Electricity trading platform user power balance matching system

Abstract

This application relates to the field of power balancing technology, specifically to a user power consumption balancing matching system for a power trading platform. The system includes a user management device that acquires user information for all users, including account number, address, and power consumption type; a power load collection device that acquires power load information for all users, including power consumption for each balancing time period, weather information for each balancing time period, and daily characteristic information; the daily characteristic information includes weekdays and non-weekdays; and a sample processing device that collects all power load information to generate user samples for each user and a total sample for all users. This application can accurately predict the power consumption of users in each time period.

Description

Technical Field

[0001] This application relates to the field of power balancing technology, and more specifically, to a user power consumption balancing matching system for a power trading platform. Background Technology

[0002] The content in this section provides only background information related to this application and may not constitute prior art.

[0003] The primary function of an electricity trading platform is to act as a bridge connecting power plants (or electricity sales companies) and electricity users (or electricity purchasers). It is used to match electricity supply (generation) with electricity demand (consumption) to maintain the stable operation of the power system and ensure a dynamic balance between total power generation and consumption.

[0004] In traditional electricity trading platforms, the platform typically forecasts and analyzes the total electricity load aggregated from all electricity users. Based on the predicted total load curve, the platform sends an overall electricity demand signal to power plants or higher-level dispatching agencies. Power plants then adjust their generation plans or output levels according to the received total demand signal to meet the aggregated electricity demand. This model essentially treats a large and diverse group of users as a single, unified load.

[0005] The aforementioned method, which uses overall user electricity consumption as the core evaluation indicator and prediction basis, ignores the significant differences and dynamic changes in electricity consumption behavior among different individual users (such as industrial users, commercial users, and residential users) or user groups at different times (such as weekdays / rest days, peak and off-peak periods) and under different climatic conditions (such as temperature, humidity, and seasonal changes). This "one-size-fits-all" aggregation approach causes the load forecasting model trained by the platform to rely excessively on historical overall data patterns, making it difficult to accurately capture and reflect the subtle and complex fluctuations in actual user electricity demand. As a result, the model suffers from severe overfitting, meaning it performs well on historical data but its accuracy decreases in actual forecasting work. Summary of the Invention

[0006] The summary section of this application is intended to provide a brief overview of the concepts, which will be described in detail in the detailed description section below. This summary section is not intended to identify key or essential features of the claimed technical solutions, nor is it intended to limit the scope of the claimed technical solutions.

[0007] Some embodiments of this application propose a user power consumption balance matching system for an electricity trading platform to solve the technical problems mentioned in the background section above.

[0008] As a first aspect of this application, some embodiments of this application provide a user electricity consumption balance matching system for an electricity trading platform, including: The user management device acquires user information for all users, including account number, address, and electricity usage type. The power load collection device acquires power load information from all users, including electricity consumption during each balanced period, weather information for each balanced period, and daily characteristic information; the daily characteristic information includes weekdays and non-weekdays. The sample processing device collects all power load information to generate user samples for each user and total samples for all users; Each user's user sample includes text tags and electricity consumption data. The text tags include address, electricity type, weather information, and daily characteristics. The electricity consumption data includes the electricity consumption for each balanced time period in the past. The total sample includes total text labels and total electricity consumption data. The total text labels include weather information and daily characteristic information, while the total electricity consumption data includes the total electricity consumption for each balanced period in the past. The sample classification device classifies user samples and total samples to add classification labels to user samples and total samples, thereby obtaining user predicted samples and total predicted samples. The electricity consumption forecasting device inputs user forecast samples and total forecast samples into a big data processing model. The big data processing model extracts user forecast information for individual users from the user forecast samples and extracts total forecast information for all users from the total forecast samples. Based on all user forecast information and total forecast information, it generates the total electricity consumption for all future balanced time periods. The information transmission device sends the total electricity consumption for all future balanced time periods to the power plant.

[0009] This application constructs a two-dimensional (individual + overall) prediction model by meticulously collecting and analyzing the electricity consumption behavior characteristics of individual users (such as electricity consumption type, geographical location, weather, date attributes, etc.) and combining them with overall load data. This effectively overcomes the shortcomings of traditional platforms that rely solely on aggregated total load for prediction, and improves the accuracy of total electricity consumption prediction for future balance periods.

[0010] Furthermore, big data processing models include: The time-series feature extraction module is used to extract the time-series features of total electricity consumption data and electricity consumption data. The tag information extraction module is used to extract the total text tags and the tag features of the text tags; The feature fusion module is used to fuse temporal features and label features to obtain fused features; The total sample prediction module regresses the total prediction information for all users from the fusion features of the total sample. The user sample prediction module regresses user prediction information for each user from the fusion features of user samples. The adversarial training module generates the total electricity consumption for all future balanced time periods based on the adversarial information of the total forecast information and the user forecast information of each user.

[0011] This application achieves refined power load forecasting by introducing a big data processing model that includes time-series feature extraction, label information extraction, feature fusion, dual-path prediction (total sample and user sample), and adversarial training. This model not only mines the time-series patterns of historical electricity consumption and the correlation features of user / environment labels, but also integrates dynamic and static attribute information through a feature fusion module. Furthermore, by utilizing the dual-path design of "total sample prediction" and "user sample prediction," combined with the dynamic calibration mechanism of the "adversarial training module," the prediction results are effectively constrained. This ensures that the sum of individual user predictions aligns with the overall prediction trend, and also uses overall information to correct potential biases in individual predictions, effectively mitigating the overfitting problem.

[0012] Furthermore, the sample classification device includes: The first type of classification module divides all user samples into several tag sample sets based on their address location and electricity usage type. The user samples in each tag sample set have the same electricity usage type and are located in the same geographical area. The second type of classification module inputs the electricity consumption data of user samples in the label sample set into the time series feature extraction module of the big data processing model to extract hidden information. Based on the similarity of the hidden information, all label sample sets are divided into several prediction sample sets, and all prediction sample sets are treated as one type of sample.

[0013] This application improves the accuracy of user group segmentation through a two-level refined classification mechanism of the sample classification device (the first type of classification module divides regions and types based on address and electricity consumption type, and the second type of classification module performs behavioral pattern clustering based on hidden information extracted from time-series features). This provides more homogeneous and discriminative training data for subsequent big data processing models (especially user sample prediction modules), effectively avoiding prediction bias caused by excessive differences within the group, and significantly improving the accuracy of individual user electricity consumption prediction.

[0014] Furthermore, the classification of sample types includes the following steps: S1: Input all total electricity consumption data into the time series feature extraction module to extract the time series features of all total electricity consumption data, and fuse the time series features of all total electricity consumption data to generate calibration features; S2: Input the electricity consumption data of the user sample into the time series feature extraction module to extract the time series features of the electricity consumption data, and calculate the similarity with the calibrated features; S3: Divide the user samples into N sample categories based on the similarity between the user samples and the labeled features; S4: Define the sample type with the largest number of user samples as the main sample, and define the remaining sample types as residual samples.

[0015] This application improves the timeliness and adaptability of user group segmentation by introducing a dynamic sample classification process based on calibration features (S1-S4). Using "calibration features" generated by fusing overall load time-series features as a dynamic benchmark, and calculating the similarity between individual user time-series features and this benchmark (S2), user clustering is achieved, which is independent of fixed regional / type labels and directly reflects the current overall system behavior pattern (S3). In particular, the mechanism of dividing users into "main samples" (the most numerous, representing the mainstream pattern) and "residual samples" (representing deviation patterns) (S4) effectively identifies and separates core groups whose electricity consumption behavior is significantly consistent with or deviates from the overall system trend. This dynamic, data-driven classification method enables the model to more sensitively capture the dominant electricity consumption patterns and abnormal fluctuations under the current system state, providing more timely and targeted input for subsequent prediction models (especially the user sample prediction module).

[0016] Furthermore, the user sample prediction module includes: The master sample prediction network generates master prediction information for user samples based on the fusion features of user samples. N-1 residual networks are used to generate residual information by taking the fusion features of different types of residual samples as input. The user information calculator uses the main prediction information of the user sample, which is defined as the main sample, as the user prediction information. The sum of the master prediction information and the residual information of the user sample, which is defined as the residual sample, is used as the user prediction information.

[0017] This application improves the accuracy and adaptability of predicting electricity consumption behavior for diverse user groups by introducing a prediction architecture consisting of a "main sample prediction network + residual network + user information calculator". This architecture, tailored to the characteristics of different sample types (main samples and various residual samples), employs a strategy where the main network handles mainstream patterns and a dedicated residual network captures specific deviation patterns. This overcomes the limitation of a single model in simultaneously handling mainstream and diverse abnormal patterns, significantly improving the prediction accuracy for electricity consumption of atypical and highly volatile residual sample users.

[0018] Furthermore, the adversarial training module includes: The separate information aggregation module adds up the user prediction information of all user samples in each equilibrium time period to obtain aggregated user information; The discriminator calculates the probability of correctly generating aggregated user information from aggregated user information and power load information; Calculate the probability of the accuracy of generating the overall forecast information from the overall forecast information and the power load information; The correct probability of the total prediction information and the correct probability of the aggregated user information are used as adversarial information. The output device uses the correct probability of aggregated user information as the aggregation weight of aggregated user information and the correct probability of total prediction information as the overall weight of total prediction information. The aggregated user information and the total prediction information are weighted and summed using aggregate weight and overall weight.

[0019] This application addresses the inconsistency between individual user prediction aggregation (aggregating user information) and overall direct prediction (total prediction information) by introducing an adversarial training module based on Generative Adversarial Networks (GAN), thereby improving the consistency and reliability of the final total electricity consumption prediction. A discriminator dynamically evaluates the credibility (correctness probability) of the prediction results of the two prediction paths (individual aggregation path and overall path) relative to the actual load (electricity load information) in each equilibrium time period. The prediction results of the two paths are then weighted and fused to generate a future total electricity consumption prediction that combines high accuracy and high robustness.

[0020] Furthermore, the temporal feature extraction module, label information extraction module, feature fusion module, and total sample prediction module undergo their first joint training to fix the parameters of the temporal feature extraction module, label information extraction module, and feature fusion module. Furthermore, the main sample prediction network and N-1 residual networks are jointly trained a second time.

[0021] This application employs an innovative two-stage joint training strategy (the first stage fixes the parameters of the basic feature extraction and fusion modules, while the second stage focuses on optimizing the user prediction network), effectively improving the overall stability, generalization ability, and training efficiency of the model. In the first joint training stage, the parameters of the temporal feature extraction module, label information extraction module, and feature fusion module are established and solidified. In the second joint training stage, resources are concentrated on optimizing the main sample prediction network and the residual network, enabling them to more accurately learn the specific electricity consumption behavior patterns and prediction compensation mechanisms of different user groups (main samples and various residual samples) on a stable, high-quality feature input basis. This ensures the long-term accuracy and reliability of user-side prediction and the resulting total electricity consumption prediction.

[0022] Furthermore, during the second joint training, a cluster center is generated for each residual network, and the clustering loss is generated by adjusting the cluster centers.

[0023] This application improves the ability of residual networks to capture and characterize the unique electricity consumption patterns of their corresponding residual sample groups by dynamically generating and optimizing dedicated cluster centers for each residual network during the second joint training, and constructing a clustering loss function accordingly.

[0024] Furthermore, a generator is defined to form an adversarial network with the discriminator. The generator outputs aggregated user information and total prediction information and is used for adversarial training against the discriminator.

[0025] This application constructs a more complete and effective generative adversarial framework by explicitly defining the generator's output as "aggregated user information" (sum of individual predictions) and "total prediction information" (overall direct prediction), and using both as input to the discriminator for adversarial training. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the user power balance matching system of the power trading platform.

[0027] Figure 2 The flowchart for prediction of the main sample.

[0028] Figure 3 This is a flowchart of the prediction process for residual samples. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments. The same reference numerals in the accompanying drawings represent the same components. It should be noted that the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the described embodiments of this application without creative effort are within the scope of protection of this application.

[0030] Compared to the embodiments shown in the accompanying drawings, feasible embodiments within the scope of this application may have fewer components, other components not shown in the drawings, different components, differently arranged components, or components with different connections, etc. Furthermore, two or more components in the drawings may be implemented in a single component, or a single component shown in the drawings may be implemented as multiple separate components.

[0031] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” and similar terms used in this specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not necessarily indicate a quantity limitation. Terms such as “upper” and “lower” are used only to indicate relative positional relationships, and these relative positional relationships may change accordingly when the absolute position of the described object changes.

[0032] Electricity trading platforms are used to match electricity supply (generation) and electricity demand (consumption) in real time to maintain the safe, stable, and efficient operation of the power system. Since neither generation nor consumption can be adjusted in a timely manner, electricity trading platforms need to predict the total electricity consumption for various future equilibrium time periods (such as 15 minutes, 30 minutes, or 1 hour). The specific methods are as follows: Example 1: Reference Figure 1 Example 1: The user electricity consumption balance matching system of the power trading platform includes a user management device, an electricity load collection device, a sample processing device, a sample classification device, an electricity consumption forecasting device, and an information transmission device. The electricity management device and the electricity load collection device are connected to the sample processing device, which is connected to the sample classification device. The electricity consumption forecasting device is connected to the sample classification device, and the information transmission device is connected to the electricity consumption forecasting device. The information transmission device sends the total electricity consumption for all future balance periods to the power plant.

[0033] The user management device and the power load collection device provide user basic information and load data to the sample processing device, respectively. The sample processing device transmits the generated samples to the sample classification device for classification. The classified samples are input into the power consumption forecasting device for forecasting calculation. The power consumption forecasting device outputs the forecasting results to the information sending device and sends them to the power plant.

[0034] The specific method is as follows: The user management device acquires user information for all users, including account number, address, and electricity usage type. For example: User A's user information is: Account Number: 202310086; Address: Building 8, Area A, High-tech Industrial Park, XX District, XX City, XX Province; Electricity type: Industrial electricity.

[0035] Electricity consumption types include not only industrial electricity consumption, but also residential electricity consumption, commercial electricity consumption, and agricultural electricity consumption.

[0036] The power load collection device acquires power load information from all users, including electricity consumption during each balanced period, weather information for each balanced period, and daily characteristic information; the daily characteristic information includes weekdays and non-weekdays. User A's user information is as follows: The balance period is from 14:00 to 15:00 on October 5th. The electricity consumption during this period is 125.8 MWh. The weather information is: temperature 32.5℃, humidity 65%. The working characteristics of the day are a weekday.

[0037] In this way, user information is used to identify users and provide a basis for static classification.

[0038] Electricity load information is used to provide dynamic electricity consumption behavior patterns and associated environmental characteristics, and is a direct data source for constructing predictive model inputs (time series features, label features).

[0039] The sample processing device collects all power load information to generate user samples for each user and total samples for all users; Each user's user sample includes predicted data and labeled data; The forecast data includes text labels and electricity consumption data. The text labels include address, electricity type, weather information, and daily characteristics. The electricity consumption data includes the electricity consumption for each balanced period in the past. The labeled data represents the actual electricity consumption for each balanced time period of the day; The total sample includes total electricity consumption data and total labeled data. Total electricity consumption data includes total text labels and total electricity consumption data. Total text labels include weather information and daily characteristic information. Total electricity consumption data includes the total electricity consumption for each balanced period in the past. The total labeled data represents the actual total electricity consumption for each balanced time period of the day; Thus, the user sample for each user includes both predicted data and labeled data.

[0040] The forecast data includes text tags and electricity consumption data.

[0041] For example, on the first day of this month, user A's text tag is: Label: Account Number: 202310086; Address: Building 8, Area A, High-tech Industrial Park, XX District, XX City, XX Province; Electricity usage type: Industrial electricity; Daily characteristics: Weekday; Weather information: Sunny Electricity consumption data is: electricity consumption per hour.

[0042] Predictive data can be used to forecast electricity consumption for each balanced period of the next day. After the next day ends, combining the actual electricity consumption data collected for the next day with today's predicted data yields a user sample. This user sample can then be used for subsequent model training.

[0043] The same applies to the total sample.

[0044] This scheme first uses user management devices and power load collection devices to collect sufficient data, which is then used to create user samples and a total sample to train the electricity consumption forecasting device. As the number of training iterations increases, the accuracy of the electricity consumption forecasting device will gradually improve.

[0045] It is foreseeable that one total sample can be collected for each forecast period (generally set to 24 hours, with power generation plans and electricity prices fluctuating daily). One user sample can also be collected for each user.

[0046] The sample classification device classifies user samples and total samples to add classification labels to user samples and total samples, thereby obtaining user predicted samples and total predicted samples.

[0047] The total sample contains information on all samples, so there is only one type and no classification is needed. However, the user sample comes from different users, and each user's electricity usage habits may be the same or different. Therefore, user samples need to be classified.

[0048] The specific plan is as follows: The sample classification device includes: The first type of classification module divides all user samples into several tag sample sets based on their address location and electricity usage type. The user samples in each tag sample set have the same electricity usage type and are located in the same geographical area. For example: User A: Address: Science and Technology Park, High-tech Zone, XX City, XX Province; Electricity type: Industrial.

[0049] User B: Address: Software Park, High-tech Zone, XX City, XX Province; Electricity usage type: Industrial.

[0050] User C: Address: Commercial Street, Old Town District, XX City, XX Province; Electricity type: Commercial.

[0051] User D: Address: Residential area of ​​the old city district of XX province, XX city; Electricity usage type: Residential.

[0052] User E: Address: Development Zone, YY City, XX Province; Electricity usage type: Industrial.

[0053] Because users A and B have the same electricity usage type and live in the same area, their user samples are assigned to the same label sample set.

[0054] Users within the same tag sample set are generally located in the same area and have the same type of electricity usage, so they are more likely to have the same electricity usage habits.

[0055] However, in reality, each user's electricity consumption habits are not necessarily similar. Therefore, further refinement is needed. This refinement is carried out on the tag sample set.

[0056] The second type of classification module inputs the electricity consumption data of user samples in the label sample set into the time series feature extraction module of the big data processing model to extract hidden information. Based on the similarity of the hidden information, all label sample sets are divided into several prediction sample sets, and all prediction sample sets are treated as one type of sample.

[0057] Thus, by adopting a two-level classification scheme, user samples can be divided into multiple sample categories.

[0058] When predicting electricity consumption, electricity consumption forecasting devices inevitably need to extract time-series features. Therefore, the time-series feature extraction module of the big data processing model within the electricity consumption forecasting device can be directly utilized to extract the hidden information of the time-series features for sample classification.

[0059] Specifically, the classification of sample types in the second type classification module includes the following steps: T1: Input all total power consumption data into the time series feature extraction module to extract the time series features of all total power consumption data, and fuse the time series features of all total power consumption data to generate calibration features.

[0060] For example, a total of 1000 samples (1000 calendar days) were collected this time. Each of these 1000 samples contains 1000 total electricity consumption data points. Inputting these 1000 total electricity consumption data points into the time series feature extraction module can extract 1000 time series features. By weighted summing of these 1000 time series features, the calibration features can be obtained.

[0061] The meaning of this calibration feature is: the electricity consumption habits of the entire region over these 1000 natural days.

[0062] T2: Input the electricity consumption data of the user sample into the time series feature extraction module to extract the time series features of the electricity consumption data, and calculate the similarity with the calibration features; T3: Divide the user samples into N sample categories based on the similarity between the user samples and the labeled features; T4: Define the sample type with the largest number of user samples as the main sample, and define the remaining sample types as residual samples.

[0063] For example, this region has a total of 10,000 users, and one user sample can be collected from each user every day. Designing a neural network model for each of these 10,000 user samples is not only costly, but also leads to data sparsity.

[0064] Therefore, these 10,000 users are categorized. The electricity consumption data of these 10,000 users are input into the time-series feature extraction module to extract the time-series features of the electricity consumption data. Then, the similarity between the data and the calibrated features is calculated, which allows for classification.

[0065] Generally, the electricity consumption data of most users has the highest similarity to the overall electricity consumption data. These users with the highest similarity are defined as primary users, and their electricity consumption samples are defined as primary samples. Electricity consumption samples with lower similarity are classified again based on their similarity to the labeled features.

[0066] In this way, the required types of residual samples can be obtained. In this scheme, the types of residual samples are set to 4.

[0067] For example, the area covered by the power trading platform can be roughly divided into two regions. Therefore, all user samples are divided into two labeled sample sets based on their addresses.

[0068] For all user samples in the first labeled sample set, the calculated temporal features show that 80% of them have a similarity of more than 80% with the temporal features of the main sample, while the remaining ones have a similarity of less than 80%. Therefore, all user samples in the first labeled sample set are divided into two sample categories based on a similarity of less than 80%.

[0069] For all user samples in the second label sample set, 50% of the calculated temporal features have a similarity of more than 80% with the temporal features of the main sample, while the remaining similarity is less than 80%. Therefore, all user samples in the second label sample set are divided into two sample categories based on a similarity of less than 80%.

[0070] Thus, a total of 2 + 2 = 4 sample categories can be obtained. Therefore, all user samples can be classified.

[0071] The above describes the sample processing method. After sample processing, a big data processing model is needed for prediction. Specifically: Big data processing models include: The time-series feature extraction module is used to extract the time-series features of total electricity consumption data and electricity consumption data. The temporal feature extraction module is an existing LSTM multilayer network that can extract the temporal features of the input data.

[0072] Since both total electricity consumption data and electricity consumption data require extraction of time-series features, a common time-series feature extraction module was set up.

[0073] ; in, This represents the computation function of the time-series feature extraction module. This represents the total electricity consumption data. This indicates the time-series characteristics of total electricity consumption data; This represents the time-series characteristics of the i-th electricity consumption data. This represents the i-th electricity consumption data.

[0074] The tag information extraction module is used to extract the total text tags and the tag features of the text tags; ; This represents the time-series characteristics of the i-th electricity consumption data. This represents the i-th electricity consumption data.

[0075] The tag information extraction module is used to extract the total text tags and the tag features of the text tags; The tag information extraction module is a pre-built dictionary encoder that can convert text tags into digital vectors.

[0076] ; This represents the calculation function of the tag information extraction module. Indicates the total text label. Tag features representing the total text tags; This represents the label feature of the i-th text label. This represents the i-th text label; Thus, the temporal feature extraction module and the label information extraction module constitute a preprocessing module. The temporal feature extraction module is used to transform continuous data into high-dimensional temporal features. The label information extraction module is used to transform text labels into numerical vectors.

[0077] The construction methods of the above-mentioned temporal feature extraction module and label information extraction module are all existing technologies, and the specific model structure will not be described further here.

[0078] The feature fusion module is used to fuse temporal features and label features to obtain fused features; ; ; This represents the fusion characteristics of the total sample. This represents the fusion feature of the i-th user sample. This indicates feature fusion.

[0079] Feature fusion can be performed in the following ways: ; This represents the concatenation operation, where W represents the weight matrix and b represents the bias vector. This represents the activation function. This represents the merged information. and This indicates two pieces of information that need to be merged.

[0080] The total sample prediction module regresses the total prediction information for all users from the fusion features of the total sample. .

[0081] This represents the overall forecast information. Represents the regression function. Indicates overall integration characteristics; The total forecast information is the power generation of all users in each average time period.

[0082] The user sample prediction module regresses user prediction information for each user from the fusion features of user samples. The user sample prediction module includes: The master sample prediction network generates master prediction information for user samples based on the fusion features of user samples. ; in, This represents the main prediction information for the i-th user. Indicating a return to the internet, This represents the fusion feature of the i-th user sample.

[0083] N-1 residual networks are used to generate residual information by taking the fusion features of different types of residual samples as input. ; This represents the residual information of the i-th user in the k-th residual network. This represents the computation function for the k-th residual network. This represents the fusion feature of the i-th user sample.

[0084] The user information calculator uses the main prediction information of the user sample, which is defined as the main sample, as the user prediction information. The sum of the master prediction information and the residual information of the user sample, defined as the residual sample, is used as the user prediction information; The information used for prediction from the main sample is: ; The user prediction information for the remaining user samples is as follows: ; This represents the user prediction information for the i-th user sample.

[0085] The user information calculator is a calculator and does not require training.

[0086] User forecast information represents the amount of electricity generated by a user in each average time period.

[0087] The essence of this solution is to use different networks to predict different types of user samples.

[0088] In other words, the prediction for each type of electricity consumption sample is divided into two parts: the main sample prediction network and the residual prediction. If the electricity consumption sample is similar to the total sample, the result of the main sample prediction network is directly used. If the electricity consumption sample is not very similar to the total sample, the corresponding residual network needs to be selected according to the requirements, and the calculated residual is added to the main prediction information to obtain the user prediction information.

[0089] The logic of the entire scheme is as follows: After collecting a number of total samples and a number of user samples, the total samples and user samples are respectively input into the temporal feature extraction module, the label information extraction module, and the feature fusion module to obtain their own fused features.

[0090] For the total sample, its own fusion features are input into the total sample prediction module to obtain the total prediction information.

[0091] For user-predicted samples, after obtaining the fused features, their own temporal features need to be extracted and input into the sample classification device for classification to obtain their own labels. Based on their own labels, the input path for the fused features is selected.

[0092] For example, if user sample A is a residual sample, after extracting the fusion features of sample A, these features need to be input into the master sample prediction network to obtain master prediction information. Then, the fusion features of sample A are input into the corresponding residual network to obtain residual information. Finally, the master prediction information and the residual information are added together to obtain the user prediction information for sample A.

[0093] After obtaining the user prediction information for all samples, summing these predictions yields the electricity consumption for each equilibrium time period. However, this electricity consumption typically differs from the total electricity consumption predicted by all samples. Therefore, an adversarial network is needed to determine which piece of information has higher confidence. Based on this, the following adversarial training module is provided.

[0094] The adversarial training module generates the total electricity consumption for all future balanced time periods based on the adversarial information of the total forecast information and the user forecast information of each user.

[0095] The adversarial training module includes: The separate information aggregation module adds up the user prediction information of all user samples in each equilibrium time period to obtain aggregated user information; For example, if we have 10 user samples, we can simply add up the electricity consumption of these 10 user samples in each balanced time period to get the total electricity consumption of these 10 user samples in each balanced time period. This information is aggregated information.

[0096] The discriminator calculates the probability of correctly generating aggregated user information from aggregated user information and power load information; Calculate the probability of the accuracy of generating the overall forecast information from the overall forecast information and the power load information; The correct probability of the total prediction information and the correct probability of the aggregated user information are used as adversarial information. The discriminator is only used to determine which is more accurate, the aggregated information or the overall prediction information. For example, if the discriminator considers the accuracy of the aggregated information to be 0.6 and the accuracy of the overall prediction information to be 0.4, then the accuracy of the aggregated information is higher.

[0097] The output device uses the correct probability of aggregated user information as the aggregation weight of aggregated user information and the correct probability of total prediction information as the overall weight of total prediction information. The aggregated user information and the total prediction information are weighted and summed using aggregate weight and overall weight.

[0098] The output unit then performs a weighted summation of aggregated user information and total forecast information to obtain a more accurate figure for electricity consumption.

[0099] ; ; in, This represents the accuracy rate calculated by the discriminator based on the total predicted information. This represents the accuracy rate calculated by the discriminator based on the aggregated information. This represents aggregated information, where i represents the user's index and I represents the total number of users. This represents the user prediction information for the i-th user.

[0100] Example 2:

[0101] Example 1: The neural network module in this scheme includes a temporal feature extraction module, a label information extraction module, a feature fusion module, a total sample prediction module, a main sample prediction network, N-1 residual networks, and a discriminator.

[0102] The temporal feature extraction module, label information extraction module, feature fusion module, total sample prediction module, main sample prediction network, N-1 residual networks, and discriminator structures provided in Example 1 are all existing technologies. Example 1 clearly defines their inputs and outputs, therefore, their outputs will not be further described. It is clear that the temporal feature extraction module is an LSTM network, the label information extraction module is a text encoder, and the feature fusion module, total sample prediction module, main sample prediction network, N-1 residual networks, and discriminator are neural networks, or parts of a neural network.

[0103] When training the temporal feature extraction module, label information extraction module, feature fusion module, total sample prediction module, main sample prediction network, N-1 residual networks, and discriminator, joint training can lead to model distortion. Therefore, distributed training is necessary. Based on this, Example 2 provides a training method. The specific method is as follows: The temporal feature extraction module, label information extraction module, feature fusion module, and total sample prediction module undergo their first joint training to fix the parameters of the temporal feature extraction module, label information extraction module, and feature fusion module. The main sample prediction network and N-1 residual networks are jointly trained a second time. During the second joint training, a cluster center is generated for each residual network, and the clustering loss is generated by adjusting the cluster centers. The generator outputs aggregated user information and total prediction information, and then it is used for adversarial training against the discriminator.

[0104] The general logic is to first use the total sample for the first joint training to obtain accurate temporal feature extraction module, label information extraction module, and feature fusion module.

[0105] Then, a second joint training is performed using user samples to obtain an accurate master sample prediction network and N-1 residual networks.

[0106] After obtaining accurate temporal feature extraction module, label information extraction module, feature fusion module, total sample prediction module, main sample prediction network, and N-1 residual networks, the discriminator and generator are then subjected to adversarial training.

[0107] To facilitate understanding, the steps for the first joint training are described below: S1: Prepare a sufficient number of total samples and divide all total samples into two independent sets: a training set and a validation set. The training set contains 7000 total samples, and the validation set contains 3000 total samples.

[0108] S2: Take a batch of total samples from the training set and input the total sample data of this batch of total samples into the time-series feature extraction module, label information extraction module, feature fusion module, and total sample prediction module to obtain the total prediction information. Calculate the loss function value based on the total prediction information and the total labeled data. Using the backpropagation algorithm, the gradient of the loss function is passed from back to front in the model chain. Use optimization algorithms (such as SGD, Adam) to update and optimize all learnable parameters of the time-series feature extraction module, label information extraction module, feature fusion module, and total sample prediction module based on the gradient information. This process completes one training iteration.

[0109] S2 repeats step 2, using all data in the training set (batch processing) to perform multiple training iterations on the model pipeline, i.e., traversing the entire training set multiple times. After each training iteration: input the validation set data that did not participate in parameter updates into the training model pipeline, allowing the model pipeline to run without updating parameters, obtaining the total prediction information of the validation set. Based on the total prediction information of the validation set and the total labeled data (true values), calculate the evaluation metrics (such as accuracy, F1 score, loss, etc.) on the validation set. Continuously repeat the above training and validation process, with the main goal of reducing the loss on the training set while monitoring the evaluation metrics on the validation set until the requirements are met.

[0110] The loss function for the first joint training is: ; Where D represents the total number of training samples, and d represents the total sample index. This represents the temporal characteristics of the d-th total sample. This represents the label feature of the d-th total sample. express and The maximum mean difference This represents the average time-series characteristic of all total samples in the current batch. This represents the average label feature of all samples in the current batch.

[0111] The reason this scheme uses the aforementioned loss function during the first joint training is primarily to optimize the consistency between temporal feature extraction and label feature extraction. The designed loss function increases the accuracy of temporal and label feature extraction. Its innovation lies in balancing the maximum mean deviation (MMD) of feature distribution differences to bring the two types of features (temporal features and label features) closer together in the high-order statistical distributions (such as variance and kurtosis) of the underlying space. At the same time, it uses a mean alignment term to eliminate the overall offset between the two types of feature vectors (e.g., label features are generally too large while temporal features are too small), thereby solving the semantic gap problem commonly encountered in fusion and avoiding the model relying solely on a single modality feature (e.g., using only temporal data while ignoring text labels). This increases the accuracy of temporal and label feature extraction.

[0112] Of course, only after the first joint training is conducted and the temporal feature extraction module is fixed can subsequent user sample type classification be performed.

[0113] Further, the training steps for the second joint training session: Z1: Prepare a sufficient number of user samples, input the electricity consumption data of all user samples into the time series feature extraction module whose parameters have been fixed after the first joint training, and extract the time series features of each user sample. Z2: Input all user samples into the sample classification device and classify all user samples into N sample types.

[0114] Z3: For user samples of each sample type, divide them into a small training set and a small validation set in a 7:3 ratio; Combine all small training sets into a complete training set, and combine all small validation sets into a complete validation set; Z4: Take a batch of user samples from the training set, and input the prediction data of this batch of user samples into the following link: temporal feature extraction module, label information extraction module, feature fusion module, main sample prediction network, N-1 residual networks, and user information calculator. Alternatively, the following link can be used: temporal feature extraction module, label information extraction module, feature fusion module, main sample prediction network, and user information calculator. Training is then performed based on the loss function.

[0115] Z4 includes the following steps: Z41: Extract a batch of user samples, and obtain the sample type for each user sample; Z42: Select the link based on the type of user sample; refer to Figure 2 and Figure 3 For example, if there are 10 types of samples, there will be 9 different residual networks. These 9 residual networks can form 10 links.

[0116] The first link consists of: temporal feature extraction module, label information extraction module, feature fusion module, main sample prediction network, and user information calculator; The second link consists of: a temporal feature extraction module, a label information extraction module, a feature fusion module, a main sample prediction network, N-1 residual networks, and a user information calculator. The third link consists of: a temporal feature extraction module, a label information extraction module, a feature fusion module, a main sample prediction network, N-1 residual networks, and a user information calculator. ...... The tenth link consists of: temporal feature extraction module, label information extraction module, feature fusion module, main sample prediction network, N-1 residual networks, and user information calculator.

[0117] Therefore, for each user sample, it is necessary to select the appropriate link based on its sample type.

[0118] Z43: The predicted data of user samples is input into the selected link to obtain user prediction information. Based on the user prediction information and labeled data, the loss function value is calculated. Using the backpropagation algorithm, the gradient calculated by the loss function is propagated from back to front in the link. Using optimization algorithms (such as SGD, Adam), all learnable parameters of the main sample prediction network and N-1 residual networks are updated and optimized based on the gradient information. This process completes one training iteration. In the N-1 residual networks, the optimization objects also include the cluster centers of each sample type.

[0119] Z5 repeats Z4, using all data in the training set (in batches) to perform multiple training iterations on the model pipeline, i.e., traversing the entire training set multiple times. After each training iteration: input the validation set data that did not participate in parameter updates into the model pipeline during training, allowing the model pipeline to run without updating parameters, and obtain the user prediction information of the validation set. Based on the user prediction information of the validation set and the labeled data (true values), calculate the evaluation metric on the validation set until the evaluation metric meets the requirements.

[0120] The specific loss function during the second joint training for: ; in, Indicates regression loss, Represents the clustering loss, where, Used to optimize the main sample prediction network and residual network. Used to optimize cluster centers and residual networks; Used to optimize the input-output mapping capability of residual networks, so that the output value of the residual network approximates the target value; This is used to optimize the geometry of the feature space so that the residual outputs of similar samples cluster around a dedicated cluster center.

[0121] ; Where C represents the total number of user samples, and c represents the index of a user sample. This represents the user prediction information for c user samples. This represents the labeled information (true value) of the c-th user sample. ; Where N represents the total number of user sample types, and k represents the index of the residual network. Let i represent the set of user samples belonging to the k-th class in the current training batch, and let i represent the index of a single user sample. This represents the k-th residual network. This represents the feature vector (main prediction information) output by the main sample prediction network for user sample i. This represents the trainable cluster center specific to the k-th class. This represents the nearest heterogeneous cluster center for user sample i. Represents the square of the Euclidean distance. This represents hyperparameters.

[0122] The optimization process of regression loss for the internal parameters of the main sample prediction network and residual network is an existing technique and will not be described here.

[0123] The key point of this application is that... The cluster centers need further updates, specifically: ; in, Indicates the new cluster center. This represents the cluster centers before the update. express The gradient information of the k-th residual network k.

[0124] After completing the first and second joint training sessions, the parameters of the temporal feature extraction module, label information extraction module, feature fusion module, total sample prediction module, main sample prediction network, and N-1 residual networks are all fixed, including the cluster centers for each sample type. At this point, user prediction information and prediction information can be generated independently. Thus, further adversarial information can be generated.

[0125] The adversarial training module includes: a separate information aggregation module, a discriminator, and an output unit. A generator is defined that forms an adversarial network with the discriminator.

[0126] The individual information aggregation module and output unit are two calculators that do not require optimization or training. The generator of the adversarial network (the generator defined above) is essentially a temporal feature extraction module, a label information extraction module, a feature fusion module, a total sample prediction module, a main sample prediction network, and N-1 residual networks.

[0127] The temporal feature extraction module, label information extraction module, feature fusion module, total sample prediction module, main sample prediction network, and N-1 residual networks were completely fixed during the first and second joint training sessions. Therefore, when implementing adversarial loss, only the discriminator needs to be trained.

[0128] The plan is as follows: SO1: Generate several real samples, several total computation samples, and several aggregated computation samples, and merge the real samples, total computation samples, and aggregated computation samples into a training set. The real samples include the electricity load information of the previous day and the electricity load information of the current day; The total calculation sample includes the previous day's power load information and the total forecast information; The aggregated calculation sample includes the previous day's electricity load information and the predicted aggregated user information; SO2: Define the total calculated sample and the aggregated calculated sample as fake samples, and randomly select one real sample and one fake sample; Real samples and fake samples are input into the discriminator, which generates the probability that a sample is real. ; ; in, This represents the discriminator model. This represents the discriminator's judgment output on the real sample. Let represent the trainable parameters of the discriminator, and x represent the real sample. Indicates a fake sample; SO3: Calculate the discriminator's loss function value and update the discriminator's network parameters based on the discriminator's loss function value.

[0129] .

[0130] Where m represents the batch size and i represents the sample index. This represents the probability that the discriminator predicts for the true situation of the i-th sample. This represents the true label of the i-th sample; SO4: Continuously iterate and optimize the discriminator until the accuracy of the discriminator is higher than the preset threshold.

[0131] The reason this application does not directly employ adversarial networks (ANNs) for model prediction is to avoid severe overfitting during repeated training. Therefore, during training, the ARN is fixed with a generator consisting of a temporal feature extraction module, a label information extraction module, a feature fusion module, a total sample prediction module, a main sample prediction network, and N-1 residual networks. Only the discriminator is trained. This increases the overall prediction accuracy of the model and achieves better results under new data characteristics.

[0132] The above are merely preferred embodiments of this application and are not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A user power balance matching system for a power trading platform, characterized by, include: The user management device acquires user information for all users, including account number, address, and electricity usage type. The power load collection device acquires power load information from all users, including electricity consumption during each balanced period, weather information during each balanced period, and daily characteristic information. The daily characteristic information includes weekdays and non-working days; The sample processing device collects all power load information to generate user samples for each user and total samples for all users; Each user's user sample includes predicted data and labeled data; The forecast data includes text labels and electricity consumption data. The text labels include address, electricity type, weather information, and daily characteristics. The electricity consumption data includes the electricity consumption for each balanced period in the past. The labeled data represents the actual electricity consumption for each balanced time period of the day; The total sample includes total sample data and total labeled data; The total sample data includes total text labels and total electricity consumption data. The total text labels include weather information and daily characteristic information, while the total electricity consumption data includes the total electricity consumption for each balanced period in the past. The total labeled data represents the actual total electricity consumption for each balanced time period of the day; The sample classification device classifies user samples and total samples to add classification labels to user samples and total samples, thereby obtaining user predicted samples and total predicted samples. The electricity consumption forecasting device inputs user forecast samples and total forecast samples into a big data processing model. The big data processing model extracts user forecast information for individual users from the user forecast samples and extracts total forecast information for all users from the total forecast samples. Based on all user forecast information and total forecast information, it generates the total electricity consumption for all future balanced time periods. The information transmission device sends the total electricity consumption for all future balanced time periods to the power plant.

2. The user of the power trading platform's electricity balance matching system according to claim 1, characterized in that, Big data processing models include: The time-series feature extraction module is used to extract the time-series features of total electricity consumption data and electricity consumption data. The tag information extraction module is used to extract the total text tags and the tag features of the text tags; The feature fusion module is used to fuse temporal features and label features to obtain fused features; The total sample prediction module regresses the total prediction information for all users from the fusion features of the total sample. The user sample prediction module regresses user prediction information for each user from the fusion features of user samples. The adversarial training module generates the total electricity consumption for all future balanced time periods based on the adversarial information of the total forecast information and the user forecast information of each user.

3. The user of the power trading platform's electricity balance matching system according to claim 2, characterized in that, The sample classification device includes: The first type of classification module divides all user samples into several tag sample sets based on their address location and electricity usage type. The user samples in each tag sample set have the same electricity usage type and are located in the same geographical area. The second type of classification module inputs the electricity consumption data of user samples in the label sample set into the time series feature extraction module of the big data processing model to extract hidden information. Based on the similarity of the hidden information, all label sample sets are divided into several prediction sample sets, and all prediction sample sets are treated as one type of sample.

4. The user of the power trading platform's electricity balance matching system according to claim 3, characterized in that, The classification of sample types includes the following steps: S1: Input all total electricity consumption data into the time series feature extraction module to extract the time series features of all total electricity consumption data, and fuse the time series features of all total electricity consumption data to generate calibration features; S2: Input the electricity consumption data of the user sample into the time series feature extraction module to extract the time series features of the electricity consumption data, and calculate the similarity with the calibrated features; S3: Divide the user samples into N sample categories based on the similarity between the user samples and the labeled features; S4: Define the sample type with the largest number of user samples as the main sample, and define the remaining sample types as residual samples.

5. The user of the power trading platform's electricity balance matching system according to claim 4, characterized in that, The user sample prediction module includes: The master sample prediction network generates master prediction information for user samples based on the fusion features of user samples. N-1 residual networks are used to generate residual information by taking the fusion features of different types of residual samples as input. The user information calculator uses the main prediction information of the user sample, which is defined as the main sample, as the user prediction information. The sum of the master prediction information and the residual information of the user sample, which is defined as the residual sample, is used as the user prediction information.

6. The user of the power trading platform's electricity balancing matching system according to claim 3, characterized in that, The adversarial training module includes: The separate information aggregation module adds up the user prediction information of all user samples in each equilibrium time period to obtain aggregated user information; The discriminator calculates the probability of correctly generating aggregated user information from aggregated user information and power load information; Calculate the probability of the accuracy of generating the overall forecast information from the overall forecast information and the power load information; The correct probability of the total prediction information and the correct probability of the aggregated user information are used as adversarial information. The output device uses the correct probability of aggregated user information as the aggregation weight of aggregated user information and the correct probability of total prediction information as the overall weight of total prediction information. The aggregated user information and the total prediction information are weighted and summed using aggregate weight and overall weight.

7. The user of the power trading platform's electricity balancing matching system according to claim 5, characterized in that, The temporal feature extraction module, label information extraction module, feature fusion module, and total sample prediction module undergo their first joint training to fix the parameters of the temporal feature extraction module, label information extraction module, and feature fusion module.

8. The user electricity consumption balance matching system of the power trading platform according to claim 7, characterized in that, The main sample prediction network and N-1 residual networks are jointly trained a second time.

9. The user electricity consumption balance matching system of the power trading platform according to claim 8, characterized in that, During the second joint training, a cluster center is generated for each residual network, and the clustering loss is generated by adjusting the cluster centers.

10. The user electricity consumption balance matching system of the power trading platform according to claim 6, characterized in that, Define a generator that forms an adversarial network with the discriminator. The generator outputs aggregated user information and total prediction information and is trained adversarially against the discriminator.

Images