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

Abstract

The invention discloses a virtual power plant user incentive generation method and system based on a deep user portrait, and relates to the technical field of power demand side management. The method comprises the steps of collecting multi-source heterogeneous data and performing fusion preprocessing to form a multi-source user data set; generating a deep user portrait according to the multi-source user data set, and generating a personalized incentive strategy by using a reinforcement learning algorithm in combination with a power grid operation state and a historical response record; determining an optimal push time period through a user behavior activity 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 the user response result and the feedback information to form closed-loop optimization. According to the method, accurate description and differential excitation of user side response behaviors are realized, privacy compliance and cost effectiveness are considered, and the adjustable capacity and scheduling reliability of the virtual power plant are enhanced while the participation rate and the response quality are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power demand side management, and in particular to a method and system for generating user incentives for a virtual power plant based on deep user profiling. Background Art

[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 depends largely on the participation rate and response quality of user-side resources. Existing technologies primarily target user-side participation through price incentives or uniform subsidies based on demographic characteristics. Economic incentives dominate, while environmental and social factors are underutilized. Differences in user motivations and the short-term-long-term utility trade-offs are insufficiently considered. Users are often described using static attributes or a small number of statistics, making it difficult to capture the temporal dependencies of energy consumption behavior, changes in device state, and dynamic responses to external grid signals. Strategies are often developed offline, making it difficult to perform online updates and adaptive optimization based on user responses and feedback. Centralized training of raw user-side data carries the risk of data out-of-domain and leakage, hindering the construction and deployment of high-quality profiling models. There is a lack of prediction of user activity periods and contexts, resulting in a mismatch between push timing and messaging, leading to insufficient click-through conversions and engagement. Summary of the Invention

[0003] In response to the problems that the existing virtual power plant user incentive methods are single and cannot accurately characterize the individual behavioral characteristics and preferences of users, the present invention proposes a virtual power plant user incentive generation method and system based on deep user profiling. It can integrate multi-source heterogeneous data, realize deep user profiling for individuals, combine federated learning to ensure privacy and security, and use reinforcement learning to achieve personalized incentive generation and behavioral activity prediction to achieve optimal push.

[0004] The present invention achieves the above-mentioned purpose through the following technical solutions:

[0005] A method for generating user incentives for a virtual power plant 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 from 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 data set;

[0007] Input the multi-source user dataset into the user profile model to generate a deep user profile consisting of a basic attribute layer, a behavioral feature layer, and a preference recognition 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 in-depth user profile, combined with the current grid operating 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] Predicting the target user's activity probability in the future time period through a user behavior activity prediction model, determining the optimal push time period for the personalized incentive strategy, generating an incentive push message using a natural language generation method, and actively pushing the incentive push message to the target user terminal within the optimal push time period;

[0010] Receive the target user's response results and feedback information to the personalized incentive strategy, and update the user portrait model and personalized incentive strategy based on the response results and feedback information to achieve closed-loop optimization.

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

[0012] The device status data includes the switch status, operation mode and strength parameters of the user device;

[0013] The environmental perception 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 scores 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 solution of the present invention, the fusion preprocessing of the collected multi-source heterogeneous data to form a structured multi-source user data set includes:

[0017] Perform unit unification, data type conversion, and deduplication operations on the user energy consumption data, device status data, environmental perception data, user interaction data, and external power grid data to ensure dimensional consistency and eliminate duplicate records or identification conflicts;

[0018] Based on the primary key relationship between users and devices, the ownership relationship of substations or feeders, and geographic coordinate information, perform spatial registration and entity mapping of external power grid data and environmental perception data;

[0019] Aligning the user energy usage data, device status data, environmental perception data, user interaction data, and external power grid data to a preset identical timestamp reference;

[0020] Sliding window interpolation is used for time-asynchronous data, including linear interpolation for continuous numerical data, nearest neighbor interpolation for discrete data, and curve fitting interpolation for smoothly varying data. Asynchronous interaction events are encoded into corresponding time slots via timestamp vectors, and data quality indicators are generated to record the interpolation method and confidence level.

[0021] Data missing data is filled in a hierarchical manner: short-term missing data is filled using the mean of historical data for the same period, long-term missing data is filled using collaborative filtering 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 anomaly patterns are identified by combining cluster anomaly detection and time series anomaly detection. Detected outliers are corrected by interpolation of neighboring feature means.

[0023] Map device state data into device state encoding vectors, perform semantic analysis on user interaction data, and convert it into encoding vectors that can be used for model training;

[0024] Based on the alignment of user identification and timestamp, all processed multi-source heterogeneous data are unified and fused to finally generate a structured multi-source user dataset with timestamp.

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

[0026] Extracting user basic features, user behavior features, and user preference features from the multi-source user dataset, and performing dimension-unified processing on each of them through a feature mapping layer to adapt to the model input format;

[0027] By superimposing timestamps and periodic position codes on a numerical vector of uniform dimension, a multi-timescale input representation is constructed. This input is fed into a Transformer encoder consisting of a multi-head self-attention mechanism, residual connections, layer normalization, and a feedforward network. This encoder then concurrently predicts the target user's response probability, response magnitude, response duration, and comfort impact within a predetermined time period in the future.

[0028] The prediction results, together 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 that outputs the user's economic sensitivity score, comfort preference weight, incentive type preference intensity, and optimal interaction time, thereby generating an in-depth user profile.

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

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

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

[0032] Locally training the user profile model based on a multi-source user dataset, performing iterative updates according to a set learning rate decrement strategy and a fixed batch size during the local training process, and setting a training weight of twice the standard weight for user samples whose response times in the user interaction data are less than a set threshold;

[0033] After local training is completed, privacy protection is performed on the locally updated model parameter gradients. The privacy protection process 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% based on a preset parameter change rate threshold and uploading them to the federated aggregation server;

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

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

[0036] During the parameter aggregation process, the set standard deviation outlier detection method is used to identify and eliminate abnormal parameter gradients in the aggregation results that deviate from the mean by more than 3 standard deviations;

[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 portrait model is continuously optimized through a set number of iterative training rounds without uploading the user's original data, and ultimately generates a deep user portrait with privacy protection capabilities and support for updates.

[0038] As a preferred solution of the present invention, the personalized incentive strategy for target users generated by using a reinforcement learning algorithm specifically includes:

[0039] The state space for reinforcement learning is defined to include deep user profiles, current grid status, historical incentive effects, and environmental perception data. The action space includes incentive type combinations, incentive reinforcement parameters, and push timing. The reward function is calculated by weighting response rate, user satisfaction, cost-effectiveness, and long-term engagement.

[0040] Before the reinforcement learning algorithm selects an action, the action space priority is set based on the preference recognition layer of the deep user profile: for users whose economic sensitivity exceeds a preset threshold, economic incentives are given priority, and pricing is calculated dynamically; 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 given priority;

[0041] During the reinforcement learning training process, 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 the incentive combination optimization algorithm to determine the selected incentive type combination and incentive reinforcement parameters;

[0042] Apply the selected incentive type combination and incentive reinforcement parameters to the target user, and generate an immediate reward value based on the target user's response rate, feedback score, and grid load regulation effect within the preset response period, and update the policy network in the reinforcement learning algorithm;

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

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

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

[0046] Calculate personalized discount rate based on user price elasticity index , the formula is:

[0047] ;

[0048] Where, is the adjustable factor, is the market adjustment factor;

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

[0050] Dynamically adjust personalized pricing based on time-sharing pricing , the formula is:

[0051]

[0052] Where, is the base price; This is the time-sharing pricing coefficient, and 1.2, 1.0 or 0.6 are selected according to the peak, flat or valley periods.

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

[0054] Calculate the comprehensive incentive effect for the selected incentive type , the formula is:

[0055] ;

[0056] Where, For users Preference weights of incentive types; is the effectiveness coefficient of the incentive type in the current state; For the The basic utility value of the incentive type; among them,

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

[0058] ;

[0059] 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] Where, is the economic incentive amount, The user's annual or monthly income level; The number of carbon credits issued for environmental protection incentives, is the environmental preference weight; Ranking points or ranking bonus values ​​provided for social incentives, weighting preferences for social incentives;

[0064] The constraints of the incentive package include:

[0065] Total cost constraint: ;in, For the The actual cost of each incentive type, A budget cap on incentive costs for current demand response activities;

[0066] Effect balance constraints: ;in, For the The basic utility value of the incentive type, is the maximum allowed threshold of the difference in utility between incentive types;

[0067] User acceptance constraints: The complexity of the incentive combination must not exceed the user's acceptable cognitive load, and the maximum number of incentive types is 3;

[0068] Maximize the comprehensive incentive utility under the premise of satisfying the total cost constraint, effect balance constraint and user acceptance constraint , output excitation type combination.

[0069] As a preferred solution of the present invention, the user behavior activity prediction model predicts the target user's activity probability in the future time period based on the 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's activity probability per hour in the next 24 hours;

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

[0071] Build 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] Execute push strategies based on the urgency of personalized incentive policies. In high-urgency scenarios, push notifications are sent immediately and supplemented with SMS notifications. In medium-urgency scenarios, push notifications are sent during the predicted optimal time period. In low-urgency scenarios, push notifications are sent at times that users are accustomed to.

[0073] Implement push frequency control, with no more than three pushes per day, and a minimum of two hours between pushes. Monitor user fatigue and suspend pushes for 24 hours after three consecutive rejections.

[0074] Call the dialogue template library to generate incentive push messages. Template types include incentive invitation, personalized suggestion and feedback collection. The template contains the user's name, incentive content, expected benefits, recommended equipment adjustment time period and savings amount. The generated incentive push messages are processed by natural language understanding, 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 collection 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 data set;

[0077] A user portrait generation module is used to input the multi-source user data set into the user portrait model to generate a deep user portrait including a basic attribute layer, a behavioral feature layer, and a preference recognition layer;

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

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

[0080] An incentive push module, configured to generate an incentive push message using a natural language generation method, and actively push the incentive push message to a target user terminal within the optimal push time period;

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

[0082] The beneficial effects of the present invention include: fusing and preprocessing multi-source heterogeneous data before inputting it into a user profile model based on the Transformer architecture, extracting behavioral features at multiple time scales and forming a deep user profile comprising a basic attribute layer, a behavioral feature layer, and a preference recognition layer. This enables parallel prediction of individual response probability, response amplitude, response duration, and comfort level, 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, preventing the original user data from leaving the domain while balancing privacy and model performance. Within a reinforcement learning framework, the deep user profile is combined with the grid operating status and historical response records to form a state space. Economic, environmental, and social incentives are combined to form an action space. Rewards are designed based on response rate, user satisfaction, cost-effectiveness, and long-term engagement, achieving integrated optimization of incentive type, intensity, and push timeframe, effectively improving regulatory effectiveness at a unit cost. A user behavior activity prediction model predicts future activity probability at the hourly scale, determines the optimal push timeframe, and uses natural language generation to generate context-matched messaging to improve push click-through rates and engagement conversion rates. After receiving user response results and feedback information, the system updates the user portrait and personalized incentive strategy, realizing full-link closed-loop optimization from strategy generation, reach execution to effect feedback, so that the model and strategy continue to evolve with changes in user behavior and external conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0083] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be derived from these drawings without inventive effort. Among them: Figure 1 is a flow chart of the method of the present invention; Figure 2 Schematic diagram of the modular structure of the system in an embodiment of the present invention. DETAILED DESCRIPTION

[0084] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the described embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field are within the scope of protection of the present invention.

[0085] like Figure 1As shown, an embodiment of the present invention provides a method for generating user incentives for a virtual power plant based on deep user portraits, including five stages: multi-source data collection and processing, deep user portrait modeling, personalized incentive strategy generation, predictive interaction and push, and closed-loop feedback optimization mechanism.

[0086] 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 data set.

[0087] Hardware device deployment:

[0088] Smart meters: Install smart meters supporting the DL / T 645 protocol for each household to collect electricity consumption data at the 1-minute level;

[0089] Environmental sensors: Temperature, humidity, and light sensors are installed in every household to collect indoor and outdoor environmental data;

[0090] Smart sockets: Configure smart sockets for major electrical appliances (air conditioners, water heaters, washing machines, etc.) to monitor device status;

[0091] User terminal: Develop mobile apps and web interfaces to collect user interaction data and preference feedback.

[0092] An edge computing gateway is deployed locally to implement data preprocessing and caching. The communication protocol uses the MQTT protocol for inter-device communication and the HTTP API for cloud data transmission. The AES-256 encryption algorithm is used to protect data transmission security.

[0093] User energy consumption data includes electrical parameters such as active power, reactive power, voltage, current, and power factor. The device interfaces include smart meters, smart sockets, and energy consumption monitoring terminals. The communication protocol is the DL / T 645 meter communication protocol or Modbus RTU protocol, and the sampling period is 1 minute for high-frequency collection.

[0094] Equipment status data includes the switch status, operating mode, set intensity parameters, power consumption, etc. of controllable equipment such as air conditioners, water heaters, and washing machines. Environmental perception data includes indoor and outdoor temperature and humidity, light intensity, wind speed, air quality, weather, season, holidays and other parameters. The sampling period is 5-minute scheduled collection and / or immediate collection triggered by key events.

[0095] User interaction data includes user operation logs, response records, feedback scores, preference settings, interface dwell time, etc. collected through mobile apps and web interfaces. The collection method is event-driven asynchronous collection, and the data format is JSON structured data, which supports real-time transmission.

[0096] External grid data includes real-time electricity prices, time-of-use electricity prices, demand response electricity prices, grid load, renewable energy output forecasts, grid dispatch instructions, etc. It is obtained every 5 minutes and supports automatic API pull.

[0097] In this embodiment, the specific process of fusing and preprocessing the collected multi-source heterogeneous data to form a structured multi-source user data set is as follows:

[0098] Unit unification, data type conversion, and deduplication: User energy consumption data (active power, reactive power, voltage, and current) are uniformly converted to unified units, such as kW, kVar, V, and A. External grid load information is unified to MW and converted to user dimensions based on the proportion of substation or feeder load.

[0099] Data type conversion: All time series are converted to float32 format, and device switch status is converted to Boolean type or 0 / 1 encoding.

[0100] Deduplication: For records with duplicate timestamps, the data with the latest timestamp is retained; if there is a source conflict, the data source is selected based on its credibility priority.

[0101] Spatial registration and entity mapping: Build a "user-metering point-device-substation / feeder" mapping table, and register external power grid data and meteorological station environmental data to the user according to the nearest neighbor of the substation's geographic coordinates; if the user moves or the device is replaced, remap the data using the device's unique ID and installation location to ensure consistent user dimensions across the entire data set.

[0102] Time alignment: User energy consumption data, device status data, environmental perception data, user interaction data, and external grid data are uniformly resampled to a unified time granularity (such as 15 minutes or 1 hour) to obtain a standardized time axis. Data with misaligned time is mapped to the starting point of the nearest window using a timestamp.

[0103] Temporal asynchronous interpolation processing: Linear interpolation is used for continuous numerical data (such as power, voltage, temperature and humidity); nearest neighbor interpolation is used for discrete data (such as device on / off status), and the value is obtained based on the timestamp closest to the target time point; curve fitting interpolation is used for smoothly changing data (such as electricity price curves), using cubic spline interpolation or Loess regression to fit the curve and then obtain the value at the target time point; asynchronous interaction events are encoded into corresponding time slots through timestamp vectors, and a data quality indicator is generated to record the interpolation method and confidence level (for example, a confidence level of 0.9 indicates that the interpolation is based on high-quality context);

[0104] Missing data is filled in a hierarchical manner: short-term missing data (continuous missing data ≤ 3 time steps) is filled by using the mean of historical data for the same period; long-term missing data (> 3 time steps) is filled by 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 does not participate in subsequent modeling.

[0105] Anomaly detection and correction: Identify single-dimensional outliers through statistical boundary analysis, i.e. is an outlier if and only if , is the mean, The standard deviation is obtained by combining cluster anomaly detection and time series anomaly detection to identify compound anomaly patterns (using K-Means clustering to detect samples whose behaviors deviate from the group center, and using LSTM time series prediction residuals to test time dimension anomalies). The detected anomalies are replaced by the mean or median of the neighboring windows, and the anomaly flag is retained for reference by the model attention mechanism.

[0106] Feature normalization and encoding: Continuous numeric features are normalized using Z-score or min-max normalization. Categorical features (such as switch state and mode type) are encoded using one-hot encoding or embedded index encoding. A missingness mask (1 = valid, 0 = padded) is generated that matches the feature dimension for model input.

[0107] Semantic parsing and vectorization: Map device status data into fixed-dimensional device status encoding vectors. Convert user interaction data (operation logs, feedback text) into text vectors using word segmentation and word embedding models (such as Word2Vec or BERT). Multiple preference labels are represented using multi-hot encoding.

[0108] An example of a device status encoding vector is as follows:

[0109] {

[0110] "Air Conditioning_Cooling": {"mode": 1, "intensity": [1,2,3,4,5]},

[0111] "Air Conditioning_Heating": {"mode": 2, "intensity": [1,2,3,4,5]},

[0112] "Air Conditioner_Off": {"mode": 0, "intensity": 0},

[0113] "waterheater_heating": {"mode": 1, "temperature": [40,45,50,55,60]},

[0114] "Water Heater_Insulation": {"mode": 2, "temperature": "current"},

[0115] "Energy Storage_Charging": {"mode": 1, "power": "variable"},

[0116] "Storage_Discharge": {"mode": 2, "power": "variable"},

[0117] "Energy Storage_Standby": {"mode": 0, "power": 0}

[0118] }

[0119] An example of encoding user preferences is as follows:

[0120] Comfort preference: One-hot coding (comfort priority / balance / energy saving priority);

[0121] Response time preference: numerical coding (immediately / within 30 minutes / within 1 hour / flexible);

[0122] Incentive type preference: multi-label coding (economic / environmental / social / convenience).

[0123] Based on the user's historical response records, supervisory labels such as response probability, amplitude, duration and comfort impact are constructed, and strict alignment with multi-source user datasets is ensured; sensitive data such as user ID and geographic information are hashed and desensitized to generate privacy metadata to ensure that subsequent federated learning complies with privacy constraints. All feature vectors are stacked by time step, and based on the alignment of user ID and timestamp, all processed multi-source heterogeneous data are unified and fused to finally generate a structured multi-source user dataset with timestamp.

[0124] S2: Input multi-source user datasets into the user portrait model to generate a deep user portrait that includes a basic attribute layer, a behavioral feature layer, and a preference recognition layer. The user portrait model is based on the Transformer architecture and is constructed through local training and parameter aggregation under the federated learning framework.

[0125] In this embodiment, multi-source user datasets are input into the user profile model to generate a deep user profile, specifically including:

[0126] The following three types of features are extracted from the multi-source user dataset:

[0127] Basic user characteristics: total number of user devices, distribution of device types, regional and substation identification, and user attribute information (e.g., residential / commercial / industrial electricity usage category);

[0128] User behavior characteristics: historical electricity consumption time series curve, historical demand response participation times, number of device operation mode changes, and interaction behavior frequency;

[0129] User preference characteristics: historical feedback ratings, comfort adjustment behavior, and response tendencies to different incentive types.

[0130] The above three types of features are processed uniformly through the feature mapping layer, mapping the different feature dimensions into a numerical vector of uniform dimension. This feature mapping layer is a fully connected network (Dense Layer), and the mapping dimension can be set to 128 to ensure consistent dimensionality when subsequently input into the model.

[0131] Timestamps and periodic position codes are superimposed on a numerical vector of uniform dimension. The position code includes absolute time information (hours, days, weeks) and periodic information (weekdays / holidays, seasonal cycles). Multi-timescale input representations are constructed, including sequence inputs of short-term windows (e.g., 24 hours), medium-term windows (e.g., 7 days), and long-term windows (e.g., 30 days). The multi-timescale sequences are fed into a Transformer encoder composed 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 between user behaviors at different time steps.

[0132] Residual connections and layer normalization improve model training stability; feedforward neural networks perform nonlinear combinations of sub-features.

[0133] The Transformer encoder outputs four types of prediction results in parallel: the target user's response probability in the future scheduled time period (the response tendency probability per hour in the next 24 hours), response amplitude (expected load adjustment capacity), response duration (expected response duration), and comfort impact (the impact score of the response on user comfort).

[0134] The prediction results, along with basic user features, behavioral features, and preference features, are fed into the user preference modeling network. The user preference modeling network is a multi-layer, fully connected neural network with three hidden layers (128-64-32 dimensions), a ReLU activation function, and a dropout ratio of 0.2 to prevent overfitting. It outputs the user's 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 preference strengths for different incentive types), and optimal interaction time (regression output, predicting the time period when users are most likely to accept push notifications).

[0135] The prediction results are integrated with the output results of the user preference modeling network to generate a deep user profile that includes the basic attribute layer, behavioral feature layer, and preference cognition layer.

[0136] The basic attribute layer includes 35 parameters, including basic user information (12 items including family size, income level, age structure, occupation type, residential type, residential area, and years of residence), device configuration information (15 items including controllable device type, capacity, brand, and purchase time), and basic electricity usage characteristics (8 items including average monthly electricity consumption, peak-to-valley ratio, load curve type, and average power factor). These parameters reflect users' electricity usage capabilities, device characteristics, and basic behavior patterns.

[0137] The behavioral feature layer includes 33 parameters, including electricity usage patterns (daily electricity usage time distribution, device usage habits, load change patterns, etc.), response behavior history (9 items, including historical participation records, response success rate, response amplitude distribution), and interaction behavior characteristics (6 items, including app usage frequency, interface dwell time, feedback positivity score, setting adjustment frequency, message click rate, and function usage diversity). These parameters are used to characterize users' temporal behavior patterns, historical response patterns, and interaction habits.

[0138] The preference cognitive layer includes comfort sensitivity (4 items, namely temperature adjustment tolerance, equipment convenience requirements, comfort priority weight, and environmental adaptability index), economic sensitivity (3 items, namely price elasticity coefficient, incentive threshold, and cost-effectiveness attention), environmental awareness level (2 items, namely carbon emission reduction attention and green energy preference), and social incentive sensitivity (1 item, namely social competition tendency index), totaling 10 parameters, which are used to reflect users' response preferences driven by different motivations.

[0139] To improve model prediction accuracy and strategy generation, this implementation uses base and task weights for each parameter: base weights (0.5–1.6) reflect the parameter's importance within the overall user profile, while task weights (0.8–2.1) dynamically adjust the influence of different parameter subsets within a specific prediction task. Weight assignment uses an attention mechanism to automatically learn the relationships between parameters and adaptively adjusts based on task type (response probability, response magnitude, comfort level impact, etc.). Time decay is introduced to maintain data freshness, and standard deviation outlier detection is used to prevent anomalous weights from interfering with the model.

[0140] Table 1 Parameter classification statistics

[0141]

[0142] From the parameter classification statistics, in this embodiment: categories such as device configuration information, basic electricity consumption characteristics, electricity consumption behavior patterns, response behavior history, comfort sensitivity and economic sensitivity account for a relatively high proportion in average weight and task weight, and are high-influence factors that the model focuses on; while categories such as social incentive sensitivity and environmental awareness, although fewer in number, also play a key role in the task of generating specific incentive strategies.

[0143] By combining the above parameter system and weight mechanism, this embodiment can accurately characterize user electricity usage characteristics and response motivations in multiple scenarios, providing high-precision input for subsequent personalized incentive strategy generation and push optimization.

[0144] The user portrait model of 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:

[0145] Receive and load the global model parameters of the user portrait model distributed by the federated aggregation server;

[0146] The user portrait model is locally trained based on a local multi-source user dataset, and iterative updates are performed according to the set learning rate decrement strategy (the initial learning rate is, for example, 0.001, which decays by 5% per round of training) and a fixed batch size (such as 32 samples / batch). For user samples whose response times in the user interaction data are less than the set threshold, the training weight is set to twice the standard weight, thereby improving the contribution of users with low response frequency in model training.

[0147] 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 each parameter gradient according to the set privacy budget upper limit, and the change rate of the model parameter gradient is calculated. Only model parameters with a change rate exceeding 5% are selected and uploaded to the federal aggregation server to reduce invalid update data transmission.

[0148] Perform the following steps on the federated aggregation server:

[0149] Based on a federated optimization algorithm (such as FedProx) that includes a regularization term, model parameter gradients uploaded by multiple user terminals are weighted and aggregated. The weighting coefficient is calculated based on the proportion of each user's local training samples to ensure that users with large sample sizes contribute more to global model updates. During the parameter aggregation process, a set standard deviation outlier detection method is used to identify and remove abnormal parameter gradients in the aggregation results that deviate from the mean by more than three standard deviations, preventing erroneous updates uploaded by a single terminal from contaminating the global model.

[0150] 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 a set number of iterative training rounds (such as 20 rounds), continuous optimization is carried out without uploading the user's original data, and finally a deep user profile with privacy protection capabilities and support for updates is generated.

[0151] S3: Based on the in-depth user profile, combined with the current grid operation status and historical response records, a reinforcement learning algorithm is used 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 period.

[0152] In one embodiment of the present invention, a reinforcement learning algorithm is used to generate a personalized incentive strategy for a target user. First, the state space, action space, and reward function of the reinforcement learning are defined.

[0153] In this embodiment, the state space includes deep user portraits (the dimension is set to 78, including user economic sensitivity, comfort preference weight, incentive type preference intensity, etc.), current grid status (including system load level, renewable energy output, electricity price level, etc.), historical incentive effects (including previous The response rate, feedback score, cost-effectiveness ratio, etc. of the incentive strategy) and environmental perception data (temporal influencing factors such as weather, season, holidays, etc.);

[0154] The action space includes incentive type combinations (select 1-3 combinations from the economic, environmental, and social incentive libraries), incentive reinforcement parameters (such as price discount rate, number of carbon credits, ranking bonus value, etc.), and push time selection (accurate to the hour level);

[0155] Reward Function By response rate , user satisfaction , cost-effectiveness ratio and long-term engagement The weighted calculation is: ;

[0156] Where, is the response rate weight, which is used to measure the immediate effect of the incentive strategy. ; is the user satisfaction weight, based on user feedback ratings, ; is the cost-benefit ratio weight, that is, the ratio of incentive cost to response value, ; is the long-term engagement weight, which is used to measure the stability of users’ continued participation. .

[0157] Before the reinforcement learning algorithm selects an action, the action space priority is set based on the preference recognition layer of the deep user profile:

[0158] When the economic sensitivity is greater than a threshold, such as 0.6, economic incentives are prioritized and pricing is dynamically calculated;

[0159] When the environmental incentive preference value in the incentive type preference intensity is greater than a threshold value such as 0.5, the environmental incentive is added;

[0160] When the social incentive preference value is greater than a threshold value such as 0.4, social incentive is added.

[0161] Financial incentives include:

[0162] Direct subsidy: linear / tiered subsidy based on response volume (yuan / kWh);

[0163] Electricity price discount: personalized electricity prices based on user price sensitivity (discount rate 0-20%);

[0164] Cashback Rewards: A cumulative cashback mechanism based on consecutive days of participation.

[0165] Non-financial incentives include:

[0166] Carbon credit rewards: Carbon credits are issued based on emission reduction contributions (credits / kgCO2);

[0167] Green certification: virtual honorary titles such as Environmental Protection Expert and Energy Saving Pioneer;

[0168] Priority benefits: Enjoy priority access to new features, exclusive customer service, and other privileged services.

[0169] Social incentives include:

[0170] Ranking competition: community energy conservation ranking, response initiative ranking;

[0171] Team Challenge: Energy-saving competitions among families and communities;

[0172] Sharing incentives: Additional rewards for sharing responses on social media.

[0173] During the reinforcement learning training process, 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 the incentive combination optimization algorithm to determine the selected incentive type combination and incentive reinforcement parameters;

[0174] Apply the selected incentive type combination and incentive reinforcement parameters to the target user, and generate an immediate reward value based on the target user's response rate, feedback score, and grid load regulation effect within the preset response period, and update the policy network in the reinforcement learning algorithm;

[0175] The reinforcement learning algorithm uses the Proximal Policy Optimization (PPO) method to maximize the cumulative discounted return:

[0176] ;

[0177] Where, is the reinforcement learning objective function, i.e. the expected cumulative discounted return value; is the policy network parameter; is the expectation operator, which represents the expected value of all possible state transitions and action selection paths; is the discount factor;

[0178] PPO ensures that the policy update amplitude is controlled and the training is stable. After the policy network converges, it outputs a personalized incentive policy for the target user.

[0179] Among them, the methods for dynamically calculating pricing include:

[0180] According to target users Calculate the user price elasticity index based on the economic sensitivity of the user, historical price response data and current economic conditions ;

[0181] Calculate personalized discount rate based on user price elasticity index , the formula is:

[0182] ;

[0183] Where, It is an adjustable factor with a default value of 0.15; It is a market adjustment factor that reflects the current supply and demand situation;

[0184] Personalized discount rate The upper limit is 20% to prevent the risk of cross-subsidy;

[0185] Dynamically adjust personalized pricing based on time-sharing pricing , the formula is:

[0186] ;

[0187] Where, is the base price; It is the time-sharing pricing coefficient, which can be 1.2, 1.0 or 0.6 according to the peak (8:00-11:00, 18:00-21:00), flat (6:00-8:00, 11:00-18:00, 21:00-23:00) and valley (23:00-6:00) time periods.

[0188] The incentive combination optimization algorithm specifically includes:

[0189] Calculate the comprehensive incentive effect for the selected incentive type , the formula is:

[0190] ;

[0191] Where, For users The preference weights of the incentive types; is the effectiveness coefficient of the incentive type in the current state; For the The basic utility value of the incentive type; among them,

[0192] Basic utility value of economic incentives The calculation formula is:

[0193] ;

[0194] Basic utility value of environmental protection incentives The calculation formula is:

[0195] ;

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

[0197] ;

[0198] Where, is the economic incentive amount, The user's annual or monthly income level; The number of carbon credits issued for environmental incentives, is the environmental preference weight; Ranking points or ranking bonus values ​​provided for social incentives, weighting preferences for social incentives;

[0199] The constraints of the incentive package include:

[0200] Total cost constraint: ;in, For the The actual cost of each incentive type, A budget cap on incentive costs for current demand response activities;

[0201] Effect balance constraints: ;in, For the The basic utility value of the incentive type, is the maximum allowed threshold of the difference in utility between incentive types;

[0202] User acceptance constraints: The complexity of the incentive combination must not exceed the user's acceptable cognitive load, and the maximum number of incentive types is 3;

[0203] Maximize the comprehensive incentive utility under the premise of meeting the total cost constraints, effect balance constraints and user acceptance constraints , output the incentive type combination. The final incentive combination should not exceed 3 incentive types.

[0204] S4: Use the user behavior activity prediction model to predict the target user's activity probability in the future time period, determine the optimal push time period for the personalized incentive strategy, use the natural language generation method to generate the incentive push message, and actively push the incentive push message to the target user terminal within the optimal push time period.

[0205] In this embodiment, the user behavior activity prediction model predicts the activity probability of the target user in the future time period based on the LSTM network. The LSTM network input includes the historical login time series (past 30 days), the time distribution of user operation behavior, and environmental perception data (weekdays, weekends, holidays and weather conditions), and outputs the prediction results of the user activity probability per hour in the next 24 hours; based on the prediction results, a high activity period of 2 consecutive hours is determined as the optimal push time period, and the push success rate is predicted based on historical response data.

[0206] Build 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.

[0207] Execute the push strategy based on the urgency of the personalized incentive strategy. In high-urgency scenarios, push immediately and supplement with SMS notifications. In medium-urgency scenarios, push during the predicted optimal push time period. In low-urgency scenarios, push at the time the user is accustomed to. Implement push frequency control, with no more than three pushes per day and a single push interval of no less than two hours. Monitor user fatigue and suspend push for 24 hours after three consecutive rejections.

[0208] Call the dialogue template library to generate incentive push messages. Template types include incentive invitation, personalized suggestion, and feedback collection. The template contains parameters such as user name, incentive content, expected benefits, recommended device adjustment time period, and savings amount. An example of the dialogue template library is as follows:

[0209] Incentive invitation category:

[0210] "Hi {username}, you can get {incentive content} by participating in the energy conservation response. It is estimated that you will save {expected income} yuan. Do you want to participate?"

[0211] Personalized suggestions:

[0212] "Based on your electricity usage habits, we recommend adjusting {device name} to {recommended setting} during {time period} to save {amount saved} yuan in electricity bills."

[0213] Feedback collection class:

[0214] "Thank you for participating in this response! Please rate this experience (1-5 stars). Your feedback will help us provide better service."

[0215] The generated incentive push messages are processed through natural language understanding, including intent recognition (willingness to participate, reasons for rejection, suggestions and feedback, complaints and consultations), entity extraction (key information such as time, device, amount, satisfaction, etc.) and sentiment analysis (positive, negative, and neutral sentiment tendency recognition) to enhance the personalization of the push content; support is provided for voice recognition (support for Chinese voice command recognition), voice synthesis (personalized voice broadcast, support for dialect adaptation), voice emotion recognition (recognition of the emotional state in the user's voice), graphical interface visualization push (charts showing response effects and benefits, intuitive device control and preference setting interface) and real-time feedback display (real-time status display of the response process) to achieve multimodal interaction.

[0216] S5: Receive the target user's response results and feedback information to the personalized incentive strategy, and update the user portrait model and personalized incentive strategy based on the response results and feedback information to achieve closed-loop optimization.

[0217] The method of this embodiment can be directly run on the virtual power plant platform of the residential community. In view of the special needs of commercial building users, it needs to be adapted and optimized based on the implementation in the residential community.

[0218] The key points of differentiated implementation are:

[0219] In terms of data collection, it connects to the building automation system (BAS) to obtain detailed operating data of air conditioning, lighting, elevators and other equipment; integrates the energy management platform (EMS) to achieve sub-item metering and refined monitoring; and adds commercial operation data: business hours, customer flow, indoor air quality, etc.

[0220] In terms of user portrait models, commercial operation characteristics (business type, operation scale, energy consumption intensity, etc.) can be added, the cost sensitivity and operation continuity requirements of commercial users can be considered, and the collaborative response capability assessment of equipment groups can be introduced.

[0221] The incentive strategy includes commercial incentives such as corporate green certification and certified carbon emission reduction (CER), designs response period optimization based on minimizing operational impact, and provides value-added services such as energy consumption analysis reports and energy-saving recommendations.

[0222] like Figure 2 FIG. 1 is another embodiment of the present invention, which provides a virtual power plant user incentive generation system based on deep user profiling, including:

[0223] The data collection 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 data set;

[0224] The user profile generation module is used to input multi-source user data sets into the user profile model to generate a deep user profile including the basic attribute layer, behavioral feature layer and preference recognition layer;

[0225] The personalized incentive strategy generation module is used to generate personalized incentive strategies for target users based on in-depth user profiles, combined with the current grid operating status and historical response records, using reinforcement learning algorithms;

[0226] The user behavior activity prediction module is used to predict the target user's activity probability in the future time period through the user behavior activity prediction model, and determine the optimal push time period for personalized incentive strategies;

[0227] The incentive push module is used to generate incentive push messages using a natural language generation method and actively push the incentive push messages to the target user terminal within the optimal push time period;

[0228] The feedback processing and updating module is used to receive the target user's response results and feedback information to the personalized incentive strategy, and update the user portrait model and personalized incentive strategy based on the response results and feedback information to achieve closed-loop optimization.

[0229] In summary, the present invention achieves accurate characterization and differentiated incentives for user-side response behaviors through a unified framework of "multi-source data fusion preprocessing - Transformer deep user profiling - federated learning privacy protection - reinforcement learning personalized incentives - active prediction optimal push - feedback closed-loop optimization", taking into account both privacy compliance and cost-effectiveness, while improving participation rate and response quality, enhancing the adjustable capacity and scheduling reliability of virtual power plants, and forming an efficient user incentive generation and interaction mechanism that is scalable for multiple scenarios.

[0230] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any person skilled in the art can easily conceive of various modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should be included within the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A virtual power plant user incentive generation method based on deep user profiling, characterized in that: The method comprises: Collect user energy consumption data, equipment status data, environmental perception data, user interaction data, and external power grid data from 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 data set; Input the multi-source user dataset into the user profile model to generate a deep user profile consisting of a basic attribute layer, a behavioral feature layer, and a preference recognition 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 in-depth user profile, combined with the current grid operating 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; Predicting the target user's activity probability in the future time period through a user behavior activity prediction model, determining the optimal push time period for the personalized incentive strategy, generating an incentive push message using a natural language generation method, and actively pushing the incentive push message to the target user terminal within the optimal push time period; Receive the target user's response results and feedback information to the personalized incentive strategy, and update the user portrait model and personalized incentive strategy based on the response results and feedback information to achieve closed-loop optimization.

2. The method for generating user incentives for a virtual power plant 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 switch status, operation mode and strength parameters of the user device; The environmental perception 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 scores 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 user incentives for a virtual power plant based on deep user profiling according to claim 1, characterized in that: The fusion preprocessing of the collected multi-source heterogeneous data to form a structured multi-source user data set includes: Perform unit unification, data type conversion, and deduplication operations on the user energy consumption data, device status data, environmental perception data, user interaction data, and external power grid data to ensure dimensional consistency and eliminate duplicate records or identification conflicts; Based on the primary key relationship between users and devices, the ownership relationship of substations or feeders, and geographic coordinate information, perform spatial registration and entity mapping of external power grid data and environmental perception data; Aligning the user energy usage data, device status data, environmental perception data, user interaction data, and external power grid data to a preset identical timestamp reference; Sliding window interpolation is used for time-asynchronous data, including linear interpolation for continuous numerical data, nearest neighbor interpolation for discrete data, and curve fitting interpolation for smoothly varying data. Asynchronous interaction events are encoded into corresponding time slots via timestamp vectors, and data quality indicators are generated to record the interpolation method and confidence level. Data missing data is filled in a hierarchical manner: short-term missing data is filled using the mean of historical data for the same period, long-term missing data is filled using collaborative filtering 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 anomaly patterns are identified by combining cluster anomaly detection and time series anomaly detection. Detected outliers are corrected by interpolation of neighboring feature means. Map device state data into device state encoding vectors, perform semantic analysis on user interaction data, and convert it into encoding vectors that can be used for model training; Based on the alignment of user identification and timestamp, all processed multi-source heterogeneous data are unified and fused to finally generate a structured multi-source user dataset with timestamp.

4. The method for generating user incentives for a virtual power plant based on deep user profiling according to claim 3 is characterized in that: Inputting the multi-source user dataset into the user profile model to generate a deep user profile specifically includes: Extracting user basic features, user behavior features, and user preference features from the multi-source user dataset, and performing dimension-unified processing on each of them through a feature mapping layer to adapt to the model input format; By superimposing timestamps and periodic position codes on a numerical vector of uniform dimension, a multi-timescale input representation is constructed. This input is fed into a Transformer encoder consisting of a multi-head self-attention mechanism, residual connections, layer normalization, and a feedforward network. This encoder then concurrently predicts the target user's response probability, response magnitude, response duration, and comfort impact within a predetermined time period in the future. The prediction results, together 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 that outputs the user's economic sensitivity score, comfort preference weight, incentive type preference intensity, and optimal interaction time, thereby generating an in-depth user profile.

5. The method for generating user incentives for a virtual power plant based on deep user profiling according to claim 4 is characterized in that: The user profile model is constructed through local training and parameter aggregation under the federated learning framework, including: Perform the following steps on the user's local terminal: Receive and load global model parameters of the user portrait model; Locally training the user profile model based on a multi-source user dataset, performing iterative updates according to a set learning rate decrement strategy and a fixed batch size during the local training process, and setting a training weight of twice the standard weight for user samples whose response times in the user interaction data are less than a set threshold; After local training is completed, privacy protection is performed on the locally updated model parameter gradients. The privacy protection process 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% based on a preset parameter change rate threshold and uploading them to the federated aggregation server; Perform the following steps on the federated aggregation server: Based on a federated optimization algorithm including a regularization term, the model parameter gradients uploaded by multiple user terminals are weighted and aggregated. The weighting coefficient is calculated based on the proportion of each user's total local training samples. During the parameter aggregation process, the set standard deviation outlier detection method is used to identify and eliminate abnormal parameter gradients in the aggregation results that deviate from the mean by more than 3 standard deviations; 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 portrait model is continuously optimized through a set number of iterative training rounds without uploading the user's original data, and ultimately generates a deep user portrait with privacy protection capabilities and support for updates.

6. The method for generating user incentives for a virtual power plant based on deep user profiling according to claim 1, characterized in that: The method of using a reinforcement learning algorithm to generate a personalized incentive strategy for target users specifically includes: The state space for reinforcement learning is defined to include deep user profiles, current grid status, historical incentive effects, and environmental perception data. The action space includes incentive type combinations, incentive reinforcement parameters, and push timing. The reward function is calculated by weighting response rate, user satisfaction, cost-effectiveness, and long-term engagement. Before the reinforcement learning algorithm selects an action, the action space priority is set based on the preference recognition layer of the deep user profile: for users whose economic sensitivity exceeds a preset threshold, economic incentives are given priority, and pricing is calculated dynamically; 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 given priority; During the reinforcement learning training process, 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 the incentive combination optimization algorithm to determine the selected incentive type combination and incentive reinforcement parameters; Apply the selected incentive type combination and incentive reinforcement parameters to the target user, and generate an immediate reward value based on the target user's response rate, feedback score, and grid load regulation effect within the preset response period, and update the policy network in the reinforcement learning algorithm; The proximal policy optimization method is used to maximize the cumulative discounted return. After the policy network converges, a personalized incentive strategy for the target user is output.

7. The method for generating user incentives for a virtual power plant based on deep user profiling according to claim 6, characterized in that: The method for dynamically calculating pricing includes: According to target users Calculate the user price elasticity index based on the economic sensitivity of the user, historical price response data and current economic conditions ; Calculate personalized discount rate based on user price elasticity index , the formula is: ； Where, is the adjustable factor, is the market adjustment factor; Personalized discount rate The upper limit is 20%; Dynamically adjust personalized pricing based on time-sharing pricing , the formula is: ； Where, is the base price; This is the time-sharing pricing coefficient, and 1.2, 1.0 or 0.6 are selected according to the peak, flat or valley periods.

8. The method for generating user incentives for a virtual power plant based on deep user profiling according to claim 6, characterized in that: The incentive combination optimization algorithm specifically includes: Calculate the comprehensive incentive effect for the selected incentive type , the formula is: ； Where, For users The preference weights of the incentive types; is the effectiveness coefficient of the incentive type in the current state; For the The basic utility value of the incentive type; among them, Basic utility value of economic incentives The calculation formula is: ； Basic utility value of environmental protection incentives The calculation formula is: ； The basic utility value of social incentives The calculation formula is: ； Where, is the economic incentive amount, The user's annual or monthly income level; The number of carbon credits issued for environmental incentives, is the environmental preference weight; Ranking points or ranking bonus values ​​provided for social incentives, weighting preferences for social incentives; The constraints of the incentive package include: Total cost constraint: ;in, For the The actual cost of each incentive type, A budget cap on incentive costs for current demand response activities; Effect balance constraints: ;in, For the The basic utility value of the incentive type, is the maximum allowed threshold of the difference in utility between incentive types; User acceptance constraints: The complexity of the incentive combination must not exceed the user's acceptable cognitive load, and the maximum number of incentive types is 3; Maximize the comprehensive incentive utility under the premise of satisfying the total cost constraint, effect balance constraint and user acceptance constraint , output excitation type combination.

9. The method for generating user incentives for a virtual power plant based on deep user profiling according to claim 1, characterized in that: The user behavior activity prediction model predicts the target user's activity probability in the future time period based on the 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's activity probability per hour in the next 24 hours; Based on the prediction results, a two-hour period of high activity is determined as the optimal push time period, and the push success rate is predicted based on historical response data. Build 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. Execute push strategies based on the urgency of personalized incentive policies. In high-urgency scenarios, push notifications are sent immediately and supplemented with SMS notifications. In medium-urgency scenarios, push notifications are sent during the predicted optimal time period. In low-urgency scenarios, push notifications are sent at times that users are accustomed to. Implement push frequency control, with no more than three pushes per day, and a minimum of two hours between pushes. Monitor user fatigue and suspend pushes for 24 hours after three consecutive rejections. Call the dialogue template library to generate incentive push messages. Template types include incentive invitation, personalized suggestion and feedback collection. The template contains the user's name, incentive content, expected benefits, recommended equipment adjustment time period and savings amount. The generated incentive push messages are processed by natural language understanding, 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.

10. A virtual power plant user incentive generation system based on deep user profiling, applied to a virtual power plant user incentive generation method based on deep user profiling according to any one of claims 1 to 9, characterized in that: The system comprises: The data collection 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 data set; A user portrait generation module is used to input the multi-source user data set into the user portrait model to generate a deep user portrait including a basic attribute layer, a behavioral feature layer, and a preference recognition layer; A personalized incentive strategy generation module is used to generate a personalized incentive strategy for the target user using a reinforcement learning algorithm based on the deep user profile, combined with the current grid operation status and historical response records; A user behavior activity prediction module is used to predict the activity probability of the target user in the future time period through the user behavior activity prediction model, and determine the optimal push time period for the personalized incentive strategy; An incentive push module, configured to generate an incentive push message using a natural language generation method, and actively push the incentive push message to a target user terminal within the optimal push time period; The feedback processing and updating module is used to receive the target user's response results and feedback information to the personalized incentive strategy, and update the user portrait model and personalized incentive strategy based on the response results and feedback information to achieve closed-loop optimization.

Images