Zone area load prediction system based on multi-factor model game

Abstract

The invention relates to a transformer area load prediction system based on a multi-factor model game, and the system comprises a data fusion module which is used for accessing and processing multi-source heterogeneous data, and carrying out the processing of the multi-source heterogeneous data, so as to obtain power utilization related information; a factor generation module; a user configuration module; the model game module is configured with a power utilization game strategy, and the power utilization game strategy is configured in a corresponding micro-grid; and the load calculation module is used for calculating the electricity consumption prediction load of the transformer area according to the comprehensive electricity consumption load of each micro-grid. Multi-source heterogeneous data are integrated, and dynamic game characteristics of user behaviors are described through dynamic factor generation, user revenue function configuration and game-driven behavior evolution, so that the defects that a traditional method is static and lacks user decision modeling are overcome. Refined prediction of transformer area load is realized, prediction adaptability and behavior interpretability are improved, and reliable support is provided for operation, planning and demand side response of a distribution network.

Description

Technical Field

[0001] This invention relates to power grid load forecasting systems, and more specifically, to a distribution area load forecasting system based on a multi-factor model game. Background Technology

[0002] With the deepening of the construction of new power systems, distribution substations, as key units at the end of the distribution network, directly affect the efficiency of distribution network operation, the effectiveness of equipment investment planning, and demand-side response strategies due to their load forecasting accuracy. Traditional load forecasting methods are mostly based on statistical modeling or machine learning fitting of historical load data and external environmental factors, which makes it difficult to fully characterize the dynamic game characteristics of user behavior under the influence of multi-source information. Especially in scenarios with a high proportion of distributed energy access and flexible interaction of user-side resources, existing methods have limitations such as insufficient forecasting accuracy, poor adaptability, and lack of modeling of user decision-making mechanisms.

[0003] A power grid load forecasting method, with publication number CN120200224B, constructs a judgment matrix, adjusts weights using the coefficient of variation, employs grey relational analysis and a fuzzy relationship matrix, and finally outputs the forecast results using a pre-trained model. While this scheme introduces multi-factor weighting and correlation analysis to improve the structure of the forecast, its core still relies on static weight allocation and historical data-driven fuzzy mapping. It fails to model users' proactive decision-making behavior under the influence of multiple factors such as electricity prices, weather, and electricity consumption habits, and does not consider the strategic interaction relationships between different users or microgrids. It cannot reflect the dynamic game process of user behavior within the distribution area, resulting in a lack of real-time adaptability and behavioral interpretability in complex interaction scenarios. Another power grid load forecasting method, with publication number CN112101663B, is based on big data and partially observable Markov decision processes, combining an orthogonal matching tracking algorithm to classify load types and achieve short-term forecasting. While this method incorporates decision-making process modeling and possesses some dynamic verification capabilities, its objective function is set as global optimization scheduling. It fails to differentiate user individual benefit objectives at the transformer substation level and does not establish a game mechanism to allow for the interaction of user strategies. Furthermore, its factor system is fixed and lacks the ability to dynamically generate and evolve factors from multi-source heterogeneous data, making it difficult to support the needs for refined and personalized transformer substation load forecasting.

[0004] The above problems indicate that existing technologies generally have the following shortcomings in distribution area load forecasting: (1) the multi-factor processing method is static, lacking a closed-loop mechanism for dynamic generation, evaluation and updating of factors; (2) user behavior modeling is lacking, and users are not regarded as rational decision-making subjects with objective benefit functions; (3) the strategic interaction between microgrids or user groups is ignored, and the game theory framework is not introduced to characterize the behavioral evolution process; (4) there is a lack of feedback correction mechanism between the forecast results and the actual electricity consumption deviation, making it difficult to continuously optimize model parameters. Therefore, this invention proposes a distribution area load forecasting system based on multi-factor model game, which aims to achieve high-precision, highly adaptable distribution area load forecasting with behavioral interpretability through data fusion, dynamic factor generation, user benefit function configuration and game-driven behavioral evolution mechanism, providing technical support for the refined operation of smart distribution networks. Summary of the Invention

[0005] In view of this, the purpose of this invention is to provide a load forecasting system for transformer substations based on a multi-factor model game.

[0006] To solve the above-mentioned technical problems, the technical solution of the present invention is: a transformer area load forecasting system based on a multi-factor model game, comprising:

[0007] The data fusion module is used to access and process multi-source heterogeneous data from the transformer area measurement system, meteorological information system, power market information system and user-side management system, and to process the multi-source heterogeneous data to obtain electricity consumption related information.

[0008] The factor generation module is configured with a preset factor splitting strategy. The factor splitting strategy is used to generate a corresponding user's electricity consumption association factor group based on electricity consumption association information. The electricity consumption association factor group includes several electricity consumption association factors.

[0009] The user configuration module configures the target revenue function corresponding to the user based on user characteristic information, and configures a parameter optimization strategy. The parameter optimization strategy adjusts the revenue parameters of the target revenue function according to the user's historical information to update the target revenue function.

[0010] The model game module is configured with an electricity consumption game strategy, which is configured in the corresponding microgrid, including:

[0011] Step S1: Obtain the target revenue function and electricity consumption correlation factor for the current user;

[0012] Step S2: Generate the user behavior of the current user based on the target revenue function and electricity consumption correlation factors;

[0013] Step S3: Substitute the user behavior of the microgrid into the preset game model to obtain the user behavior of the next time window;

[0014] Step S4: Calculate the comprehensive power load for the next time window using a preset load calculation algorithm;

[0015] Step S5: Update the electricity consumption correlation factor using the preset factor update strategy and return to step S2;

[0016] The load calculation module calculates the predicted load of the distribution area based on the comprehensive power load of each microgrid.

[0017] Furthermore, the factor splitting strategy includes:

[0018] Step A1: Configure a feature extraction sub-strategy to extract features from electricity consumption-related information to generate several electricity consumption-related features;

[0019] Step A2: Traverse the electricity consumption-related features to match the corresponding anchoring features from the preset anchoring feature library;

[0020] Step A3: Determine the corresponding anchoring factor based on the anchoring features using the anchoring generation sub-strategy;

[0021] Step A4: Configure the factor model construction sub-strategy to construct a factor splitting model based on the anchor factor;

[0022] Step A5: Input the electricity consumption correlation information into the factor splitting model to obtain dynamic factors;

[0023] Step A6: Calculate the split matching value of the dynamic factor and the anchoring factor using a preset split evaluation algorithm. When the split matching value is higher than the first benchmark matching value, use the dynamic factor or anchoring factor as the output electricity consumption correlation factor; when the split matching value is lower than the first benchmark matching value, generate model adjustment parameters to adjust the factor splitting model and return to step A5. Through a closed-loop strategy of feature extraction, anchoring matching, model construction, and split evaluation, adaptive electricity consumption correlation factors are dynamically generated, solving the problem of a fixed factor system. The factor splitting model can be adjusted in real time, ensuring factor effectiveness, providing accurate data support for subsequent load forecasting, and improving the reliability of the forecasting basis.

[0024] Furthermore, the feature extraction sub-strategy is configured with a dynamic feature library, which stores several feature structure groups, and each feature structure group is configured with an index thread set.

[0025] The feature extraction sub-strategy includes:

[0026] Step A11: Preprocess each piece of electricity consumption association data in progress to unify the data format of each type of electricity consumption association data;

[0027] Step A12: Match each electricity consumption-related data according to the index clue set to determine the closest structural feature group for each electricity consumption-related data;

[0028] Step A13: Re-edit the power consumption association data according to the structural feature group so that the power consumption association data has the corresponding structural features;

[0029] Step A14: Identify electricity consumption-related data using a pre-built feature recognition model to obtain corresponding electricity consumption-related features. Relying on a dynamic feature library and index thread set, standardize the data format and match structural feature groups to regulate electricity consumption-related data. Accurately extract features through the feature recognition model, avoiding feature distortion caused by data heterogeneity, and improving the accuracy and adaptability of electricity consumption-related feature extraction.

[0030] Furthermore, the anchoring generation sub-strategy includes:

[0031] Step A31: Configure the anchor matching algorithm to calculate the anchor matching value for each anchor feature;

[0032] Step A32: Filter anchor features with anchor matching values ​​higher than the preset benchmark anchor value to construct an anchor feature group;

[0033] Step A33: Calculate the attenuation influence value of each anchoring feature. The attenuation influence value is the sum of the attenuation correlation sub-values, which reflect the correlation between anchoring features in the anchoring feature group.

[0034] Step A34: Remove the anchoring feature with the highest attenuation impact value from the anchoring feature group and return to step A33 until the preset attenuation mutual exclusion condition is met.

[0035] Step A35: Obtain the corresponding anchoring factors by matching anchoring features with a pre-defined anchoring factor library. Through anchoring match value filtering and attenuation influence value calculation, eliminate redundant anchoring features with strong correlations and construct an efficient anchoring feature group. This ensures the independence and effectiveness of the anchoring factors, reduces interfering factors, and provides high-quality basic factors for the construction of the factor splitting model.

[0036] Furthermore, the factor generation module also includes a splitting construction unit, which is used to construct a factor feature network. The factor feature network includes several node factors, each node factor having node factor information, including node type, node features, and node index. Factor correlation lines are configured between the node factors, reflecting the correlation between the node factors. The factor model construction sub-strategy includes:

[0037] Step A41: Determine the position of each anchoring factor in the factor feature network and generate the transmission break value;

[0038] Step A42: Calculate the transmission cost value corresponding to the factor correlation line of adjacent node factors. If the transmission break value is higher than the transmission cost value, obtain the corresponding node factor and use the difference between the transmission break value and the transmission cost value as the new transmission break value, until all node factors with transmission break values ​​can no longer be transmitted.

[0039] Step A43: Obtain the identified node factors to generate a factor splitting model. Based on the factor feature network, node factors are screened by passing breakpoint values ​​and passing cost values, and a targeted factor splitting model is constructed. Fully utilize the correlation of node factors to make the model fit the characteristics of electricity consumption-related information, improving the targeting and rationality of factor splitting.

[0040] Furthermore, the splitting evaluation algorithm is used to weight the correlation matching item, content matching item, and topology influence item to obtain the splitting matching value. The correlation matching item reflects the degree of correlation between the determined node factors, the content matching item reflects the degree of matching between electricity consumption correlation information and the corresponding node factor, and the topology influence item reflects the degree of influence of the node factor on the topological relationship in the factor splitting model. By comprehensively evaluating the three indicators of weighted correlation matching, content matching, and topology influence, the factor splitting effect is comprehensively assessed. This avoids the one-sidedness of single-dimensional evaluation, ensures that the output electricity consumption correlation factors are highly adapted to the electricity consumption correlation information, and guarantees the reliability of subsequent predictions.

[0041] Furthermore, the user configuration module is configured with a set of objective functions, which includes several objective revenue functions. Each objective revenue function is configured with user matching clues. The user configuration module matches user matching clues based on user feature information to determine the objective revenue function with the highest matching degree. By using the set of objective functions and user matching clues, the objective revenue function is accurately matched based on user feature information. This enables differentiated configuration of revenue functions, aligning with the electricity consumption characteristics of different users, providing a personalized basis for user behavior modeling, and improving the adaptability of behavior prediction.

[0042] Furthermore, the parameter optimization strategy includes:

[0043] Step B1: Input the electricity consumption correlation factors from the user's historical information into the target revenue function of the current user to generate an array of expected electricity consumption revenues;

[0044] Step B2: Retrieve the corresponding user behavior from the expected electricity revenue array, and generate predicted electricity consumption parameter information based on the user behavior;

[0045] Step B3: Compare the predicted electricity consumption parameters with the actual electricity consumption parameters from historical data to obtain electricity consumption deviation information;

[0046] Step B4: Based on the electricity consumption deviation information, retrieve the deviation correction table corresponding to the target revenue function to obtain adjustment variables and update the revenue parameters of the target revenue function. Based on historical user information, obtain the deviation by comparing expected revenue with actual data, and adjust the revenue parameters. Form a dynamic optimization closed loop for the target revenue function, correct function deviations, and improve the accuracy of the revenue function in depicting users' actual electricity consumption behavior.

[0047] Furthermore, the load calculation algorithm obtains the load value through weighted predicted load and corrected load. The predicted load value is obtained by summing the comprehensive electricity load of all microgrids, and the corrected load value is obtained by summing the cumulative deviation load and the loss-corrected load. The cumulative deviation load reflects the superposition deviation of the target revenue function, and the loss-corrected load reflects the load loss corresponding to each microgrid. By combining the predicted load value and the corrected load value, both the cumulative deviation and load loss factors are taken into account. The weighted calculation optimizes the comprehensive electricity load result, compensating for the shortcomings of simply summing and ignoring errors, thus improving the accuracy and completeness of the load calculation.

[0048] Furthermore, the model game module includes a model building unit, which constructs the game model. This model is trained using historical information as samples to evolve the correlation between user behaviors, thereby determining user behavior in the next time window. Training the game model with historical information as samples evolves the correlation between user behaviors. This accurately characterizes the interaction patterns of user behavior within the microgrid, providing a scientific basis for predicting behavior in the next time window and improving the rationality and dynamic adaptability of behavior prediction.

[0049] The main technical effects of this invention are reflected in the following aspects: It integrates multi-source heterogeneous data, and through dynamic factor generation, user benefit function configuration, and game-driven behavioral evolution, it characterizes the dynamic game-like characteristics of user behavior, overcoming the shortcomings of traditional methods that are static and lack user decision-making modeling. This enables refined forecasting of distribution area loads, improves forecast adaptability and behavioral interpretability, and provides reliable support for distribution network operation, planning, and demand-side response. Attached Figure Description

[0050] Figure 1 : System architecture schematic diagram of the present invention;

[0051] Figure 2 : Flowchart of the electricity game strategy of this invention;

[0052] Figure 3 : Flowchart of the factor splitting strategy of this invention;

[0053] Figure 4 : Flowchart of the anchoring generation sub-strategy of this invention;

[0054] Figure 5 A flowchart illustrating the construction of sub-strategies using a clearly defined factor model;

[0055] Figure 6 : Flowchart of the parameter optimization strategy of this invention. Detailed Implementation

[0056] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings, so that the technical solution of the present invention can be more easily understood and mastered.

[0057] Reference Figure 1 As shown, a transformer area load forecasting system based on a multi-factor game model includes:

[0058] The data fusion module is used to access and process multi-source heterogeneous data from the transformer substation measurement system, meteorological information system, power market information system, and user-side management system. It processes this multi-source heterogeneous data to obtain electricity consumption-related information. Through a series of standardized processes, it obtains comprehensive electricity consumption-related information that reflects the intrinsic relationship between user electricity consumption behavior and external influencing factors. Multi-source heterogeneous data refers to data from different systems with varying formats and structures, but all related to transformer substation electricity consumption. The processing flow includes four consecutive stages: data access, data preprocessing, data alignment, and data fusion, ensuring accurate and usable results. The data access stage employs a multi-protocol adaptive access mechanism, supporting stable connections with various data source systems via Ethernet, RS485, LoRa wireless, and 5G communication interfaces. It uses standard communication protocols such as Modbus, IEC61850, and MQTT to parse data transmission formats and achieve real-time data acquisition. During data collection, a data integrity verification mechanism is employed, requiring the number of data fields to match the preset limit and that key numerical fields are not empty. If verification fails, three retransmission requests are triggered. If incomplete data is still not obtained, a missing data log is recorded, and the missing time period is marked. The data preprocessing stage includes three sub-steps: data cleaning, data standardization, and data completion. The data alignment stage uses a sliding window-based timestamp synchronization algorithm. The first step determines a uniform time granularity of 15 minutes, with each time window starting at 00:00, 15:00, 30:00, and 45:00. The second step extracts the original timestamps from each data source and converts them to Unix timestamp format. The third step calculates the time difference between the original timestamp of each data source and the window's starting timestamp. If the absolute value of the time difference is less than or equal to 7.5 minutes, the data is assigned to the current time window; otherwise, it is discarded or assigned to the nearest adjacent window. The fourth step deduplicates data of the same type within each time window, retaining data whose timestamp is closest to the window's center time. The data fusion stage employs a hybrid fusion strategy combining static weighted fusion and dynamic intelligent fusion.

[0059] Reference Figure 3 As shown, the factor generation module generates electricity consumption-related factor groups for each user based on the electricity consumption correlation information output by the data fusion module, using a preset factor splitting strategy. Each electricity consumption-related factor group consists of a set of independent factors that accurately reflect the quantitative relationship between user electricity consumption behavior and various influencing factors. Each factor has a clear physical meaning and quantitative indicator, and can be directly used for subsequent user behavior modeling and game theory analysis. The specific execution steps are as follows:

[0060] Step A1 configures a feature extraction sub-strategy to extract features from electricity consumption-related information to generate several electricity consumption-related features. This sub-strategy relies on a dynamic feature library to achieve accurate feature extraction for various types of electricity consumption-related information. The dynamic feature library is a database that can be updated in real-time according to changes in the type of electricity consumption-related information and the needs of actual application scenarios. The stored feature structure groups consist of several electricity consumption-related data templates with similar data structures and feature dimensions. Each feature structure group corresponds to an index clue set, which is a set of key feature identifiers used for quickly matching electricity consumption-related data, including data type identifiers, numerical ranges, thresholds, time dimensions, and attributes. The specific execution steps of the feature extraction sub-strategy are as follows: Step A11 preprocesses each piece of electricity consumption-related data in the electricity consumption-related information, unifying the format of different data types to JSON format, standardizing the units of numerical data (power to kilowatts, voltage to kilovolts, current to amperes), and converting textual data into numerical codes, such as converting weather type identifiers (sunny, cloudy, rainy) into different numerical values. Step A12 calculates the matching degree between each electricity consumption-related data point and each feature structure group in the dynamic feature library based on the key feature identifiers in the index clue set. The matching degree calculation adopts a multi-dimensional weighted summation method, prioritizing the matching of feature structure groups with consistent data type identifiers and numerical range thresholds, and determining the closest structural feature group for each electricity consumption-related data point. Step A13 re-edits the electricity consumption-related data according to the template format of the determined structural feature group, supplementing missing feature fields and adjusting the data arrangement order, so that the edited electricity consumption-related data fully possesses the structural features corresponding to the structural feature group. Step A14 identifies the electricity consumption-related data through a pre-built feature recognition model to obtain the corresponding electricity consumption-related features. This feature recognition model is a deep learning model based on convolutional neural networks. The input training samples are the pre-processed and structurally edited dataset of electricity consumption-related data from the past 3 years. Each sample contains the data itself and the corresponding manually labeled feature labels. The output form is the feature vector corresponding to the electricity consumption-related data, and the dimension of the feature vector is consistent with the feature dimension of the feature structure group in the dynamic feature library. The model training process uses the gradient descent optimization algorithm and the cross-entropy loss function. The dataset is divided into training set and test set in an 8:2 ratio. The training set is used for iterative updating of model parameters, and the test set is used to verify the recognition accuracy of the model. Training is stopped when the recognition accuracy of the test set is higher than 96% for 10 consecutive training cycles, and the optimal model parameters are saved for actual feature extraction.

[0061] Step A2 iterates through all generated electricity consumption-related features and matches the corresponding anchor features from the preset anchor feature library. The anchor feature library is a pre-constructed standardized feature set containing the features corresponding to all known electricity consumption influencing factors. Each anchor feature has been verified by a large amount of historical data to have stable correlation and recognizability. The matching process is achieved by calculating the similarity between the electricity consumption-related features and each feature in the anchor feature library. The similarity calculation adopts the cosine similarity algorithm to ensure that the anchor feature that best matches the electricity consumption-related features is found.

[0062] Step A3 determines the corresponding anchoring factor based on the anchoring features using the anchoring generation sub-strategy. The specific execution steps of the anchoring generation sub-strategy are as follows: Step A31 configures an anchoring matching algorithm to calculate the anchoring matching value for each anchoring feature. The formula for calculating the anchoring matching value is... Where M i ω is the anchor matching value of the i-th anchor feature. k These are the weight coefficients of the k-th matching dimension, which are determined using the analytic hierarchy process (AHP) to satisfy... S ik It is the similarity between the i-th anchor feature and the electricity consumption-related feature in the k-th dimension. The similarity is calculated using the cosine similarity formula. It is the feature vector of the i-th anchored feature. This is the feature vector of the electricity consumption correlation feature in the k-th dimension, where n is the total number of matching dimensions, including numerical similarity, structural similarity, and temporal correlation similarity, totaling 5 dimensions. Step A32 sets the preset benchmark anchoring value to 0.6, determined through historical data statistics. Anchoring features with matching values ​​higher than this benchmark are selected to construct an anchoring feature group, ensuring that the features in the anchoring feature group have a strong correlation with the electricity consumption correlation feature. Step A33 calculates the attenuation influence value of each anchoring feature. The attenuation influence value is the sum of the attenuation correlation sub-values. The attenuation correlation sub-value is a quantitative indicator reflecting the strength of the correlation between two anchoring features, calculated using the following formula:

[0063]

[0064] C ij It is the decay correlation sub-value between the i-th anchor feature and the j-th anchor feature, cov(F) i ,F j ) is the covariance of the historical data sequences corresponding to the two anchor features. It is the standard deviation of the historical data sequence of the i-th anchor feature. The standard deviation of the historical data sequence of the j-th anchor feature is given by the formula for calculating the attenuation effect value:

[0065]

[0066] , m is the total number of anchor features in the anchor feature group, D i The larger the value, the stronger the correlation between the anchoring feature and other anchoring features, and the higher the redundancy. Step A34 pre-sets the attenuation mutual exclusion condition, which means that the attenuation influence value of all anchoring features in the anchoring feature group is less than 0.3. The anchoring feature with the highest attenuation influence value is removed from the anchoring feature group, and the process returns to step A33 to recalculate the attenuation influence values ​​of the remaining anchoring features until the attenuation mutual exclusion condition is met, ensuring that the features in the anchoring feature group have good independence. Step A35 obtains the corresponding anchoring factor by accurately matching the anchoring feature with a pre-set anchoring factor library. The anchoring factor library is a pre-constructed mapping database of anchoring features and anchoring factors, where each anchoring feature corresponds to a unique anchoring factor with a clear physical meaning.

[0067] Step A4 configures a factor model construction sub-strategy to construct a factor splitting model based on anchored factors. The factor generation module also includes a splitting construction unit, which is used to construct a factor feature network. The factor feature network is an undirected weighted graph containing several node factors. Each node factor has node factor information. Node types include environmental factors, electricity consumption factors, market factors, etc. Node features are the quantitative attribute values ​​of node factors. Node indexes are unique identifiers used to locate the position of a node in the network. Factor correlation lines are configured between node factors. The weight values ​​of the factor correlation lines reflect the correlation coefficients between node factors. The correlation coefficients are calculated using Pearson correlation analysis, with values ​​ranging from -1 to 1. The larger the absolute value, the stronger the correlation. The specific execution steps of the factor model construction sub-strategy are as follows: Step A41: Locate the specific position of each anchored factor in the factor feature network using the node index, and then generate the transmission breakage value for each anchored factor. The formula for calculating the transmission breakage value is...

[0068]

[0069] T i M is the transmission break value of the i-th anchoring factor. i is the anchoring matching value corresponding to the i-th anchoring factor, p is the total number of anchoring factors, and T0 is the preset basic transmission break value, which is 1.0. Step A42 calculates the transmission cost value corresponding to the factor correlation line of adjacent node factors. The formula for calculating the transmission cost value is C. ij =1-|r ij |, where C ij It is the transitive cost of the factor correlation line between the i-th node factor and the j-th node factor, r ijThis is the correlation coefficient between two node factors. If the transmission break value is higher than the transmission cost value, the adjacent node factor is included in the candidate set, and the transmission break value is updated to the difference between the current value and the transmission cost value. Then, the above process is repeated with the adjacent node factor as the new starting point until the transmission break value of all candidate node factors is no higher than the transmission cost value of their adjacent node factors. Step A43 collects all node factors included in the candidate set and the initial anchoring factor, and constructs a factor splitting model according to the correlation between factors. This model can accurately split dynamic factors based on the input electricity consumption correlation information.

[0070] Step A5 inputs the preprocessed and structured electricity consumption information into the constructed factor splitting model. Through the factor mapping relationship and quantification calculation within the model, dynamic factors are output. Dynamic factors are factors that are generated in real time based on the current electricity consumption information and reflect the specificity of the current electricity consumption scenario. Together with the anchoring factors, they constitute a candidate factor set.

[0071] Step A6 calculates the split matching values ​​of dynamic factors and anchoring factors using a preset split evaluation algorithm. The split evaluation algorithm obtains the split matching value by weighting the content matching item, topological influence item, and related matching item. The calculation formula is S = α × A + β × B + γ × C, where S is the split matching value, α, β, and γ are weight coefficients satisfying α + β + γ = 1. The optimal values ​​are determined to be 0.3, 0.5, and 0.2 respectively using a grid search method. The related matching item A reflects the degree of association between the determined node factors, and the calculation formula is...

[0072]

[0073] k is the total number of node factors, r ij It is the correlation coefficient between the i-th and j-th node factors. Content matching term B reflects the degree of matching between electricity consumption correlation information and the corresponding node factors, and is calculated using the following formula:

[0074]

[0075] n is the number of samples of electricity-related data, x i It is the actual value of the i-th sample. These are the predicted values ​​obtained through node factor prediction. The topological influence term C reflects the degree of influence of node factors corresponding to the topological relationships in the factorization model, and is calculated using the following formula:

[0076]

[0077] d i d is the degree of the i-th node factor in the model topology, i.e., the number of connections to other node factors. maxThis represents the maximum value of the node factors in the model. The first baseline matching value is a threshold of 0.75 determined through historical data statistics. When the split matching value is higher than the first baseline matching value, it indicates that the factor splitting effect is good. This dynamic factor or anchoring factor is then included in the electricity consumption correlation factor group as the output. When the split matching value is lower than the first baseline matching value, model adjustment parameters are generated. The adjustment parameters include the correction value of the weight coefficient of the node factors and the correction amount of the base value of the transmission break value. By adjusting the topology and parameters of the factor splitting model, the process returns to step A5 to extract dynamic factors until the split matching value meets the requirements.

[0078] The user configuration module, based on the electricity consumption-related factor set output by the factor generation module, combines user basic information with grid operation constraints to construct a personalized user configuration system. It generates a user configuration parameter set containing electricity consumption constraint parameters, response preference parameters, and equipment adaptation parameters. This parameter set is the core link between user-side demand and grid-side control, directly supporting subsequent user behavior modeling, game theory analysis, and precise control command generation. User basic information includes user identity type, electricity address, access phase, list of electrical equipment, operating parameters, and special electricity consumption needs. Identity types include residential, industrial, commercial, agricultural, and special industry users. Access phases are automatically identified and labeled as phase A, phase B, or phase C by the transformer area measurement system. Special electricity needs, such as uninterrupted power supply requirements for medical users and continuous production power requirements for industrial users, require user self-declaration and system verification.

[0079] The core execution logic of the user configuration module adopts a closed-loop process of user profile construction, configuration rule matching, parameter dynamic generation, and configuration effect verification to ensure that the generated configuration parameters are both adaptable to user needs and compliant with power grid operation. The specific execution steps are as follows.

[0080] Step B1 generates user electricity consumption profiles through a sub-strategy based on user profiling. These profiles are quantitative descriptions of user electricity consumption behavior characteristics, needs, preferences, and constraints. The construction process relies on a user electricity consumption feature clustering model, which is an unsupervised learning model based on an improved K-means algorithm. The input training samples are time-series data of electricity consumption-related factors from the past two years and corresponding encoded user basic information. Each sample contains features such as electricity power fluctuation factors, electricity price response factors, and environmental sensitivity factors for 90 consecutive days, as well as encoded information such as user identity type and device power level. The output format is user cluster labels and corresponding feature description vectors. The dimension of the feature description vectors is consistent with the number of electricity consumption-related factors. The model training process first normalizes the input samples and uses the K-means++ algorithm to initialize cluster centers to avoid local optima. The number of clusters is determined using the elbow rule, and the silhouette coefficient is used as the evaluation index for clustering performance. The formula for calculating the silhouette coefficient is:

[0081]

[0082] , of which SC i It is the silhouette coefficient of the i-th sample, a i b is the average distance between this sample and other samples in the same cluster. i This is the average distance between the sample and all samples in the nearest heterogeneous cluster. Training stops when the average silhouette coefficient is higher than 0.7 and the change in cluster centers is less than 0.001 for five consecutive iterations, and the optimal clustering model is saved. Based on the trained model, the current user data is clustered, and combined with the user's basic information, a complete user electricity consumption profile is generated, including electricity consumption pattern labels, demand priority labels, and constraint level labels. Among them, electricity consumption pattern labels are such as residential peak-valley type and industrial and commercial stable type, demand priority labels are divided into 1 to 5 levels according to importance, and constraint level labels are divided into three categories: strong constraint, medium constraint, and weak constraint according to the power grid regulation requirements.

[0083] Step B2 matches corresponding basic configuration rules from a pre-set configuration rule library based on the user's electricity consumption profile. The configuration rule library is a standardized set of rules integrating power grid safety operation regulations, electricity market rules, and user rights protection requirements. It is constructed using a combination of decision trees and fuzzy logic. The rule library is divided into three categories according to configuration type: basic guarantee, market response, and energy-saving optimization. Basic guarantee rules are formulated for the user's core electricity needs. For example, the power supply reliability guarantee rule for medical users specifies that the voltage fluctuation range of their access circuit is ±5% of the rated voltage, and the power interruption time does not exceed 50 milliseconds. Market response rules are linked to the electricity market mechanism. For example, the user's electricity consumption period adjustment rule under time-of-use pricing specifies the power adjustment threshold for different price ranges. Energy-saving optimization rules are combined with the user's equipment energy efficiency parameters. For example, the temperature adjustment and energy consumption correlation rule for air conditioning equipment. The matching process is achieved through rule matching degree calculation. The matching degree calculation formula is as follows: Where R i λ is the matching degree of the i-th rule. t The weight coefficients for the t-th class of labels are determined using the analytic hierarchy process (AHP) to be 0.5, 0.3, and 0.2 for the basic support, market response, and energy-saving optimization categories, respectively. it It is the fit between the i-th rule and the t-th tag in the user profile. The fit is calculated by weighting semantic similarity and numerical similarity, and the top 3 rules with the highest matching degree are selected to form the initial rule set.

[0084] Step B3 generates initial configuration parameters based on the initial rule set and electricity consumption-related factor group through a parameter generation sub-strategy. The parameter generation adopts a weighted factor mapping method, and the core calculation formula is as follows: Where Param is the generated initial configuration parameter vector, ω s It is the weight coefficient of the s-th rule, calculated according to the matching degree normalization, which satisfies Rule s It is the parameter constraint vector of the s-th rule, Factor s This is the power consumption correlation factor vector associated with the s-th rule, where ° indicates element-wise multiplication of the vectors. Specific parameter generation includes three dimensions: power consumption constraint parameters such as maximum allowable power, minimum operating power, and voltage fluctuation threshold, where the maximum allowable power is calculated using the formula P. max =ω1P 1max F 1load +ω2P 2max F 2price +ω3P 3max F 3env P 1max It is the power limit in the basic protection rules, F 1load It is the load factor, P 2max It is the power cap in the market response rules, F 2price It is the electricity price factor, P 3max It is the power limit in the energy-saving optimization rules, F 3env These are environmental factors; response preference parameters, such as the incentive threshold for user participation in demand response and response delay time, etc., with the incentive threshold being positively correlated with the electricity price response factor; and equipment adaptation parameters, such as the communication protocol standard for equipment access and control signal interface parameters, etc., which are determined based on the user equipment list and power grid access specifications.

[0085] Step B4 verifies the compliance and validity of the initial configuration parameters using a configuration verification algorithm. The verification indicators include three core metrics: grid constraint compliance, user requirement satisfaction, and economic optimization. The formula for calculating grid constraint compliance is as follows:

[0086]

[0087] Where Com is the compliance level, m is the number of power grid constraint parameters, and Param j It is the j-th configuration parameter value, Limit j This corresponds to the power grid constraint threshold, and the compliance rate must be no less than 0.95; the formula for calculating user demand satisfaction is... Where Sat is the satisfaction level, and η is the degree of satisfaction. k It is the weight of the k-th level demand, Req k This refers to the parameter requirements corresponding to the k-th level of demand. Ind is an indicator function that takes the value 1 if the condition is met and 0 otherwise, and the satisfaction level must be no less than 0.9. The formula for calculating the economic optimization degree is as follows:

[0088]

[0089] Where Eco is the degree of optimization, and C is the degree of optimization. base This is the estimated electricity cost without configuration, C configThis refers to the estimated electricity cost after configuration, with an optimization degree of no less than 0.05. When all three indicators meet the requirements, the initial configuration parameters are output as the final user configuration parameter set. If any indicator fails to meet the requirements, a parameter adjustment strategy is triggered. The rule weights and factor mapping coefficients are corrected using a particle swarm optimization algorithm. The adjusted parameters are then returned to step B4 for verification until all indicators meet the requirements. The user configuration module also has a dynamic parameter update function. When a change in the user's electricity consumption-related factor group exceeds 20% or when power grid operation constraints are adjusted, the above closed-loop process is automatically triggered to update the configuration parameters, ensuring the real-time adaptability of the configuration parameters. Simultaneously, the module allows users to query configuration parameter content via mobile terminals and modify response preference parameters within the compliant scope. Modification requests are included in the parameter update process after system verification, achieving two-way interaction between user needs and system configuration.

[0090] Reference Figure 2 As shown, the game theory module, based on the electricity consumption-related factor set from the factor generation module and the user configuration parameter set from the user configuration module, constructs a two-agent non-cooperative game model between the user side and the power grid side. By solving the game equilibrium strategy, it generates an optimal control reference scheme that satisfies both user electricity consumption preferences and power grid operation objectives. This scheme is the core basis for achieving coordinated user-driven electricity consumption and precise power grid control, and can be directly transformed into specific electricity dispatch instructions and incentive mechanisms. The user-side game agents aim for "minimizing electricity costs + optimizing electricity experience," while the power grid-side game agents aim for "minimizing network losses + balancing loads + maximizing power supply reliability." The two parties achieve dynamic equilibrium through strategy interaction.

[0091] The core execution logic of the model game module adopts a closed-loop process of defining the strategy space of the game subjects, constructing the payoff function, designing the equilibrium strategy, solving the game result verification, and ensuring that the generated equilibrium strategy is both feasible and optimal. The specific execution steps are as follows.

[0092] Step C1 clarifies the game participants and their constraints. The participants are divided into user-side participants and grid-side participants. User-side participants are grouped according to their electricity consumption pattern labels in the user configuration parameter set. Users with the same electricity consumption pattern label are grouped into a single user cluster, such as residential peak-valley clusters or industrial / commercial stable clusters. Each cluster integrates the electricity demand and strategy preferences of its internal users through a cluster agent node, achieving clustered game theory to reduce computational complexity. The constraints for user-side participants are determined based on the user configuration parameter set, including the maximum allowable power constraint P. u,max ≤P config,max Voltage fluctuation threshold constraint V u,min ≤V u ≤V u,max Response delay constraint T u,delay ≤T config,delay, where P u,max It is the actual maximum power consumption on the user side, P config,max It is the maximum allowable power in the user-configured parameter set, V u It is the actual voltage at the user access point, V u,min and V u,max It refers to the upper and lower limits of voltage fluctuation in the configuration parameters, T. u,delay This is the actual user response latency, T config,delay This refers to the response delay threshold in the configuration parameters. The main constraints on the grid side are determined based on grid operation procedures, including total power constraints for distribution areas. Three-phase load imbalance constraint ΔI≤10%, network loss constraint L≤L max Where U is the total number of users, P grid,max It is the maximum allowable power supply of the transformer area, ΔI is the three-phase current imbalance, and L is the actual network loss of the transformer area. max This is the maximum permissible network loss in the area.

[0093] Step C2 constructs the strategy space for both parties. The strategy space is the set of all feasible strategies for each party, described using a combination of discretization and continuous representation. The user-side party's strategy space includes three categories: electricity consumption time adjustment strategy, power regulation strategy, and demand response participation strategy. The electricity consumption time adjustment strategy divides a 24-hour day into 96 time slices with 15-minute intervals, represented by a binary variable x. u,t Indicates whether user u uses electricity during time slice t, x u,t =1 indicates electricity consumption, x u,t =0 indicates no power is needed; the power regulation strategy is based on the power constraints in the user-configured parameters, within [P u,min ,P u,max Continuously adjustable within an interval, using a continuous variable P. u,t This represents the actual power consumption of user u in time slice t; the demand response participation strategy uses the participation variable r. u,t It means that r u,t The value ∈ [0,1] indicates a stronger willingness to participate in grid demand response and a greater potential for adjustment. The strategy space of the grid-side entities includes three categories: electricity price incentive strategies, power regulation command strategies, and power supply reliability assurance strategies. Electricity price incentive strategies, based on the benchmark electricity price, use a floating coefficient δ... t Adjustment, δ t ∈[0.8,1.5],δ t <1 represents the incentive electricity price, δ t >1 represents the constrained electricity price; the power regulation command strategy uses C. u,t This indicates that the power regulation target for user u in time slice t must meet the user-side power constraints; the power supply reliability assurance strategy uses the reserve capacity allocation variable S. t express, Ensure stable power supply in case of emergencies.

[0094] Step C3 designs the payoff functions for both parties. The payoff function is a core indicator for quantifying the benefits of implementing the strategy, and it is constructed using a multi-objective weighted summation method. The user-side party's payoff function U... u It includes three parts: electricity cost savings, electricity experience guarantee benefits, and demand response incentive benefits, calculated using the formula U. u =α1U u,cost +α2U u,exp +α3U u,dr α1, α2, and α3 are weighting coefficients, determined through a user response preference survey combined with the analytic hierarchy process, with default values ​​of 0.5, 0.3, and 0.2 respectively. Electricity cost savings benefits. C u,base This is the baseline electricity cost when the user does not participate in the game, p t It is the benchmark electricity price for time slice t; electricity consumption experience guarantee benefits. T u It is the core set of electricity consumption time slices for user u, P u,prefer It represents the user's preferred power consumption; demand response incentive revenue. k t This is the demand response incentive coefficient for time slice t, which is positively correlated with the grid load stress level. The main revenue function U_{\text{grid}} on the grid side includes three parts: revenue from reduced grid losses, revenue from load balancing, and revenue from power supply security. The calculation formula is U_{\text{grid}}. grid =β1U grid,loss +β2U grid,balance +β3U grid,safe β1, β2, and β3 are weighting coefficients, determined according to the grid operation priority, with default values ​​of 0.4, 0.3, and 0.3 respectively. The benefit U from reduced grid losses. grid,loss =L base -L,L base It is the baseline network loss before the game; load balancing revenue.

[0095]

[0096] P grid,target This refers to the target load value for the transformer area; and the benefits of power supply security.

[0097]

[0098] Ind(.) is an indicator function that takes the value 1 if the condition is met, and 0 otherwise.

[0099] Step C4 uses a game equilibrium solution algorithm to find the Nash equilibrium strategy. A Nash equilibrium strategy is a stable strategy combination where neither party can improve their own payoff by unilaterally adjusting their strategy, given that the other's strategy remains unchanged. The solution process relies on an improved particle swarm optimization algorithm, which introduces a game learning factor to enhance convergence speed and the optimality of the solution. The algorithm's input training samples consist of a dataset of user electricity consumption strategies, grid control strategies, and corresponding payoff function values ​​from the past year. Each sample contains x values ​​from the user side. u,t P u,t r u,t δ on the grid side t C u,t S t The output format is the optimal strategy combination and its corresponding payoff value. The training process first initializes the particle swarm, with each particle representing a strategy combination. The particle dimension is the total number of policy variables for both sides. Then, the fitness value of each particle is calculated, which is the weighted sum of the payoff functions for both sides, F = γU. u +(1-γ)U grid γ∈[0.4,0.6] is the fairness coefficient; then the particle swarm optimization formula v is updated iteratively. i,d (t+1)=ωv i,d (t)+c1r1(p i,d -x i,d (t))+c2r2(p g,d -x i,d (t)) and x i,d (t+1)=x i,d (t)+v i,d (t+1) Update particle velocity and position, where v i,d ω is the velocity of the i-th particle in the d-th dimension, c1 and c2 are learning factors, r1 and r2 are random numbers in the interval [0,1], and p is the velocity of the d-th particle. i,d It is the optimal position of an individual particle, p g,d It is the optimal position of the group; finally, the convergence condition is set as the change of the optimal fitness value of the group in 10 consecutive iterations being less than 0.001. When the condition is met, training stops and the corresponding optimal policy combination is output as the Nash equilibrium policy.

[0100] Step C5 verifies the feasibility and stability of the Nash equilibrium strategy using a game result verification algorithm. The verification indicators include three aspects: strategy constraint satisfaction rate, payoff stability, and user acceptance. The formula for calculating the strategy constraint satisfaction rate is as follows: M is the total number of constraints, Strategy m It is the value of the strategy under the m-th constraint, Constraint mThis is the feasible region of the m-th constraint, and the constraint satisfaction rate must be no less than 0.98; the return stability is calculated using the return volatility coefficient CV = σ. U / μ U , σ U It is the standard deviation of the revenues of both parties over the next 24 hours, μ U The average return must be no higher than 0.1. User acceptance is determined through sampling surveys combined with fuzzy comprehensive evaluation. 20% of users in the corresponding user cluster are invited to rate the strategy, and the average rating must be no lower than 4 points (out of 5). When all three indicators meet the requirements, the Nash equilibrium strategy is output as the final game result. If any indicator is not met, the weight coefficients of the payoff function and the strategy space boundary are adjusted, and the process returns to step C4 to solve for equilibrium until all indicators meet the requirements.

[0101] The core function of the load calculation module is to construct a multi-dimensional load calculation system based on the power consumption correlation factor set output by the factor generation module and the user configuration parameter set output by the user configuration module, combined with real-time power grid operation data. This system accurately generates user-level and transformer area-level load forecast results and load characteristic analysis reports. These results provide core data support for power grid dispatch planning, power supply capacity assessment, and fault prediction, directly determining the economic efficiency and reliability of power grid operation. User-level load calculation focuses on the power consumption variation patterns of a single user or a cluster of similar users. Transformer area-level load calculation aggregates user-level loads and superimposes transformer area losses to form a comprehensive load profile for the entire transformer area. Real-time power grid operation data, including transformer area bus voltage, current, power factor, and feeder load, is accessed and updated in real-time through the data fusion module.

[0102] The core execution logic of the load calculation module adopts a closed-loop process of data preprocessing, load calculation model construction, load stratification calculation, load characteristic analysis results verification and optimization to ensure the accuracy and practicality of the load calculation results. The specific execution steps are as follows.

[0103] Step D1 involves targeted preprocessing of the input data. The core objective of this preprocessing is to improve data quality and provide a reliable foundation for subsequent calculations. This preprocessing includes three sub-steps: data filtering, time series alignment, and outlier correction. Data filtering is based on power consumption constraint parameters in the user configuration parameter set. It retains characteristic factors strongly correlated with the load in the power consumption-related factor group, such as power fluctuation factors, environmental sensitivity factors, and electricity price response factors, while removing irrelevant features such as equipment communication protocol encoding to reduce computational redundancy. Time series alignment uses a uniform time granularity of 15 minutes, synchronizing the timestamps of user configuration parameter update records and real-time grid data in the power consumption-related factor group to the Unix timestamp format to ensure one-to-one correspondence of data within the same time slice. Outlier correction employs an improved 3σ algorithm. Considering the time series characteristics of load-related data, it uses a sliding time window instead of a fixed time period to calculate statistics. The window size is set to 24 hours, and the calculation formula is |xt -μ t,w |>3σ t,w , where x t It is the data to be detected at time slice t, μ t,w σ is the mean of the data within the sliding window containing time slice t. t,w This corresponds to the standard deviation of the data within the window. Data identified as outliers are replaced with the weighted mean of data from two adjacent valid time slices. The weights are set to 0.6 and 0.4 based on the time distance, i.e., x′ t =0.6x t-1 +0.4x t+1 , x′ t These are the corrected outliers.

[0104] Step D2 constructs a hierarchical load calculation model, which consists of a user-level basic load calculation sub-model, a user-level dynamic load calculation sub-model, and a transformer area-level total load aggregation sub-model, realizing hierarchical progressive calculation from users to transformer areas. The user-level basic load calculation sub-model adopts a statistical regression algorithm based on historical data, and the core calculation formula is P. u,base,t =k u ×μ u,T ×(1+α u ×F u,env,t ), where P u,base,t It is the base load of user u in time slice t, k u This is the electricity consumption pattern correction factor for user u. The value is 1.2 for residential peak-valley users and 0.95 for industrial and commercial stable users. It is determined by the electricity consumption pattern label in the user configuration parameter set. u,T α is the average load of user u during the same period over the past 3 months. u This is the environmental sensitivity coefficient for user u, ranging from 0.01 to 0.05, calculated from the environmental sensitivity factor, F. u,env,t It is the normalized value of the environmental factors for user u in time slice t.

[0105] The user-level dynamic load calculation sub-model employs a Long Short-Term Memory (LSTM) network model, specifically designed to capture the long-short-term dependencies in time-series data. The input training samples are preprocessed data from the past year, with each sample containing a vector of electricity consumption-related factors for 48 consecutive time slices and the corresponding actual dynamic load value for each time slice. The output is a predicted dynamic load value for the next 24 time slices. During model training, the dataset is divided into training and validation sets in an 8:2 ratio. The input data is normalized to the [0,1] interval using Min-Max. The model structure consists of an input layer, hidden layers, and an output layer. The number of neurons in the input layer matches the dimension of the electricity consumption-related factors. There are two hidden layers with 64 neurons each, using ReLU as the activation function. The output layer has 24 neurons, and the loss function is mean squared error.

[0106]

[0107] y i This represents the actual dynamic load value. Let N be the predicted value, N be the number of samples, and Adam be the optimizer with an initial learning rate of 0.001. Training stops when the validation set loss decreases by less than 0.0001 for five consecutive epochs, and the optimal model parameters are saved. The total user-level load is the sum of the base load and the dynamic load, i.e., P. u,t =P u,base,t +P u,dynamic,t P u,dynamic,t This represents the dynamic load value output by the Long Short-Term Memory (LSTM) network model.

[0108] The sub-model for total load aggregation at the transformer area level is obtained by aggregating the total load of all users and superimposing transformer area losses. The core calculation formula is as follows: Where P grid,t It is the total load of the transformer area in time slice t, γ t It is the transformer area loss coefficient for time slice t, which is positively correlated with the total load of the transformer area. Calculate, S grid η is the rated capacity of the transformer in the distribution area. u,phase It is the access phase correction coefficient for user u. The values ​​for phase A, phase B, and phase C are 1.02, 1.0, and 0.98, respectively, and are determined by the access phase information in the user configuration parameter set. U is the total number of users in the transformer area.

[0109] Step D3 performs hierarchical load calculation. First, based on the user-level basic load calculation sub-model, the basic load of each user is calculated by combining the user's historical load data and current environmental factors. Then, the preprocessed electricity consumption correlation factor group is input into the trained long short-term memory network model to obtain the dynamic load prediction value of each user. The basic load is then superimposed to obtain the user-level total load. Finally, through the transformer area-level total load aggregation sub-model, combined with the access phase correction coefficient and real-time loss coefficient, the transformer area-level total load is aggregated to obtain the transformer area-level total load. At the same time, user-level load curves and transformer area-level load curves are generated, with time slices as the horizontal axis and load values ​​as the vertical axis.

[0110] Step D4 involves conducting load characteristic analysis, extracting key characteristic indicators based on the load calculation results, and generating a load characteristic analysis report to provide in-depth data support for power grid regulation and control. Core characteristic indicators include the peak-to-valley load factor difference and load forecast accuracy. The load factor calculation formula is as follows:

[0111]

[0112] λ is the daily load factor, P t P represents the load value for each time slot. maxThe maximum load value of the day; the peak-to-valley difference is calculated using the formula ΔP = P max -P min P min Minimum load value for the day; the formula for calculating load forecast accuracy is: P t,pred It is the predicted load value, P t,actual This represents the actual load value for the same period. The analysis report is generated on a daily, weekly, and monthly basis, clearly indicating the changing trends of load characteristic indicators within each period, as well as the main user clusters and influencing factors corresponding to high load periods.

[0113] Step D5 verifies and optimizes the load calculation results, setting a verification threshold system. The load prediction accuracy must be no less than 92%, and the calculation deviation of the load factor and peak-valley difference must not exceed 5%. When the prediction accuracy is lower than the threshold, the long short-term memory network model is updated using transfer learning, and recent load data from similar transformer areas is introduced to supplement the training samples, quickly correcting the model parameters. When the calculation deviation of the load characteristic indicators exceeds the standard, the transformer area loss coefficient γ is re-verified. t The calculation is based on the power factor adjustment loss coefficient formula in the real-time operation data of the power grid; the optimized load calculation results and characteristic analysis report are simultaneously pushed to the model game module and the power grid dispatching system to complete the closed loop.

[0114] The load calculation module has real-time update and traceability functions. It synchronizes input data and updates load calculation results every 15 minutes and automatically triggers the model optimization process every 24 hours. At the same time, it records the data source parameter settings and model version of all load calculation processes, supports querying and tracing historical calculation results, and provides data basis for power grid operation analysis and fault diagnosis.

[0115] Of course, the above are just typical examples of the present invention. In addition, the present invention may have many other specific embodiments. All technical solutions formed by equivalent substitution or equivalent transformation fall within the scope of protection claimed by the present invention.

Claims

1. A load forecasting system for transformer substations based on a multi-factor game model, characterized in that, include: The data fusion module is used to access and process multi-source heterogeneous data from the transformer area measurement system, meteorological information system, power market information system and user-side management system, and to process the multi-source heterogeneous data to obtain electricity consumption related information. The factor generation module is configured with a preset factor splitting strategy. The factor splitting strategy is used to generate a corresponding user's electricity consumption association factor group based on electricity consumption association information. The electricity consumption association factor group includes several electricity consumption association factors. The user configuration module configures the target revenue function corresponding to the user based on user characteristic information, and configures a parameter optimization strategy. The parameter optimization strategy adjusts the revenue parameters of the target revenue function according to the user's historical information to update the target revenue function. The model game module is configured with an electricity consumption game strategy, which is configured in the corresponding microgrid, including: Step S1: Obtain the target revenue function and electricity consumption correlation factor for the current user; Step S2: Generate the user behavior of the current user based on the target revenue function and electricity consumption correlation factors; Step S3: Substitute the user behavior of the microgrid into the preset game model to obtain the user behavior of the next time window; Step S4: Calculate the comprehensive power load for the next time window using a preset load calculation algorithm; Step S5: Update the electricity consumption correlation factor using the preset factor update strategy and return to step S2; The load calculation module calculates the predicted load of the distribution area based on the comprehensive power load of each microgrid.

2. The transformer area load forecasting system based on multi-factor model game theory according to claim 1, characterized in that, The factor splitting strategy includes: Step A1: Configure a feature extraction sub-strategy to extract features from electricity consumption-related information to generate several electricity consumption-related features; Step A2: Traverse the electricity consumption-related features to match the corresponding anchoring features from the preset anchoring feature library; Step A3: Determine the corresponding anchoring factor based on the anchoring features using the anchoring generation sub-strategy; Step A4: Configure the factor model construction sub-strategy to construct a factor splitting model based on the anchor factor; Step A5: Input the electricity consumption correlation information into the factor splitting model to obtain dynamic factors; Step A6: Calculate the split matching value of the dynamic factor and the anchoring factor using the preset split evaluation algorithm. When the split matching value is higher than the first benchmark matching value, use the dynamic factor or the anchoring factor as the output electricity consumption correlation factor. When the split matching value is lower than the first benchmark matching value, generate model adjustment parameters to adjust the factor split model and return to step A5.

3. The transformer area load forecasting system based on a multi-factor model game as described in claim 2, characterized in that, The feature extraction sub-strategy is configured with a dynamic feature library, which stores several feature structure groups, and each feature structure group is configured with an index thread set. The feature extraction sub-strategy includes: Step A11: Preprocess each piece of electricity consumption association data in progress to unify the data format of each type of electricity consumption association data; Step A12: Match each electricity consumption-related data according to the index clue set to determine the closest structural feature group for each electricity consumption-related data; Step A13: Re-edit the power consumption association data according to the structural feature group so that the power consumption association data has the corresponding structural features; Step A14: Identify electricity consumption-related data through a pre-built feature recognition model to obtain the corresponding electricity consumption-related features.

4. The transformer area load forecasting system based on a multi-factor model game as described in claim 2, characterized in that, The anchoring generation strategy includes: Step A31: Configure the anchor matching algorithm to calculate the anchor matching value for each anchor feature; Step A32: Filter anchor features with anchor matching values ​​higher than the preset benchmark anchor value to construct an anchor feature group; Step A33: Calculate the attenuation influence value of each anchoring feature. The attenuation influence value is the sum of the attenuation correlation sub-values, which reflect the correlation between anchoring features in the anchoring feature group. Step A34: Remove the anchoring feature with the highest attenuation impact value from the anchoring feature group and return to step A33 until the preset attenuation mutual exclusion condition is met. Step A35: Obtain the corresponding anchoring factor by matching the anchoring feature with the preset anchoring factor library.

5. A transformer area load forecasting system based on a multi-factor model game as described in claim 2, characterized in that, The factor generation module further includes a splitting construction unit, which is used to construct a factor feature network. The factor feature network includes several node factors, each node factor having node factor information, including node type, node features, and node index. Factor correlation lines are configured between the node factors, reflecting the correlation between the node factors. The factor model construction sub-strategy includes: Step A41: Determine the position of each anchoring factor in the factor feature network and generate the transmission break value; Step A42: Calculate the transmission cost value corresponding to the factor correlation line of adjacent node factors. If the transmission break value is higher than the transmission cost value, obtain the corresponding node factor and use the difference between the transmission break value and the transmission cost value as the new transmission break value, until all node factors with transmission break values ​​can no longer be transmitted. Step A43: Obtain the node factors that are determined at this time to generate the factor splitting model.

6. The transformer area load forecasting system based on multi-factor model game theory according to claim 1, characterized in that, The splitting evaluation algorithm is used to weight the correlation matching item, content matching item, and topology influence item to obtain the splitting matching value. The correlation matching item reflects the degree of correlation between the determined node factors, the content matching item reflects the degree of matching between electricity consumption correlation information and the corresponding node factor, and the topology influence item reflects the degree of influence of the node factor on the topological relationship in the factor splitting model.

7. The transformer area load forecasting system based on a multi-factor model game as described in claim 1, characterized in that, The user configuration module is configured with a set of target functions, which includes several target revenue functions. Each target revenue function is configured with user matching clues. The user configuration module matches user matching clues based on user feature information to determine the target revenue function with the highest matching degree.

8. The transformer area load forecasting system based on a multi-factor model game as described in claim 7, characterized in that, The parameter optimization strategy includes: Step B1: Input the electricity consumption correlation factors from the user's historical information into the target revenue function of the current user to generate an array of expected electricity consumption revenues; Step B2: Retrieve the corresponding user behavior from the expected electricity revenue array, and generate predicted electricity consumption parameter information based on the user behavior; Step B3: Compare the predicted electricity consumption parameters with the actual electricity consumption parameters from historical data to obtain electricity consumption deviation information; Step B4: Retrieve the deviation correction table corresponding to the target revenue function based on the electricity consumption deviation information to obtain the adjustment variables and update the revenue parameters of the target revenue function.

9. A transformer area load forecasting system based on a multi-factor model game as described in claim 1, characterized in that, The load calculation algorithm obtains the weighted predicted load value and the corrected load value. The predicted load value is obtained by summing the comprehensive electricity load of all microgrids. The corrected load value is obtained by summing the cumulative deviation load and the loss corrected load. The cumulative deviation load reflects the superposition deviation of the target revenue function, and the loss corrected load reflects the load loss corresponding to each microgrid.

10. A transformer area load forecasting system based on a multi-factor model game as described in claim 1, characterized in that, The model game module includes a model building unit, which is used to build the game model. The game model is trained using historical information as samples to evolve the correlation between user behaviors in order to determine the user behavior in the next time window.

Images