is the variance of the kth data window, is the total number of historical data windows, is the mean of all historical data windows.
[0044] Furthermore, by mapping the load data's mean and variance statistics to preset drift thresholds in real time, a dynamic drift threshold is generated. This overcomes the lack of sensitivity of traditional fixed thresholds in complex power usage scenarios. Compared to single-metric detection, this design significantly improves its adaptability to non-stationary load data, such as seasonal fluctuations and abnormal power consumption.
[0045] Furthermore, the sliding window mechanism of the ADWIN algorithm only retains historical data points within the current window for mean and variance calculation, eliminating the need for full data storage and backtracking. This reduces computational complexity from O(N²) to O(N), enabling second-level response times on edge computing devices such as smart meters, meeting the real-time monitoring needs of the power IoT.
[0046] Example 2, Figure 2 The present invention provides a distribution load forecasting system, which includes a data acquisition module, an initial forecasting module, a propagation module, a control module and an identification module, and there are connections between the modules; Data collection module, used to obtain user load data and auxiliary features and build load profiles; The initial prediction module is used to select a model subset from the pre-trained model library based on the load profile, migrate the selected model, and output the initial load forecast result; The propagation module is used to construct a load association graph based on the spatial adjacency and behavioral similarity between users and propagate load features in the load association graph; The control module is used to control the target user's model based on the propagated load characteristics and the initial load forecast results; The identification module is used to obtain the distribution changes of user load data after regulation and identify behavioral mutations.
[0047] The above formulas are all dimensionless and numerical calculations. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters in the formulas are set by technicians in this field according to actual conditions.
[0048] The above embodiments may be implemented in whole or in part through software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments may be implemented in whole or in part in the form of a computer program product.
[0049] Those skilled in the art will appreciate that the modules and algorithm steps of each example 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 performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel 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.
[0050] In addition, each functional module in each embodiment of the present application may be integrated into one processing module, or each module may exist physically separately, or two or more modules may be integrated into one module.
[0051] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.
[0052] Finally: 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 in the scope of protection of the present invention.
Claims
1. A method for predicting power distribution load, characterized in that: include: Obtain user load data and auxiliary features to build a load profile; Based on the load profile, a model subset is selected from the pre-trained model library, and the selected model is migrated to output the initial load forecast results; Construct a load association graph based on the spatial adjacency and behavioral similarity between users, and propagate load features in the load association graph; According to the load characteristics after propagation and the initial load forecast results, the model of the target user is regulated; Obtain changes in user load data distribution after regulation and identify behavioral mutations.
2. A distribution load forecasting method according to claim 1, characterized in that: The user's load data and auxiliary features are obtained to construct a load profile, specifically as follows: Clean and process the collected load data; Obtain the user's auxiliary features and combine them with load data to extract load features; A load feature vector is constructed according to the load characteristics, and the load feature vectors are merged to form a load profile.
3. A distribution load forecasting method according to claim 2, characterized in that: According to the load profile, a model subset is selected from the pre-trained model library, as follows: Establish a model library containing multiple pre-trained models; Each user's individual load profile is feature-vectorized and similarity is calculated with each model feature vector in the pre-trained model library using the Manhattan distance method. The similarity between the user's load profile features and each model feature is output; According to the similarity ranking, the top K most similar models are selected to form the initial model subset.
4. A distribution load forecasting method according to claim 3, characterized in that: The selected model is migrated and the initial load forecast result is output 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 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; Build a domain discriminator based on shared feature representation to identify source domain features and target domain features; Extract the feature representation of the source domain sample and the target domain features, and output the mean difference in the shared feature space; Switch the initial model subset in the overall migration model based on the mean difference, and assign weights to the selected models based on the score of each model in the similarity calculation; The prediction results of the model subsets are weighted averaged according to the weights and the initial load prediction results are output.
5. A distribution load forecasting method according to claim 4, characterized in that: The load association graph is constructed based on the spatial adjacency and behavioral similarity between users, as follows: Based on the power grid topology data, the distribution network access relationship between each user is extracted and the initial spatial adjacency matrix is established; In 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 characteristics are extracted and the behavior similarity matrix within each time window is calculated; In each time window, the time-varying spatial adjacency matrix and the behavioral similarity matrix are fused by weighted fusion to form a joint adjacency matrix, namely the load association graph.
6. A distribution load forecasting method according to claim 5, characterized in that: In each time window, the weights of the initial spatial adjacency matrix are dynamically adjusted to form a time-varying spatial adjacency matrix, 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; Perform statistical analysis on each user's load data using the standard deviation algorithm to calculate the user's load fluctuation within the current time window; Dynamically adjust the weights of the initial spatial adjacency matrix according to load volatility and load similarity; If the load volatility of a user exceeds a preset threshold within a certain period, the correlation with surrounding users should be enhanced, that is, the weight value increases; If the load similarity between a user and other users exceeds a preset threshold within a certain time window, the connection strength between them will be enhanced, that is, the weight value will increase.
7. A distribution load forecasting method according to claim 6, characterized in that: The load characteristics propagation in the load correlation diagram are as follows: Map the load profile features of each user node to a high-dimensional space as the initial node features; According to the initial node features, the connection strength of each pair of neighbor nodes is dynamically adjusted through the graph self-attention mechanism; The adjusted node features are input into the graph neural network for propagation. After the graph neural network passes through several layers of graph convolution, the enhanced feature representation of each user node is obtained.
8. A distribution load forecasting method according to claim 7, characterized in that: The model of the target user is regulated based on the propagated load characteristics combined with the initial load forecast results, as follows: The enhanced feature representation and the initial load forecast result are concatenated to form the input feature representation; The input feature representation is input into the pre-trained transfer model, and through the incremental learning method, the mean square error between the current prediction result and the actual load is output as the loss function; According to the loss function, the parameters in the migration model are gradually updated based on the gradient descent method, and the preliminary load forecast results are output; According to the mean absolute error between the load forecast results before and after updating the migration model, the learning rate in the gradient descent method is dynamically adjusted to obtain a revised parameter update model; The model is updated based on the revised parameters and the load forecast results are output.
9. A distribution load forecasting method according to claim 8, characterized in that: The method of obtaining the distribution change of user load data after regulation and identifying behavioral mutations is as follows: The acquired user load data is input into the ADWIN drift detection algorithm in time series; 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; Mapping the mean and variance of the load data to a preset drift threshold, and outputting a dynamic drift threshold based on the mapping relationship between the mean and variance of the load data and the preset drift threshold; When ADWIN detects that the mean and variance differences between the distribution of the current window data and the previous window data exceed the drift threshold, a drift alarm is issued; When distribution drift is detected, the incremental learning mechanism is triggered and the parameters are updated in the parameter update model.
10. A power distribution load forecasting system using a power distribution load forecasting method according to any one of claims 1 to 9, characterized in that: It includes data acquisition module, initial prediction module, propagation module, regulation module and identification module, and there are connections between modules; Data collection module, used to obtain user load data and auxiliary features and build load profiles; The initial prediction module is used to select a model subset from the pre-trained model library based on the load profile, migrate the selected model, and output the initial load forecast result; The propagation module is used to construct a load association graph based on the spatial adjacency and behavioral similarity between users and propagate load features in the load association graph; The control module is used to control the target user's model based on the propagated load characteristics and the initial load forecast results; The identification module is used to obtain the distribution changes of user load data after regulation and identify behavioral mutations.
Citation Information
Patent Citations
Power consumption load prediction method based on multivariate user behavior portrait
CN111582548A
Low-voltage transformer area load prediction method based on neural network
CN118523291A
Load prediction method based on data enhancement strategy and residual error self-correction learning model
CN115936176A
Day-ahead power load prediction method and device, terminal and storage medium
CN116205359A
Household ultra-short-term load prediction method
CN120561623A
Cited By
Active load prediction method and device for low-voltage distribution transformer area
CN120955651A
