Virtual power plant user incentive generation method and system based on deep user portrait

Abstract

The application discloses a virtual power plant user incentive generation method and system based on deep user portrait, and relates to the technical field of power demand side management. The method comprises the following steps: collecting multi-source heterogeneous data and performing fusion preprocessing to form a multi-source user dataset; generating a deep user portrait according to the multi-source user dataset, combining a power grid operation state and historical response records, and generating an individualized incentive strategy by using a reinforcement learning algorithm; determining an optimal push time period by using a user behavior active prediction model, generating an incentive push message by using a natural language generation method, and actively pushing the incentive push message to a user terminal; and updating the user portrait and the incentive strategy according to user response results and feedback information to form a closed-loop optimization. The application realizes accurate characterization and differentiated incentives of user side response behaviors, takes into account privacy compliance and cost efficiency, improves participation rate and response quality, and enhances the adjustable capacity and scheduling reliability of the virtual power plant.

Description

Technical Field

[0001] This invention relates to the field of power demand-side management technology, and in particular to a method and system for generating virtual power plant user incentives based on deep user profiles. Background Technology

[0002] Virtual power plants participate in electricity market transactions and grid dispatch by aggregating distributed controllable loads, energy storage, and distributed power sources. Their operational effectiveness largely depends on the participation rate and response quality of user-side resources. Current technologies for user-side participation mainly rely on price incentives or uniform subsidies based on group statistical characteristics. Economic incentives dominate, while environmental and social motivations are insufficiently utilized. The differences in user motivations and the short-term-long-term utility trade-offs are not adequately considered. Users are often described using static attributes or a limited number of statistical quantities, making it difficult to capture the temporal dependence of energy consumption behavior, changes in equipment status, and dynamic response patterns to external grid signals. Strategies are mostly formulated offline, making online updates and adaptive optimization based on user response results and feedback information difficult. Centralized training of raw user data carries the risk of data out-of-domain leakage, hindering the construction and deployment of high-quality user profiling models. There is a lack of prediction of user activity periods and contexts, leading to mismatches between push notification timing and wording, resulting in insufficient click conversion and participation. Summary of the Invention

[0003] To address the shortcomings of existing virtual power plant user incentive methods, such as their simplistic approach and inability to accurately characterize individual user behavior and preferences, this invention proposes a virtual power plant user incentive generation method and system based on deep user profiling. This method integrates multi-source heterogeneous data to create deep user profiles for individuals, combines federated learning to ensure privacy and security, and uses reinforcement learning to generate personalized incentives and behavioral activity prediction to achieve optimal push notifications.

[0004] The present invention achieves the above objectives through the following technical solutions:

[0005] A method for generating user incentives for virtual power plants based on deep user profiling, the method comprising:

[0006] Collect user energy consumption data, equipment status data, environmental perception data, user interaction data, and external power grid data of target users within the virtual power plant, and perform fusion preprocessing on the collected multi-source heterogeneous data to form a structured multi-source user dataset;

[0007] The multi-source user dataset is input into the user profile model to generate a deep user profile including a basic attribute layer, a behavioral feature layer, and a preference cognition layer. The user profile model is based on the Transformer architecture and is constructed through local training and parameter aggregation under the federated learning framework.

[0008] Based on the deep user profile, combined with the current power grid operation status and historical response records, a reinforcement learning algorithm is used to generate a personalized incentive strategy for the target user. The personalized incentive strategy includes economic incentives, environmental incentives and / or social incentives, and the corresponding incentive parameters and push time period are output.

[0009] The user behavior activity prediction model is used to predict the probability of target users' activity in the future time period, the optimal push time period of the personalized incentive strategy is determined, the incentive push message is generated using natural language generation method, and the incentive push message is actively pushed to the target user's terminal within the optimal push time period.

[0010] The system receives the target user's response and feedback information to the personalized incentive strategy, and updates the user profile model and personalized incentive strategy based on the response and feedback information to achieve closed-loop optimization.

[0011] As a preferred embodiment of the present invention, the user energy consumption data includes active power, reactive power, voltage, and current;

[0012] The device status data includes the user equipment's on / off status, operating mode, and intensity parameters;

[0013] The environmental sensing data includes indoor and outdoor temperature and humidity, light intensity, wind speed, and air quality;

[0014] The user interaction data includes user operation logs, response records, feedback ratings, and preference settings;

[0015] The external power grid data includes real-time electricity prices, time-of-use electricity prices, demand response electricity prices, and power grid load information.

[0016] As a preferred embodiment of the present invention, the step of fusing and preprocessing the collected multi-source heterogeneous data to form a structured multi-source user dataset includes:

[0017] The user energy consumption data, equipment status data, environmental perception data, user interaction data, and external power grid data are subjected to unit unification, data type conversion, and deduplication operations to ensure consistency of units and eliminate duplicate records or identification conflicts.

[0018] Based on the user-device primary key relationship, the distribution area or feeder affiliation relationship, and geographical coordinate information, perform spatial registration and entity mapping between external power grid data and environmental sensing data;

[0019] The user energy consumption data, equipment status data, environmental perception data, user interaction data, and external power grid data are uniformly aligned to the same preset timestamp reference.

[0020] For asynchronous data, sliding window interpolation is used, where: linear interpolation is used for continuous numerical data; nearest neighbor interpolation is used for discrete data; curve fitting interpolation is used for smoothly changing data; asynchronous interactive events are encoded into the corresponding time slots through timestamp vectors, and data quality identifiers that record the interpolation method and confidence level are generated.

[0021] To address missing data, a tiered imputation process is implemented: short-term missing data is imputed using the average of historical data from the same period, long-term missing data is imputed using a collaborative filtering method based on the similarity of user energy consumption behavior, and missing data caused by equipment failure is marked as invalid data.

[0022] Single-dimensional outliers are identified through statistical boundary analysis, and composite outlier patterns are identified by combining clustering outlier detection and time-series outlier detection. The detected outliers are corrected by interpolation using the mean of neighboring features.

[0023] Device status data is mapped to device status encoding vectors, semantic parsing of user interaction data is performed, and the data is converted into encoding vectors that can be used for model training.

[0024] Based on the alignment of user identifiers and timestamps, all processed multi-source heterogeneous data are uniformly merged to finally generate a structured multi-source user dataset with timestamps.

[0025] As a preferred embodiment of the present invention, the multi-source user dataset is input into a user profile model to generate a deep user profile, specifically including:

[0026] User basic features, user behavior features, and user preference features are extracted from the multi-source user dataset, and dimensionality is unified through a feature mapping layer to adapt to the model input format.

[0027] By overlaying timestamps and periodic location codes onto a numerical vector of uniform dimension, a multi-timescale input representation is constructed and fed into a Transformer encoder consisting of a multi-head self-attention mechanism, residual connections, layer normalization, and a feedforward network. This allows for the parallel prediction of the target user's response probability, response amplitude, response duration, and degree of comfort impact during a predetermined time period in the future.

[0028] The prediction results, along with the extracted user basic features, user behavior features, and user preference features, are input into the user preference modeling network. The user preference modeling network is a multi-layer fully connected neural network used to output the user's economic sensitivity score, comfort preference weight, incentive type preference strength, and optimal interaction time, thereby generating a deep user profile.

[0029] As a preferred embodiment of the present invention, the user profiling model is constructed through local training and parameter aggregation under a federated learning framework, including:

[0030] Perform the following steps on the user's local terminal:

[0031] Receive and load the global model parameters of the user profile model;

[0032] The user profile model is trained locally based on a multi-source user dataset. During the local training process, iterative updates are performed according to a set learning rate reduction strategy and a fixed batch size. For user samples in the user interaction data with fewer than a set threshold of response times, the training weight is set to twice the standard weight.

[0033] After local training is completed, privacy protection processing is performed on the locally updated model parameter gradients. The privacy protection processing includes adding Gaussian noise with an amplitude within 10% of the absolute value of the gradient to each parameter gradient according to the set privacy budget upper limit, and only selecting model parameters with a change rate exceeding 5% to be uploaded to the federated aggregation server according to the preset parameter change rate threshold.

[0034] Perform the following steps on the federated aggregation server:

[0035] The model parameter gradients uploaded by multiple user terminals are weighted and aggregated based on a federated optimization algorithm with regularization terms. The weighting coefficients are calculated based on the proportion of each user's total local training samples.

[0036] During the parameter aggregation process, a set standard deviation outlier detection method is used to identify and remove outlier parameter gradients in the aggregation results that deviate from the mean by more than 3 times the standard deviation.

[0037] The aggregated global model parameters are distributed to each user's local terminal, replacing the old local parameters and used for the next round of model updates. The user profile model is continuously optimized through iterative training of a set number of rounds without uploading the user's original data, and finally generates a deep user profile with privacy protection capabilities and supports updates.

[0038] As a preferred embodiment of the present invention, the step of generating a personalized incentive strategy for the target user using a reinforcement learning algorithm specifically includes:

[0039] The state space of reinforcement learning is defined to include deep user profiles, current power grid status, historical incentive effects, and environmental awareness data; the action space includes combinations of incentive types, incentive reinforcement parameters, and push time selection; the reward function is calculated by weighting response rate, user satisfaction, cost-effectiveness ratio, and long-term engagement.

[0040] Before the reinforcement learning algorithm selects an action, the preference cognition layer based on the deep user profile sets the action space priority: for users whose economic sensitivity is greater than a preset threshold, economic incentives are selected first, and the pricing is dynamically calculated; for users whose corresponding values ​​of environmental incentives or social incentives in the incentive type preference intensity exceed a preset threshold, environmental incentives or social incentives are selected first.

[0041] During reinforcement learning training, the incentive action to be executed is selected based on the current state space, and the cost-effectiveness balance between different incentive types is adjusted through an incentive combination optimization algorithm to determine the selected incentive type combination and incentive reinforcement parameters.

[0042] The selected combination of incentive types and incentive reinforcement parameters are applied to the target user, and an instant reward value is generated based on the target user's response rate, feedback score and power grid load regulation effect within a preset response period, and the policy network in the reinforcement learning algorithm is updated.

[0043] The proximal strategy optimization method is used to maximize the cumulative discount reward. After the strategy network converges, a personalized incentive strategy for the target user is output.

[0044] As a preferred embodiment of the present invention, the method for dynamically calculating pricing includes:

[0045] Based on target users Calculate the user price elasticity index based on economic sensitivity, historical price response data, and current economic conditions. ;

[0046] Personalized discount rate calculated based on user price elasticity index. The formula is:

[0047] ;

[0048] In the formula, It is an adjustable factor. As a market adjustment factor;

[0049] Personalized discount rate The upper limit is 20%;

[0050] Dynamically adjust personalized pricing based on time-based pricing. The formula is:

[0051]

[0052] In the formula, The benchmark price; The time-of-use pricing coefficient is selected as 1.2, 1.0, or 0.6 depending on the peak, off-peak, or valley periods.

[0053] As a preferred embodiment of the present invention, the incentive combination optimization algorithm specifically includes:

[0054] Calculate the overall incentive utility for the selected incentive type. The formula is:

[0055] ;

[0056] In the formula, For users to the first Preference weights for different incentive types; The effectiveness coefficient of the incentive type in the current state; For the first The basic utility values ​​of each incentive type; among them,

[0057] The basic utility value of economic incentives The calculation formula is:

[0058] ;

[0059] The basic utility value of environmental protection incentives The calculation formula is:

[0060] ;

[0061] The basic utility value of social incentives The calculation formula is:

[0062] ;

[0063] In the formula, For economic incentive amount, This refers to the user's annual or monthly income level. The number of carbon credits issued as environmental incentives. Weighting based on environmental preferences; Leaderboard points or ranking bonuses provided as social incentives Weights for social incentive preferences;

[0064] The constraints of the incentive combination include:

[0065] Total cost constraint: ;in, For the first The actual cost of this type of incentive The upper limit of the incentive cost budget set for the current demand response activities;

[0066] Effect balance constraint: ;in, For the first The basic utility value of each type of incentive The maximum permissible threshold for the difference in utility between incentive types;

[0067] User acceptance constraints: The complexity of incentive combinations shall not exceed the cognitive load acceptable to users, and the maximum number of incentive types is 3.

[0068] Maximize the overall incentive utility while satisfying the total cost constraint, the effect balance constraint, and the user acceptance constraint. Output stimulus type combination.

[0069] As a preferred embodiment of the present invention, the user behavior activity prediction model predicts the activity probability of a target user in a future time period based on an LSTM network. The LSTM network input includes historical login time series, user operation behavior time distribution and environmental perception data, and outputs the prediction result of the user activity probability per hour in the next 24 hours.

[0070] Based on the prediction results, a two-hour period of high activity was determined as the optimal push time period, and the push success rate was predicted based on historical response data.

[0071] Construct a push timing decision tree to determine whether the target user is online or offline based on their current status, and decide whether to push immediately or predict the next online time.

[0072] The push strategy is executed according to the urgency of the personalized incentive strategy. In high-urgency scenarios, push is sent immediately and supplemented by SMS notification. In medium-urgency scenarios, push is sent during the predicted best push time period. In low-urgency scenarios, push is sent at the time that the user is accustomed to.

[0073] Implement push frequency control, with no more than 3 pushes per day and an interval of no less than 2 hours between each push. Monitor user fatigue and suspend pushes for 24 hours after 3 consecutive rejections.

[0074] The system calls upon a dialogue template library to generate incentive push messages. Template types include incentive invitations, personalized suggestions, and feedback collection. Each template contains the user's name, incentive content, expected benefits, suggested device adjustment period, and amount saved. The system performs natural language understanding processing on the generated incentive push messages, including intent recognition, entity extraction, and sentiment analysis. It supports speech recognition, speech synthesis, speech emotion recognition, graphical interface visualization push, and real-time feedback display.

[0075] A virtual power plant user incentive generation system based on deep user profiling, the system comprising:

[0076] The data acquisition and fusion module is used to collect user energy consumption data, equipment status data, environmental perception data, user interaction data and external power grid data of target users in the virtual power plant, and perform fusion preprocessing to form a structured multi-source user dataset;

[0077] The user profile generation module is used to input the multi-source user dataset into the user profile model to generate a deep user profile including a basic attribute layer, a behavioral feature layer, and a preference cognition layer.

[0078] The personalized incentive strategy generation module is used to generate a personalized incentive strategy for the target user based on the deep user profile, combined with the current power grid operation status and historical response records, and using a reinforcement learning algorithm.

[0079] The user behavior activity prediction module is used to predict the activity probability of target users in the future time period through the user behavior activity prediction model, and to determine the optimal push time period for the personalized incentive strategy.

[0080] The incentive push module is used to generate incentive push messages using natural language generation methods, and actively push the incentive push messages to the target user terminal within the optimal push time period.

[0081] The feedback processing and update module is used to receive the target user's response results and feedback information to the personalized incentive strategy, and update the user profile model and personalized incentive strategy based on the response results and feedback information to achieve closed-loop optimization.

[0082] The beneficial effects of this invention are as follows: Multi-source heterogeneous data is fused and preprocessed before being input into a user profile model based on the Transformer architecture. This extracts behavioral features across multiple time scales, forming a deep user profile encompassing a basic attribute layer, a behavioral feature layer, and a preference cognition layer. This enables parallel prediction of the impact on individual response probability, response amplitude, response duration, and comfort, improving the predictability and controllability of user responses. The user profile model is constructed through local training and parameter aggregation within a federated learning framework, avoiding the out-of-domain transmission of raw user data and balancing privacy and model performance. Under a reinforcement learning framework, the deep user profile, along with power grid operation status and historical response records, constitutes the state space. Economic, environmental, and social incentives form the action space, and response rate, user satisfaction, cost-effectiveness ratio, and long-term engagement are used as rewards. This achieves integrated optimization decisions regarding incentive type, intensity, and push timing, effectively improving the adjustment effectiveness per unit cost. A user behavior activity prediction model predicts future activity probabilities on an hourly scale, determining the optimal push time period. Combined with natural language generation, this forms matching motivations and contextual messages, improving push click-through rates and engagement conversion rates. After receiving user response results and feedback information, the system updates user profiles and personalized incentive strategies, realizing closed-loop optimization of the entire chain from strategy generation, outreach execution to effect feedback, so that the model and strategy can continuously evolve with changes in user behavior and external conditions. Attached Figure Description

[0083] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0084] Figure 1 This is a flowchart of the method of the present invention;

[0085] Figure 2 This is a schematic diagram of the modular structure of the system in an embodiment of the present invention. Detailed Implementation

[0086] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention are within the scope of protection of the present invention.

[0087] like Figure 1As shown, this is an embodiment of the present invention, which provides a method for generating user incentives for virtual power plants based on deep user profiles, including five stages: multi-source data acquisition and processing, deep user profile modeling, personalized incentive strategy generation, predictive interaction and push, and closed-loop feedback optimization mechanism.

[0088] S1: Collect multi-source heterogeneous data of target users in the virtual power plant, including user energy consumption data, equipment status data, environmental perception data, user interaction data and external power grid data, and perform fusion preprocessing on the collected multi-source heterogeneous data to form a structured multi-source user dataset.

[0089] Hardware deployment:

[0090] Smart meters: Install smart meters that support the DL / T 645 protocol in each household to achieve 1-minute-level electricity consumption data collection;

[0091] Environmental sensors: Temperature and humidity sensors and light sensors are installed in each household to collect indoor and outdoor environmental data;

[0092] Smart sockets: Configure smart sockets for major electrical appliances (air conditioners, water heaters, washing machines, etc.) to monitor the status of the devices;

[0093] User terminals: Develop mobile apps and web interfaces, and collect user interaction data and preference feedback.

[0094] The edge computing gateway is deployed locally to achieve data preprocessing and caching; the communication protocol adopts MQTT protocol for inter-device communication and HTTP API for cloud data transmission; and the AES-256 encryption algorithm is used to protect the security of data transmission.

[0095] User energy consumption data includes electrical parameters such as active power, reactive power, voltage, current, and power factor. The equipment interfaces include smart meters, smart sockets, and energy consumption monitoring terminals. The communication protocol is the DL / T 645 meter communication protocol or the Modbus RTU protocol, and the sampling period is 1 minute high-frequency acquisition.

[0096] Equipment status data includes the on / off status, operating mode, set intensity parameters, and power consumption of controllable devices such as air conditioners, water heaters, and washing machines. Environmental sensing data includes parameters such as indoor and outdoor temperature and humidity, light intensity, wind speed, air quality, weather, season, and holidays. The sampling period is 5 minutes for timed collection and / or real-time collection triggered by key events.

[0097] User interaction data includes user operation logs, response records, feedback ratings, preference settings, and interface dwell time collected through mobile apps and web interfaces; the collection method is event-driven asynchronous collection, the data format is JSON structured data, and real-time transmission is supported.

[0098] External grid data includes real-time electricity prices, time-of-use electricity prices, demand response prices, grid load, renewable energy output forecasts, and grid dispatch instructions. Data is retrieved every 5 minutes and supports automatic API retrieval.

[0099] In this embodiment, the specific process of fusing and preprocessing the collected multi-source heterogeneous data to form a structured multi-source user dataset is as follows:

[0100] Unit unification, data type conversion and deduplication: Convert user energy consumption data (active power, reactive power, voltage, current) to a unified unit, such as kW, kVar, V, A; unify external power grid load information to MW and convert it to user dimension according to the load ratio of distribution area or feeder.

[0101] Data type conversion: All time series are converted to float32 format, and device on / off status is converted to Boolean or 0 / 1 encoding.

[0102] Deduplication: For records with duplicate timestamps, retain the data with the latest timestamp; if there is a source conflict, select the data source based on its reliability priority.

[0103] Spatial registration and entity mapping: Construct a mapping table of "user-metering point-equipment-transformer area / feeder" to register external power grid data and meteorological station environmental data to users according to the nearest neighbor of the geographical coordinates of the transformer area; if the user moves or the equipment is replaced, the mapping is re-mapped through the unique ID of the equipment and the installation location to ensure the consistency of the user dimension of all data.

[0104] Time alignment: User energy consumption data, equipment status data, environmental perception data, user interaction data, and external power grid data are uniformly resampled to a uniform time granularity (such as 15 minutes or 1 hour) to obtain a standardized time axis. Data with inconsistent time are mapped to the starting point of the nearest window through timestamps.

[0105] Asynchronous time-based interpolation processing: Continuous numerical data (such as power, voltage, temperature, and humidity) uses linear interpolation; discrete data (such as equipment on / off status) uses nearest neighbor interpolation, taking the value based on the timestamp closest to the target time point; smoothly changing data (such as electricity price curves) uses curve fitting interpolation, using cubic spline interpolation or Loess regression to fit the curve and obtain the value at the target time point; asynchronous interactive events are encoded into the corresponding time slots via timestamp vectors, and a data quality identifier recording the interpolation method and confidence level is generated (e.g., a confidence level of 0.9 indicates that the interpolation is based on a high-quality context);

[0106] Missing data imputation is tiered: short-term missing data (continuous missing data ≤ 3 time steps) is imputed using the average of historical data from the same period; long-term missing data (> 3 time steps) is imputed using a collaborative filtering method based on the similarity of user energy consumption behavior; missing data caused by equipment failure is marked as invalid data and is not included in subsequent modeling.

[0107] Anomaly Detection and Correction: Identifying one-dimensional outliers through statistical boundary analysis, i.e. It is an outlier if and only if , The mean, The standard deviation is used; combined with cluster anomaly detection and time-series anomaly detection, composite anomaly patterns are identified (K-Means clustering is used to detect samples whose behavior deviates from the group center, and LSTM time-series prediction residuals are used to test time dimension anomalies). The detected outliers are replaced with the mean or median of the nearest window, and the anomaly flag is retained for reference by the model's attention mechanism.

[0108] Feature normalization and encoding: Continuous numerical features are normalized using Z-Score or min-max normalization, while categorical features (such as on / off states and pattern types) are encoded using One-Hot encoding or embedded index encoding. A missing mask (1 = valid, 0 = filled) consistent with the feature dimensions is generated for reference when inputting into the model.

[0109] Semantic parsing and vectorization: Device status data is mapped to fixed-dimensional device status encoding vectors; user interaction data (operation logs, feedback text) is converted into text vectors through word segmentation and word embedding models (such as Word2Vec or BERT); preference-type multi-labels are represented through multi-hot encoding.

[0110] The following is an example of a device status encoding vector:

[0111] {

[0112] "Air Conditioning_Cooling": {"mode": 1, "intensity": [1,2,3,4,5]},

[0113] "Air Conditioning_Heating": {"mode": 2, "intensity": [1,2,3,4,5]},

[0114] "Air Conditioner_Off": {"mode": 0, "intensity": 0},

[0115] "Water heater_heating": {"mode": 1, "temperature": [40,45,50,55,60]},

[0116] "Water heater_insulation": {"mode": 2, "temperature": "current"},

[0117] "Energy Storage_Charging": {"mode": 1, "power": "variable"},

[0118] "Energy Storage_Discharge": {"mode": 2, "power": "variable"},

[0119] "Energy Storage_Standby": {"mode": 0, "power": 0}

[0120] }

[0121] Here is an example of user preference coding:

[0122] Comfort preference: One-hot coding (comfort priority / balance / energy saving priority);

[0123] Response time preference: Numerical coding (immediate / within 30 minutes / within 1 hour / flexible);

[0124] Incentive type preference: multi-label coding (economic / environmental / social / convenience).

[0125] Based on users' historical response records, supervised labels such as response probability, magnitude, duration, and comfort impact are constructed and strictly aligned with multi-source user datasets. Sensitive data such as user identifiers and geographic information are hashed and anonymized to generate privacy metadata, ensuring that subsequent federated learning complies with privacy constraints. All feature vectors are stacked according to time steps. Based on the alignment of user identifiers and timestamps, all processed multi-source heterogeneous data are uniformly fused to finally generate a structured multi-source user dataset with timestamps.

[0126] S2: Input multi-source user datasets into the user profile model to generate a deep user profile including a basic attribute layer, a behavioral feature layer, and a preference cognition layer. The user profile model is based on the Transformer architecture and is constructed through local training and parameter aggregation under the federated learning framework.

[0127] In this embodiment, a multi-source user dataset is input into the user profile model to generate a deep user profile, specifically including:

[0128] Extract the following three types of features from multi-source user datasets:

[0129] Basic user characteristics: total number of user devices, distribution of device types, regional and transformer area identification, and user attribute information (such as residential / commercial / industrial electricity categories).

[0130] User behavior characteristics: historical electricity consumption time-series curves, historical number of times of demand response participation, number of times equipment operation mode changes, and frequency of interaction behavior;

[0131] User preference characteristics: historical feedback ratings, comfort adjustment behavior, and response tendencies to different types of incentives.

[0132] The above three types of features are processed uniformly through a feature mapping layer, mapping different feature dimensions into numerical vectors of a unified dimension. This feature mapping layer is a fully connected network (Dense Layer), and the mapping dimension can be set to 128 dimensions to ensure consistent dimensionality when inputting into the model.

[0133] A timestamp and periodic location encoding are overlaid on a numerical vector of uniform dimension. The location encoding includes absolute time information (hour, day, week) and periodic information (weekday / holiday, seasonal cycle). Multi-timescale input representations are constructed, including sequence inputs with short-term windows (e.g., 24 hours), medium-term windows (e.g., 7 days), and long-term windows (e.g., 30 days). These multi-timescale sequences are then fed into a Transformer encoder consisting of a multi-head self-attention mechanism, residual connections, layer normalization, and a feedforward network. The multi-head self-attention mechanism is used to capture the dependencies of user behavior at different time steps.

[0134] Residual connections and layer normalization improve the stability of model training; feedforward neural networks perform nonlinear combinations of sub-features.

[0135] The Transformer encoder outputs four types of prediction results in parallel: the probability of response to the target user in the future predetermined time period (the probability of response tendency per hour in the next 24 hours), the response magnitude (the expected load adjustment capacity), the response duration (the expected duration of response maintenance), and the degree of comfort impact (the degree of impact of the response on user comfort score).

[0136] The prediction results, along with user basic features, user behavior features, and user preference features, are input into the user preference modeling network. The user preference modeling network is a multi-layer fully connected neural network, including three hidden layers (128-64-32 dimensions), with ReLU activation and a Dropout ratio of 0.2 to prevent overfitting. It outputs user economic sensitivity score (sigmoid activation, outputting a sensitivity score between 0 and 1), comfort preference weights (softmax activation, outputting the importance weights of each comfort factor), incentive type preference strength (multi-label classification, outputting the preference strength for different incentive types), and optimal interaction time (regression output, predicting the time period when users are most likely to accept push notifications).

[0137] The prediction results are fused with the output of the user preference modeling network to generate a deep user profile that includes a basic attribute layer, a behavioral feature layer, and a preference cognition layer.

[0138] The basic attribute layer includes 35 parameters, including basic user information (12 items such as family size, income level, age structure, occupation type, housing type, housing area, and length of residence), equipment configuration information (15 items such as controllable equipment type, capacity, brand, and purchase time), and basic electricity consumption characteristics (8 items such as average monthly electricity consumption, peak-valley electricity consumption ratio, load curve type, and average power factor). These parameters reflect the user's electricity consumption capacity, equipment characteristics, and basic behavioral patterns.

[0139] The behavioral characteristics layer includes 18 items such as electricity consumption behavior patterns (daily electricity consumption time distribution, equipment usage habits, load change patterns, etc.), response behavior history (9 items such as historical participation records, response success rate, response amplitude distribution, etc.), and interaction behavior characteristics (6 items such as App usage frequency, interface dwell time, feedback positivity rating, setting adjustment frequency, message click rate, and function usage diversity), totaling 33 parameters, used to characterize users' temporal behavior patterns, historical response patterns, and interaction habits.

[0140] The preference perception layer includes 10 parameters, including comfort sensitivity (temperature regulation tolerance, equipment ease of use requirements, comfort priority weight, and environmental adaptability index), economic sensitivity (price elasticity coefficient, incentive threshold, and cost-benefit attention), environmental awareness level (carbon emission reduction attention and green energy preference), and social incentive sensitivity (social competition tendency index), which are used to reflect users' response preferences under different motivations.

[0141] To improve the model's prediction accuracy and strategy generation effectiveness, this embodiment designs basic weights and task weights for each parameter: basic weights (0.5–1.6) reflect the importance of the parameter in the overall user profile, while task weights (0.8–2.1) are used to dynamically adjust the influence of different parameter subsets in a specific prediction task. Weight allocation employs an attention mechanism to automatically learn the relationships between parameters and adaptively adjusts them based on task type (response probability, response amplitude, comfort impact, etc.). Temporal decay is introduced to maintain data timeliness, and outlier detection using standard deviation is used to prevent abnormal weights from interfering with the model.

[0142] Table 1 Parameter Classification Statistics

[0143]

[0144] From the parameter classification statistics, in this embodiment, the categories of equipment configuration information, basic electricity consumption characteristics, electricity consumption behavior patterns, response behavior history, comfort sensitivity and economic sensitivity have a high proportion in terms of average weight and task weight, and are high-impact factors that the model focuses on; while the categories of social incentive sensitivity and environmental awareness, although fewer in number, also play a key role in the generation task of specific incentive strategies.

[0145] By combining the above parameter system and weighting mechanism, this embodiment can accurately characterize users' electricity consumption characteristics and response motivations in multiple scenarios, providing high-precision input for subsequent personalized incentive strategy generation and push optimization.

[0146] The user profile model in this embodiment is constructed through local training and parameter aggregation under the federated learning framework. The following steps are performed on the user's local terminal:

[0147] Receive and load the global model parameters of the user profile model distributed by the federated aggregation server;

[0148] The user profile model is trained locally using a multi-source user dataset. Iterative updates are performed according to a set learning rate reduction strategy (initial learning rate such as 0.001, decaying by 5% per training round) and a fixed batch size (such as 32 samples / batch). For user samples whose number of responses in the user interaction data is less than a set threshold, the training weight is set to twice the standard weight to increase the contribution of low-response frequency users in model training.

[0149] After local training is completed, privacy protection processing is performed on the locally updated model parameter gradients: Gaussian noise with an amplitude within 10% of the absolute value of the gradient is added to the gradient of each parameter according to the set privacy budget upper limit, and the rate of change of the model parameter gradient is calculated. Only model parameters with a rate of change exceeding 5% are selected to be uploaded to the federated aggregation server to reduce invalid update data transmission.

[0150] Perform the following steps on the federated aggregation server:

[0151] A federated optimization algorithm with regularization (such as FedProx) is used to weight and aggregate model parameter gradients uploaded from multiple user terminals. The weighting coefficients are calculated based on the proportion of each user's local training samples to ensure that users with a large sample size contribute more to the global model update. During the parameter aggregation process, a standard deviation outlier detection method is used to identify and remove outlier parameter gradients in the aggregation result that deviate from the mean by more than three standard deviations, preventing erroneous updates uploaded by a single terminal from polluting the global model.

[0152] The aggregated global model parameters are distributed to each user's local terminal, replacing the old local parameters and used for the next round of model updates. Through iterative training with a set number of rounds (such as 20 rounds), continuous optimization is achieved without uploading the user's original data, ultimately generating a deep user profile with privacy protection capabilities and support for updates.

[0153] S3: Based on the deep user profile, combined with the current power grid operation status and historical response records, use reinforcement learning algorithms to generate personalized incentive strategies for target users, including economic incentives, environmental incentives and / or social incentives, and output the corresponding incentive parameters and push time periods.

[0154] In one embodiment of the present invention, to generate a personalized incentive strategy for a target user using a reinforcement learning algorithm, the first step is to define the state space, action space, and reward function of the reinforcement learning algorithm.

[0155] In this embodiment, the state space includes a deep user profile (with 78 dimensions, including user economic sensitivity, comfort preference weight, incentive type preference intensity, etc.), the current grid state (including system load level, renewable energy output, electricity price level, etc.), and historical incentive effects (including previous...). The data includes the response rate, feedback score, cost-effectiveness ratio, etc. of the incentive strategy and environmental perception data (time-series influencing factors such as weather, season, and holidays).

[0156] The action space includes incentive type combinations (select 1–3 combinations from the economic, environmental and social incentive libraries), incentive enhancement parameters (such as price discount rate, carbon credits, leaderboard bonus value, etc.), and push time selection (accurate to the hour level).

[0157] reward function From response rate User satisfaction Cost-benefit ratio and long-term engagement The weighted calculation yields the following formula: ;

[0158] In the formula, Response rate weights are used to measure the immediate effect of incentive strategies. ; Weighted by user satisfaction, based on user feedback ratings. ; This is the cost-benefit ratio weight, which is the ratio of incentive cost to response value. ; This is a long-term engagement weight, used to measure the stability of users' continued engagement. .

[0159] Before the reinforcement learning algorithm selects an action, the preference cognition layer based on the deep user profile sets the action space priority:

[0160] When the economic sensitivity is greater than a threshold such as 0.6, economic incentives are given priority, and pricing is dynamically calculated.

[0161] When the environmental incentive preference value in the incentive type preference intensity is greater than a threshold such as 0.5, an environmental incentive is added;

[0162] When the social incentive preference value is greater than a threshold such as 0.4, social incentives are added.

[0163] Economic incentives include:

[0164] Direct subsidies: linear / tiered subsidies based on response volume (yuan / kWh);

[0165] Electricity price discount: Personalized electricity pricing based on user price sensitivity (discount rate 0-20%);

[0166] Cashback Rewards: A cumulative cashback mechanism based on the number of consecutive days of participation.

[0167] Non-economic incentives include:

[0168] Carbon credit rewards: Carbon credits are awarded based on emission reduction contributions (credits / kgCO2).

[0169] Green certification includes virtual honorary titles such as "Environmental Protection Expert" and "Energy Saving Pioneer."

[0170] Priority benefits: Enjoy priority access to new features, dedicated customer service, and other privileged services.

[0171] Social incentives include:

[0172] Ranking competition: Community energy conservation ranking, response initiative ranking;

[0173] Team Challenge: Energy-saving competitions between households and communities;

[0174] Sharing Incentives: Additional rewards for sharing your results on social media.

[0175] During reinforcement learning training, the incentive action to be executed is selected based on the current state space, and the cost-effectiveness balance between different incentive types is adjusted through an incentive combination optimization algorithm to determine the selected incentive type combination and incentive reinforcement parameters.

[0176] The selected combination of incentive types and incentive reinforcement parameters are applied to the target user, and an instant reward value is generated based on the target user's response rate, feedback score and power grid load regulation effect within a preset response period, and the policy network in the reinforcement learning algorithm is updated.

[0177] Reinforcement learning algorithms employ proximal policy optimization (PPO) to maximize cumulative discount rewards.

[0178] ;

[0179] In the formula, To reinforce the learning objective function, namely the expected cumulative discounted reward value; For policy network parameters; The expectation operator represents the expected value for all possible state transitions and action selection paths; Discount factor;

[0180] PPO ensures that the policy update magnitude is controlled and training is stable. After the policy network converges, a personalized incentive policy for the target user is output.

[0181] The methods for dynamically calculating pricing include:

[0182] Based on target users Calculate the user price elasticity index based on economic sensitivity, historical price response data, and current economic conditions. ;

[0183] Personalized discount rate calculated based on user price elasticity index. The formula is:

[0184] ;

[0185] In the formula, This is an adjustable factor, with a default value of 0.15; It serves as a market adjustment factor, reflecting the current supply and demand situation;

[0186] Personalized discount rate The upper limit is 20% to prevent the risk of cross-subsidization;

[0187] Dynamically adjust personalized pricing based on time-based pricing. The formula is:

[0188] ;

[0189] In the formula, The benchmark price; The time-of-use pricing coefficient is 1.2, 1.0, or 0.6, depending on the peak (8:00-11:00, 18:00-21:00), average (6:00-8:00, 11:00-18:00, 21:00-23:00), and trough (23:00-6:00) time periods.

[0190] The specific incentive combinatorial optimization algorithm includes:

[0191] Calculate the overall incentive utility for the selected incentive type. The formula is:

[0192] ;

[0193] In the formula, For users to the first Preference weights for different incentive types; The effectiveness coefficient of the incentive type in the current state; For the first The basic utility values ​​of each incentive type; among them,

[0194] The basic utility value of economic incentives The calculation formula is:

[0195] ;

[0196] The basic utility value of environmental protection incentives The calculation formula is:

[0197] ;

[0198] The basic utility value of social incentives The calculation formula is:

[0199] ;

[0200] In the formula, For economic incentive amount, This refers to the user's annual or monthly income level. The number of carbon credits issued as environmental incentives. Weighting based on environmental preferences; Leaderboard points or ranking bonuses provided as social incentives Weights for social incentive preferences;

[0201] The constraints of the incentive combination include:

[0202] Total cost constraint: ;in, For the first The actual cost of this type of incentive The upper limit of the incentive cost budget set for the current demand response activities;

[0203] Effect balance constraint: ;in, For the first The basic utility value of each type of incentive The maximum permissible threshold for the difference in utility between incentive types;

[0204] User acceptance constraints: The complexity of incentive combinations shall not exceed the cognitive load acceptable to users, and the maximum number of incentive types is 3.

[0205] Maximize the overall incentive effect while satisfying the constraints of total cost, effect balance, and user acceptance. Output the combination of incentive types. The final incentive combination should select no more than 3 incentive types.

[0206] S4: Predict the probability of target users' activity in the future time period through a user behavior activity prediction model, determine the optimal push time period for personalized incentive strategies, generate incentive push messages using natural language generation methods, and actively push incentive push messages to target user terminals within the optimal push time period.

[0207] In this embodiment, the user behavior activity prediction model predicts the probability of target users' activity in the future time period based on the LSTM network. The LSTM network input includes historical login time series (past 30 days), user operation behavior time distribution, environmental perception data (weekdays, weekends, holidays and weather conditions), and outputs the prediction result of the user activity probability per hour in the next 24 hours. Based on the prediction result, a high-activity period of 2 consecutive hours is determined as the best push time period, and the push success rate is predicted based on historical response data.

[0208] Construct a push timing decision tree to determine whether the target user is online or offline based on their current status, and decide whether to push immediately or predict the next online time.

[0209] The push strategy is executed according to the urgency of the personalized incentive policy. In high-urgency scenarios, push notifications are sent immediately and supplemented by SMS notifications. In medium-urgency scenarios, push notifications are sent during the predicted optimal push time period. In low-urgency scenarios, push notifications are sent at the time that the user is accustomed to. Push frequency is controlled, with no more than 3 pushes per day and an interval of no less than 2 hours between each push. User fatigue is monitored, and push notifications are suspended for 24 hours after 3 consecutive rejections.

[0210] The system uses a dialogue template library to generate incentivized push messages. Template types include incentivized invitations, personalized suggestions, and feedback collection. Templates include parameters such as user name, incentive content, expected benefits, suggested device adjustment period, and savings amount. An example of the dialogue template library is shown below.

[0211] Incentive-based invitations:

[0212] "Hi {username}, you can now participate in the energy-saving response and receive {incentive content}, which is expected to save you {expected benefits} yuan. Do you want to participate?"

[0213] Personalized suggestions:

[0214] Based on your electricity usage habits, we recommend adjusting the device name to the recommended setting during the specified time period. This can save you [amount] yuan in electricity costs.

[0215] Feedback collection category:

[0216] Thank you for participating in this response! Please rate this experience (1-5 stars). Your feedback will help us provide better service.

[0217] The generated incentive push messages undergo natural language understanding processing, including intent recognition (participation willingness, reasons for refusal, suggestions and feedback, complaints and inquiries), entity extraction (key information such as time, device, amount, satisfaction, etc.), and sentiment analysis (identification of positive, negative, and neutral sentiment tendencies) to enhance the personalization of push content. It supports speech recognition (supports Chinese voice command recognition), speech synthesis (personalized voice broadcast, supporting dialect adaptation), voice emotion recognition (identifying the emotional state in the user's voice), graphical interface visualization push (charts showing response effects and benefits, intuitive device control and preference settings interface), and real-time feedback display (real-time status display of the response process) to achieve multimodal interaction.

[0218] S5: Receive the target user's response results and feedback information on the personalized incentive strategy, and update the user profile model and personalized incentive strategy based on the response results and feedback information to achieve closed-loop optimization.

[0219] The method in this embodiment can be run directly on a virtual power plant platform in residential communities. However, for the specific needs of commercial building users, it needs to be adapted and optimized based on the implementation in residential communities.

[0220] Key points for implementing differentiation include:

[0221] In terms of data acquisition, it connects to the Building Automation System (BAS) to obtain detailed operating data of equipment such as air conditioning, lighting, and elevators; it integrates the Energy Management System (EMS) to achieve itemized metering and refined monitoring; and it adds commercial operation data such as business hours, customer flow, and indoor air quality.

[0222] In terms of user profile models, business operation characteristics (business type, operation scale, energy consumption intensity, etc.) can be added, taking into account the cost sensitivity and operational continuity requirements of business users, and introducing an assessment of the collaborative response capability of equipment groups.

[0223] The incentive strategy includes commercial incentives such as green certification and carbon emission reduction certifications (CERs), optimizes response time based on minimizing operational impact, and provides value-added services such as energy consumption analysis reports and energy-saving suggestions.

[0224] like Figure 2 As shown, another embodiment of the present invention provides a virtual power plant user incentive generation system based on deep user profiling, comprising:

[0225] The data acquisition and fusion module is used to collect user energy consumption data, equipment status data, environmental perception data, user interaction data and external power grid data of target users in the virtual power plant, and perform fusion preprocessing to form a structured multi-source user dataset;

[0226] The user profile generation module is used to input multi-source user datasets into the user profile model to generate a deep user profile including a basic attribute layer, a behavioral feature layer, and a preference cognition layer.

[0227] The personalized incentive strategy generation module is used to generate personalized incentive strategies for target users based on deep user profiles, combined with the current power grid operation status and historical response records, and using reinforcement learning algorithms.

[0228] The user behavior activity prediction module is used to predict the probability of target users' activity in the future time period through the user behavior activity prediction model, and to determine the optimal push time period for personalized incentive strategies.

[0229] The incentive push module is used to generate incentive push messages using natural language generation methods and actively push the incentive push messages to the target user's terminal within the optimal push time period.

[0230] The feedback processing and update module is used to receive the response results and feedback information of the target users to the personalized incentive strategy, and update the user profile model and personalized incentive strategy based on the response results and feedback information to achieve closed-loop optimization.

[0231] In summary, this invention achieves accurate characterization and differentiated incentives for user-side response behavior through a unified framework of "multi-source data fusion preprocessing—Transformer deep user profiling—federated learning for privacy protection—reinforcement learning for personalized incentives—active prediction for optimal push—feedback closed-loop optimization," balancing privacy compliance and cost-effectiveness. While improving participation rate and response quality, it also enhances the adjustable capacity and scheduling reliability of the virtual power plant, forming a highly efficient user incentive generation and interaction mechanism that is scalable across multiple scenarios.

[0232] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various variations or substitutions within the technical scope disclosed in this application, and these should all be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for generating user incentives for virtual power plants based on deep user profiling, characterized in that, The method includes: Collect user energy consumption data, equipment status data, environmental perception data, user interaction data, and external power grid data of target users within the virtual power plant, and perform fusion preprocessing on the collected multi-source heterogeneous data to form a structured multi-source user dataset; The multi-source user dataset is input into the user profile model to generate a deep user profile including a basic attribute layer, a behavioral feature layer, and a preference cognition layer. The user profile model is based on the Transformer architecture and is constructed through local training and parameter aggregation under the federated learning framework. Based on the deep user profile, combined with the current power grid operation status and historical response records, a reinforcement learning algorithm is used to generate a personalized incentive strategy for the target user. The personalized incentive strategy includes economic incentives, environmental incentives and / or social incentives, and the corresponding incentive parameters and push time period are output. The personalized incentive strategy for generating target users using reinforcement learning algorithms specifically includes: The state space of reinforcement learning is defined to include deep user profiles, current power grid status, historical incentive effects, and environmental awareness data; the action space includes combinations of incentive types, incentive reinforcement parameters, and push time selection; the reward function is calculated by weighting response rate, user satisfaction, cost-effectiveness ratio, and long-term engagement. Before the reinforcement learning algorithm selects an action, the preference cognition layer based on the deep user profile sets the action space priority: for users whose economic sensitivity is greater than a preset threshold, economic incentives are selected first, and the pricing is dynamically calculated; for users whose corresponding values ​​of environmental incentives or social incentives in the incentive type preference intensity exceed a preset threshold, environmental incentives or social incentives are selected first. During reinforcement learning training, the incentive action to be executed is selected based on the current state space, and the cost-effectiveness balance between different incentive types is adjusted through an incentive combination optimization algorithm to determine the selected incentive type combination and incentive reinforcement parameters. The selected combination of incentive types and incentive reinforcement parameters are applied to the target user, and an instant reward value is generated based on the target user's response rate, feedback score and power grid load regulation effect within a preset response period, and the policy network in the reinforcement learning algorithm is updated. The proximal strategy optimization method is used to maximize the cumulative discount reward. After the strategy network converges, a personalized incentive strategy for the target user is output. The user behavior activity prediction model is used to predict the probability of target users' activity in the future time period, the optimal push time period of the personalized incentive strategy is determined, the incentive push message is generated using natural language generation method, and the incentive push message is actively pushed to the target user's terminal within the optimal push time period. The system receives the target user's response and feedback information to the personalized incentive strategy, and updates the user profile model and personalized incentive strategy based on the response and feedback information to achieve closed-loop optimization.

2. The method for generating virtual power plant user incentives based on deep user profiling according to claim 1, characterized in that, The user energy consumption data includes active power, reactive power, voltage, and current; The device status data includes the user equipment's on / off status, operating mode, and intensity parameters; The environmental sensing data includes indoor and outdoor temperature and humidity, light intensity, wind speed, and air quality; The user interaction data includes user operation logs, response records, feedback ratings, and preference settings; The external power grid data includes real-time electricity prices, time-of-use electricity prices, demand response electricity prices, and power grid load information.

3. The method for generating virtual power plant user incentives based on deep user profiling according to claim 1, characterized in that, The process of fusing and preprocessing the collected multi-source heterogeneous data to form a structured multi-source user dataset includes: The user energy consumption data, equipment status data, environmental perception data, user interaction data, and external power grid data are subjected to unit unification, data type conversion, and deduplication operations to ensure consistency of units and eliminate duplicate records or identification conflicts. Based on the user-device primary key relationship, the distribution area or feeder affiliation relationship, and geographical coordinate information, perform spatial registration and entity mapping between external power grid data and environmental sensing data; The user energy consumption data, equipment status data, environmental perception data, user interaction data, and external power grid data are uniformly aligned to the same preset timestamp reference. For asynchronous data, sliding window interpolation is used, where: linear interpolation is used for continuous numerical data; nearest neighbor interpolation is used for discrete data; curve fitting interpolation is used for smoothly changing data; asynchronous interactive events are encoded into the corresponding time slots through timestamp vectors, and data quality identifiers that record the interpolation method and confidence level are generated. To address missing data, a tiered imputation process is implemented: short-term missing data is imputed using the average of historical data from the same period, long-term missing data is imputed using a collaborative filtering method based on the similarity of user energy consumption behavior, and missing data caused by equipment failure is marked as invalid data. Single-dimensional outliers are identified through statistical boundary analysis, and composite outlier patterns are identified by combining clustering outlier detection and time-series outlier detection. The detected outliers are corrected by interpolation using the mean of neighboring features. Device status data is mapped to device status encoding vectors, semantic parsing of user interaction data is performed, and the data is converted into encoding vectors that can be used for model training. Based on the alignment of user identifiers and timestamps, all processed multi-source heterogeneous data are uniformly merged to finally generate a structured multi-source user dataset with timestamps.

4. The method for generating virtual power plant user incentives based on deep user profiling according to claim 3, characterized in that, The multi-source user dataset is input into the user profiling model to generate a deep user profile, specifically including: User basic features, user behavior features, and user preference features are extracted from the multi-source user dataset, and dimensionality is unified through a feature mapping layer to adapt to the model input format. By overlaying timestamps and periodic location codes onto a numerical vector of uniform dimension, a multi-timescale input representation is constructed and fed into a Transformer encoder consisting of a multi-head self-attention mechanism, residual connections, layer normalization, and a feedforward network. This allows for the parallel prediction of the target user's response probability, response amplitude, response duration, and degree of comfort impact during a predetermined time period in the future. The prediction results, along with the extracted user basic features, user behavior features, and user preference features, are input into the user preference modeling network. The user preference modeling network is a multi-layer fully connected neural network used to output the user's economic sensitivity score, comfort preference weight, incentive type preference strength, and optimal interaction time, thereby generating a deep user profile.

5. The method for generating virtual power plant user incentives based on deep user profiling according to claim 4, characterized in that, The user profiling model is constructed under a federated learning framework through local training and parameter aggregation, including: Perform the following steps on the user's local terminal: Receive and load the global model parameters of the user profile model; The user profile model is trained locally based on a multi-source user dataset. During the local training process, iterative updates are performed according to a set learning rate reduction strategy and a fixed batch size. For user samples in the user interaction data with fewer than a set threshold of response times, the training weight is set to twice the standard weight. After local training is completed, privacy protection processing is performed on the locally updated model parameter gradients. The privacy protection processing includes adding Gaussian noise with an amplitude within 10% of the absolute value of the gradient to each parameter gradient according to the set privacy budget upper limit, and only selecting model parameters with a change rate exceeding 5% to be uploaded to the federated aggregation server according to the preset parameter change rate threshold. Perform the following steps on the federated aggregation server: The model parameter gradients uploaded by multiple user terminals are weighted and aggregated based on a federated optimization algorithm with regularization terms. The weighting coefficients are calculated based on the proportion of each user's total local training samples. During the parameter aggregation process, a set standard deviation outlier detection method is used to identify and remove outlier parameter gradients in the aggregation results that deviate from the mean by more than 3 times the standard deviation. The aggregated global model parameters are distributed to each user's local terminal, replacing the old local parameters and used for the next round of model updates. The user profile model is continuously optimized through iterative training of a set number of rounds without uploading the user's original data, and finally generates a deep user profile with privacy protection capabilities and supports updates.

6. The method for generating virtual power plant user incentives based on deep user profiling according to claim 1, characterized in that, The method for dynamically calculating pricing includes: Based on target users Calculate the user price elasticity index based on economic sensitivity, historical price response data, and current economic conditions. ; Personalized discount rate calculated based on user price elasticity index. The formula is: ； In the formula, It is an adjustable factor. As a market adjustment factor; Personalized discount rate The upper limit is 20%; Dynamically adjust personalized pricing based on time-based pricing. The formula is: ； In the formula, The benchmark price; The time-of-use pricing coefficient is selected as 1.2, 1.0, or 0.6 depending on the peak, off-peak, or valley periods.

7. The method for generating virtual power plant user incentives based on deep user profiling according to claim 1, characterized in that, The incentive combination optimization algorithm specifically includes: Calculate the overall incentive utility for the selected incentive type. The formula is: ； In the formula, For users to the first Preference weights for different incentive types; The effectiveness coefficient of the incentive type in the current state; For the first The basic utility values ​​of each incentive type; among them, The basic utility value of economic incentives The calculation formula is: ； The basic utility value of environmental protection incentives The calculation formula is: ； The basic utility value of social incentives The calculation formula is: ； In the formula, For economic incentive amount, This refers to the user's annual or monthly income level. The number of carbon credits issued as environmental incentives. Weighting based on environmental preferences; Leaderboard points or ranking bonuses provided as social incentives Weights for social incentive preferences; The constraints of the incentive combination include: Total cost constraint: ;in, For the first The actual cost of this type of incentive The upper limit of the incentive cost budget set for the current demand response activities; Effect balance constraint: ;in, For the first The basic utility value of each type of incentive The maximum permissible threshold for the difference in utility between incentive types; User acceptance constraints: The complexity of incentive combinations shall not exceed the cognitive load acceptable to users, and the maximum number of incentive types is 3. Maximize the overall incentive utility while satisfying the total cost constraint, the effect balance constraint, and the user acceptance constraint. Output stimulus type combination.

8. The method for generating virtual power plant user incentives based on deep user profiling according to claim 1, characterized in that, The user behavior activity prediction model is based on an LSTM network to predict the activity probability of a target user in the future time period. The LSTM network input includes historical login time series, user operation behavior time distribution and environmental perception data, and outputs the prediction result of the user activity probability per hour in the next 24 hours. Based on the prediction results, a two-hour period of high activity was determined as the optimal push time period, and the push success rate was predicted based on historical response data. Construct a push timing decision tree to determine whether the target user is online or offline based on their current status, and decide whether to push immediately or predict the next online time. The push strategy is executed according to the urgency of the personalized incentive strategy. In high-urgency scenarios, push is sent immediately and supplemented by SMS notification. In medium-urgency scenarios, push is sent during the predicted best push time period. In low-urgency scenarios, push is sent at the time that the user is accustomed to. Implement push frequency control, with no more than 3 pushes per day and an interval of no less than 2 hours between each push. Monitor user fatigue and suspend pushes for 24 hours after 3 consecutive rejections. The system calls upon a dialogue template library to generate incentive push messages. Template types include incentive invitations, personalized suggestions, and feedback collection. Each template contains the user's name, incentive content, expected benefits, suggested device adjustment period, and amount saved. The system performs natural language understanding processing on the generated incentive push messages, including intent recognition, entity extraction, and sentiment analysis. It supports speech recognition, speech synthesis, speech emotion recognition, graphical interface visualization push, and real-time feedback display.

9. A virtual power plant user incentive generation system based on deep user profiling, applied to the virtual power plant user incentive generation method based on deep user profiling as described in any one of claims 1-8, characterized in that, The system includes: The data acquisition and fusion module is used to collect user energy consumption data, equipment status data, environmental perception data, user interaction data and external power grid data of target users in the virtual power plant, and perform fusion preprocessing to form a structured multi-source user dataset; The user profile generation module is used to input the multi-source user dataset into the user profile model to generate a deep user profile including a basic attribute layer, a behavioral feature layer, and a preference cognition layer. The personalized incentive strategy generation module is used to generate a personalized incentive strategy for the target user based on the deep user profile, combined with the current power grid operation status and historical response records, and using a reinforcement learning algorithm. The user behavior activity prediction module is used to predict the activity probability of target users in the future time period through the user behavior activity prediction model, and to determine the optimal push time period for the personalized incentive strategy. The incentive push module is used to generate incentive push messages using natural language generation methods, and actively push the incentive push messages to the target user terminal within the optimal push time period. The feedback processing and update module is used to receive the target user's response results and feedback information to the personalized incentive strategy, and update the user profile model and personalized incentive strategy based on the response results and feedback information to achieve closed-loop optimization.

Images