System for constructing high-precision user behavior portrait through improved federal learning algorithm

Abstract

The technical scheme of the invention discloses a system for constructing a high-precision user behavior portrait through an improved federal learning algorithm, and the system is characterized in that the system comprises a multi-source privacy protection data obtaining module; a federated learning modeling module based on dynamic weighting and contribution proving; a multi-dimensional user behavior portrait module; a personalized vehicle network interaction strategy generation module; a dynamic execution and fair excitation module; and a feedback and model self-optimization module. According to the method, deep analysis and portrait construction can be carried out on historical data of user charging and discharging behavior habits, a high-precision and high-coverage user behavior portrait model is generated through an improved federated machine learning algorithm, the real-time income of users is improved, the user viscosity is enhanced, and the power grid stability and the renewable energy consumption rate are further improved.

Description

Technical Field

[0001] This invention relates to a system for constructing high-precision user behavior profiles using an improved federated learning algorithm, belonging to the application technology field of vehicle-to-grid interaction charging and discharging guidance for new energy vehicles. Background Technology

[0002] With the accelerated low-carbon transformation of the energy system and the large-scale popularization of electric vehicles, vehicle-to-grid (V2G) applications have emerged to address issues such as widening peak-valley differences in the power grid and insufficient renewable energy absorption. These applications enable bidirectional energy flow between vehicles and the power grid. To effectively smooth grid load, improve the utilization rate of charging facilities and equipment, and ensure the safe and stable operation of the power system, the business models of V2G technology are constantly developing, and corresponding policies and pricing mechanisms are becoming increasingly mature. This expands the application scenarios for user participation in V2G and promotes user engagement.

[0003] Chinese invention patent application CN119944606A, published on May 6, 2025, discloses a method and system for assessing the reliability of a regional power distribution network that considers vehicle-to-grid interaction. It constructs a vehicle-to-grid interaction model by simulating behavioral data through data collection and processing. However, this patent application does not consider users' charging habits, travel patterns, or other biased data, and does not predict the charging and discharging time of vehicle-to-grid interaction.

[0004] Chinese invention patent application CN119906067A, published on April 29, 2025, discloses a vehicle-to-grid interaction control method and system based on charging pile data. It calculates the vehicle-to-grid interaction suitability index for each charging pile by collecting power data of each controllable vehicle, power load data of the distribution network, and charging power data of each charging pile. However, this invention patent application fails to promote user participation in vehicle-to-grid interaction or improve the utilization rate of vehicle charging facilities and equipment.

[0005] Chinese invention patent application CN119813304A, published on April 11, 2025, discloses an intelligent charging and discharging control method and system for electric vehicles based on vehicle-to-grid (V2G) interaction. This method acquires information about charging piles in the area to be regulated, constructs a user charging behavior analysis model based on battery information, assesses the responsiveness of V2G resources, and formulates an orderly charging control scheme. However, the predictive model of user charging behavior constructed based on battery information cannot distinguish between battery consumption as discharge or travel-related energy expenditure, leading to significant errors in predicting user participation in V2G interaction. Summary of the Invention

[0006] The purpose of this invention is to provide a new method for guiding vehicle-to-grid (V2G) interactive charging and discharging, so as to enable users to actively participate in V2G interactive charging and discharging and dynamically adapt to user needs and grid dispatching objectives.

[0007] To achieve the above objectives, the technical solution of this invention discloses a system for constructing high-precision user behavior profiles using an improved federated learning algorithm, characterized by comprising: Multi-source privacy-preserving data acquisition module: used to acquire raw data required for model training and policy formulation in a distributed manner while protecting user privacy. All raw data involving user personal privacy and vehicle privacy are used only for local feature extraction and model training. The federated learning modeling module based on dynamic weighting and contribution proof is used to collaboratively train a high-precision, high-coverage global behavior profile model using distributed local data obtained through the multi-source privacy-preserving data acquisition module, while ensuring data privacy and fair incentives. Multi-dimensional user behavior profiling module: used to accurately classify and extract features of user behavior, providing a basis for personalized guidance strategies. Based on structured vectors containing feature values, it accurately quantifies user behavior and outputs the prediction results of the profile through the generated high-quality global behavior profile model. Personalized vehicle-to-everything (V2X) interaction strategy generation module: This module transforms the output of the multi-dimensional user behavior profile module into an executable, highly personalized, and precise charging and discharging guidance strategy, enabling multi-objective dynamic optimization. Dynamic Execution and Fair Incentive Module: This module is used to accurately execute guidance strategies while ensuring grid security and core user interests. Based on a fair incentive algorithm, it quantifies and rewards users' long-term, multi-dimensional contributions, thereby building a sustainable vehicle-grid interaction ecosystem.

[0008] Preferably, the implementation of dynamic weighting includes the following steps: The metadata obtained through the multi-source privacy-preserving data acquisition module is designed in a refined manner and evaluated by data volume, data quality score, local training loss reduction value, and data distribution difference. A modified linear unit and a threshold are introduced. When the data distribution difference is less than the threshold, no labeling is performed. Only when the data distribution difference is greater than the threshold is a suppression term obtained. Based on the suppression term, the modified linear unit is used to calculate dynamic weights to suppress abnormal clients that may introduce noise. The final weights are obtained by controlling the smoothness of the weight distribution through function normalization. The obtained final weights are used for dynamic weighted aggregation during the training of the global behavior profile model.

[0009] Preferably, a federated learning mechanism is adopted to enhance contribution proof, optimizing the system's sustainability and fairness, including the following steps: Quantifying contribution: The server maintains a global validation set, and the data distribution is as even as possible to represent all user types. After each round of aggregation, the performance improvement of the new global model is calculated. A gradient-based approximation algorithm is used to reduce computational overhead, so that the model does not need to be retrained. The contribution is approximated by analyzing the marginal gain of performance during the aggregation process. Update and use long-term contribution profiles: Emphasize recent contributions through a decay factor and output long-term contribution profiles for each client; Long-term contribution profiles provide a positive feedback loop for dynamic weighting: long-term contribution incentivizes users to provide higher quality and more stable data, thereby improving the data source quality in the dynamic weighting aggregation process, forming a closed loop of improved data quality -> more accurate model -> fairer contribution assessment -> stronger incentives -> further improvement in data quality.

[0010] Preferably, the charging and discharging guidance strategy is generated by a dynamic strategy generation engine based on profile awareness, which combines profile-driven dynamic reconstruction of the objective function, a rolling optimization framework, and interpretable output.

[0011] Preferably, it also includes a feedback and model self-optimization module: used to continuously collect policy execution results and perform closed-loop optimization, including model updates, policy tuning, and profile iteration.

[0012] Preferably, the raw data includes: Battery status information, real-time location information, and trip plan information of the vehicle terminal; Charging and discharging power information, electricity price information, and charging pile status information measured at the edge of the charging pile; Real-time load information, renewable energy output information, and nodal price information on the power grid side; User's charging and discharging preference settings and historical records.

[0013] Preferably, the client calculates the difference between the divergence and the maximum mean of the local data feature distribution and the output distribution of the previous round of global model on the common anchor dataset, so as to reflect the difference in the data distribution.

[0014] Preferably, the features extracted by the multi-dimensional user behavior profiling module include daily average charging time distribution C1, average single charging amount C2, preferred common charging power C3, charging location regularity C4, daily average driving mileage T1, commuting regularity T2, long-distance travel frequency T3, trip predictability T4, historical electricity price response rate R1, historical incentive response rate R2, average response time R3, subjective preference settings R4, battery health V1, and average energy consumption V2.

[0015] Preferably, the multi-dimensional user behavior profiling module learns from the extracted features through cluster analysis, defining and identifying six typical user groups as shown in the table below: .

[0016] Preferably, in the dynamic strategy generation engine, a portrait-driven dynamic multi-objective optimization framework is established, including the following steps: For economically rational users, the profile is automatically configured with adaptive weights, and the objective function is approximately to maximize user benefits. For users who contribute to environmental protection, the grid load demand item is refined into green electricity consumption matching degree, and a high-weight grid load demand is configured. The objective function is approximately to maximize the use of green electricity. For conservative users, a high weight is assigned to the battery stress accumulation model, with the objective function focusing on minimizing battery loss. It also allows for personalized settings of constraints based on different user types.

[0017] Preferably, in the dynamic policy generation engine, rolling optimization based on model predictive control is performed. The model predictive control framework is used to perform rolling optimization within a finite time domain, including the following steps: Based on user travel plans and historical behavior baselines, predict the load baseline and schedulable time windows for the next few hours; at the current moment, based on the predicted information and real-time status, solve the personalized objective function to obtain the optimal charging and discharging power sequence for the future period; execute only the first control command at the current moment, and at the next moment, re-predict and optimize based on the latest system status, rolling forward.

[0018] Preferably, the dynamic execution and fair incentive module adopts the following steps: Each user maintains a dynamic response reputation score: based on historical response records such as the number of successful responses and the total number of invitations received, a multi-dimensional comprehensive incentive formula is formed to comprehensively measure user value, as shown in the following formula:

[0019] In the formula, This represents the total incentive reward ultimately given to user i. This refers to the total discharge amount that user i actually feeds back to the grid in a single vehicle-to-grid interaction event, or the equivalent adjusted amount of electricity generated by adjusting charging behavior according to guidance during a specific time period. For real-time node electricity prices, To assess long-term data contribution, a contribution proof mechanism is used, with a range of 0 to 1, to quantify the marginal contribution of user i to improving the performance of the global model. For the response quality coefficient, the multiplier factor is greater than or equal to 1, by and Joint decision, For user i, a real-time response reputation score. To ensure the timeliness of this response for user i, The cost of battery degradation resulting from this charge-discharge cycle is represented by β, where β is the amplification factor. The compensation coefficient; The state-based real-time protection mechanism monitors the user's vehicle SOC, battery temperature, and the minimum guaranteed charge set by the user in real time. During the execution of the discharge command, if the real-time SOC drops to the safety buffer margin or the battery temperature is abnormal, the user is automatically and silently removed from the current discharge queue and switched to charging or silent mode.

[0020] This invention can perform in-depth analysis and profile construction of historical data on users' charging and discharging behavior habits. By improving the federated machine learning algorithm, it generates a high-precision and high-coverage user behavior profile model, improves users' real-time benefits, enhances user stickiness, and thus further improves grid stability and renewable energy consumption rate.

[0021] Compared with existing technical solutions, the present invention has the following beneficial effects: 1) By introducing a dynamic weighted aggregation mechanism, the problem of extremely heterogeneous data in vehicle-to-everything (V2X) interaction scenarios is effectively solved. This mechanism comprehensively evaluates the data quality, training effect, and distribution differences of the client, and intelligently adjusts its weight in model aggregation, thereby significantly improving the accuracy and group coverage of user behavior profiles. This enables the model to not only accurately depict mainstream users, but also to keenly identify and learn niche but valuable charging and discharging behavior patterns.

[0022] 2) It innovatively integrates a contribution proof mechanism into the federated learning framework, achieving a closed-loop linkage between data value and model performance. This mechanism can quantitatively evaluate the marginal contribution of each user's data to the global model optimization and record it in a long-term archive. This not only provides a fair and transparent basis for subsequent economic incentives, but also fundamentally builds a virtuous cycle that continuously incentivizes users to contribute high-quality data, ensuring the long-term evolutionary capability and system vitality of the profiling model.

[0023] 3) Based on the improved algorithm described above, the constructed user behavior profile combines global consensus with individual specificity. Thanks to the high-quality global model, the profile feature extraction is more accurate. At the same time, the dynamic weighting mechanism ensures that users with different behavioral patterns can be fully represented in the profile system, thus laying a solid and reliable cognitive foundation for generating highly personalized vehicle-to-everything (V2X) interaction strategies and achieving a leap from "group profiling" to "precise individual characterization".

[0024] 4) This improved algorithm enhances model performance while maintaining computational and communication efficiency. By employing a client selection strategy, high-value clients are prioritized for training, and lightweight metadata is uploaded to replace some parameters. This effectively reduces communication overhead and computational resource consumption in a single training round, accelerating model convergence and enabling efficient and low-cost maintenance of high-precision profiles across a large user base. Attached Figure Description

[0025] Figure 1 This is a flowchart illustrating an embodiment of the present invention. Detailed Implementation

[0026] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined by the appended claims.

[0027] This invention discloses a system for constructing high-precision user behavior profiles using an improved federated learning algorithm, comprising the following modules: Multi-source privacy-preserving data acquisition module: Under the premise of protecting user privacy, it acquires the raw data required for model training and policy formulation in a distributed manner.

[0028] In this embodiment of the invention, the required raw data includes: 1) Battery status information (e.g., SOC, SOH), real-time location information, and trip plan information of the vehicle terminal; 2) Charging and discharging power information, electricity price information, and status information of the charging pile measured at its edge; 3) Real-time load information, renewable energy output information, and nodal price information on the grid side; 4) User-side charging and discharging preference settings, historical records, etc.

[0029] Of the raw data mentioned above, all data involving user privacy and vehicle privacy are used only for local feature extraction and model training, thus eliminating the risk of privacy leaks.

[0030] This federated learning modeling module, based on dynamic weighting and contribution proof, collaboratively trains a high-precision, high-coverage global behavioral profiling model using locally distributed data, while ensuring data privacy and fair incentives. Unlike traditional federated learning algorithms that assume independent and identically distributed data and use weighted averaging, this module addresses the issue of user behavior data exhibiting extreme non-independent and identically distributed characteristics in vehicle-to-everything (V2X) scenarios. Simple averaging leads to slow global model convergence, low accuracy, and an inability to cover niche but valuable groups. This invention improves upon the traditional federated averaging algorithm by using dynamic weighted aggregation.

[0031] First, this invention features a refined design for metadata, evaluating it through data volume, data quality score, local training loss reduction, and data distribution dissimilarity. In particular, the data distribution dissimilarity involves the client calculating the divergence and maximum mean difference between the local data feature distribution and the output distribution of the previous global model on the common anchor dataset. High values ​​for divergence and maximum mean difference indicate a significant difference between the client's data pattern and the global mainstream pattern.

[0032] Secondly, this invention performs dynamic weight calculation. A modified linear unit and a threshold are introduced. When the distribution difference is less than the threshold, no labeling is performed; only when the difference is too large is a suppression term obtained. This ensures that beneficial, complementary clients with slight differences are rewarded, while anomalous clients that may introduce noise are suppressed, thus achieving management of data heterogeneity.

[0033] Finally, the final weights of this invention are controlled by function normalization to control the smoothness of the weight distribution.

[0034] Unlike federated learning, which lacks an effective incentive mechanism and motivates users to contribute high-quality data, traditional incentive methods based on data volume or response power consumption fail to fairly reflect the true contribution of different data to model quality. Therefore, this invention provides an improved federated learning mechanism that adds a contribution proof mechanism, optimizing system sustainability and fairness. First, the contribution is quantified. The server needs to maintain a global validation set, with data distributed as evenly as possible to represent all user types. After each round of aggregation, the performance improvement of the new global model is calculated. A gradient-based approximation algorithm is used to reduce computational overhead, thus eliminating the need to retrain the model. The contribution is approximated by analyzing the marginal gain of performance during the aggregation process, which is more efficient.

[0035] Secondly, the long-term contribution profile is updated and utilized. By using a decay factor to emphasize recent contributions, a long-term contribution profile is generated for each client.

[0036] Finally, the long-term contribution profile provides a positive feedback loop for dynamic weighting. Long-term contributions incentivize users to provide higher-quality and more stable data, thereby improving the data source quality in the dynamic weighting aggregation process, forming a closed loop of improved data quality -> more accurate model -> fairer contribution assessment -> stronger incentives -> further improvement in data quality.

[0037] Multi-dimensional User Behavior Profiling Module: Used for accurate classification and feature extraction of user behavior, providing a basis for personalized guidance strategies. The following feature vectors are extracted from local data: 1) Charging Features: C1 – Daily average charging time distribution, i.e., charging probability during peak, off-peak, and valley periods; C2 – Average single charge amount; C3 – Preference for commonly used charging power (slow charging / fast charging); C4 – Regularity of charging locations, such as home charging, workplace, and public charging stations; 2) Travel Features: T1 – Daily average mileage; T2 – Commuting regularity; T3 – Frequency of long-distance travel; T4 – Trip predictability; 3) Response and Willingness Features: R1 – Historical electricity price response rate; R2 – Historical incentive response rate (e.g., response to additional subsidies); R3 – Average response time; R4 – Subjective preference settings, such as minimum guaranteed SOC, maximum acceptable depth of discharge, etc.; 4) Vehicle Status Features: V1 – Battery health; V2 – Average energy consumption.

[0038] The above-mentioned label definitions are used to divide the user groups. Specifically, cluster analysis can be used to learn the feature vectors mentioned above to identify six typical user groups, as shown in the table below.

[0039]

[0040] Based on a structured vector containing all the aforementioned feature values, user behavior is precisely quantified, and predictive user profiles are output, such as "probability distribution of charging demand in the next 24 hours" and "response probability of participating in V2G discharge." Unlike simply using a single feature, this approach combines features from three dimensions—behavioral habits (C / T), subjective intentions (R), and objective states (V)—to create a more comprehensive and accurate user profile. The accuracy of the profile directly depends on the high-quality global model generated by the improved federated learning. Through a dynamic weighting mechanism, it can better learn the patterns of minority groups, thus ensuring high profile coverage and avoiding the neglect of "niche" users by traditional methods.

[0041] The personalized vehicle-to-grid interaction strategy generation module transforms the output of the preceding modules into executable, highly personalized, and precise charging and discharging guidance strategies, achieving multi-objective dynamic optimization of the power grid, users, and battery life. Traditional optimization models typically use fixed weights, which cannot adapt to the diverse and dynamic needs of different users. For example, a purely economic model may not guide environmentally conscious users, while a model with strong grid scheduling may deter conservative users. Therefore, this invention aims to create a profile-aware dynamic strategy generation engine. Through profile-driven dynamic reconstruction of the objective function, a rolling optimization framework, and interpretable output, the combination of these three elements ensures that the guidance strategy is not only technically optimal but also optimal in terms of user experience and commercial feasibility. This can serve as a key to promoting the large-scale application of vehicle-to-grid interaction. First, a user profile-driven dynamic multi-objective optimization framework is established. Based on the user profile, the system dynamically reconstructs the model. For example, for economically rational users, the system automatically configures adaptive weights for the profile, with the objective function approximating maximizing user benefits. For environmentally conscious users, the system refines the grid load demand item into green electricity consumption matching degree and configures a high-weight grid load demand, with the objective function approximating maximizing the use of green electricity. For conservative users, the system configures a high weight for the battery stress accumulation model, with the objective function focusing on minimizing battery loss. Constraints can be personalized, such as setting a SOC safety boundary, setting an SOC operation window of 40%~90% for conservative users, and a more aggressive window of 20%~95% for high-value potential users; and establishing response time constraints, such as setting a shorter response confirmation time window for socially responsible users.

[0042] Secondly, rolling optimization based on model predictive control is performed. Static optimization cannot cope with real-time changes, such as sudden changes in travel plans and fluctuations in renewable energy output. This invention employs a model predictive control framework to perform rolling optimization within a finite time domain. Based on user travel plans and historical behavior baselines, the load baseline and dispatchable time windows for the next few hours are predicted. At the current moment, based on the predicted information and real-time status, a personalized objective function is solved to obtain the optimal charging and discharging power sequence for the future period. Only the first control command at the current moment is executed; at the next moment, prediction and optimization are re-performed based on the latest system state, rolling forward.

[0043] Finally, the interpretability of the generation and presentation strategy is crucial. Simple instructions like "discharge at 14:00" might confuse users and reduce adoption rates. The optimization results are processed by generating natural language descriptions and quantified benefit explanations. For example, the instruction might be: "We suggest you postpone your planned 19:00 charging to 23:00." The explanation would be: 1) After 23:00, the electricity price drops to 0.3 yuan / kWh, and this charging is expected to save you 8.5 yuan in electricity costs; 2) The power grid is currently experiencing peak evening load tension, and your delayed charging contributes to grid stability, earning you an additional 2 yuan incentive. The total benefit is 10.5 yuan. By directly linking the optimization results to user profile concerns, the transparency of the strategy and user acceptance are greatly improved.

[0044] Dynamic Execution and Fair Incentive Module: Under the premise of ensuring grid security and core user interests, this module precisely executes guidance strategies and, based on a fair incentive algorithm, quantifies and rewards users' long-term, multi-dimensional contributions, building a sustainable vehicle-grid interaction ecosystem. Traditional incentive schemes only compensate based on the electricity provided by users, neglecting data contribution and response quality, making it difficult to incentivize long-term user participation. A user with consistent behavior and high-quality data but occasional participation may gain more benefits than a user with random behavior and poor-quality data but frequent participation, but the latter's contribution to model training may be lower. Furthermore, it fails to distinguish between the timeliness and reliability of responses. This invention ensures system reliability and user experience through dynamic execution criteria: The first step is to maintain a dynamic response reputation score for each user. Based on historical response records such as the number of successful responses and the total number of invitations received, a multi-dimensional comprehensive incentive formula is formed to fully measure user value, as shown in the following formula:

[0045] In the formula, This represents the total incentive reward ultimately given to user i. This refers to the total discharge amount that user i actually feeds back to the grid in a single vehicle-to-grid interaction event, or the equivalent adjusted amount of electricity generated by adjusting charging behavior according to guidance during a specific time period. For real-time node electricity prices, To assess long-term data contribution, a contribution proof mechanism is used, with a range of 0 to 1, to quantify the marginal contribution of user i to improving the performance of the global model. For the response quality coefficient, the multiplier factor is greater than or equal to 1, by and Joint decision, For user i, a real-time response reputation score. To ensure the timeliness of this response for user i, This refers to the cost of battery degradation resulting from this charge / discharge cycle.

[0046] Data contribution coefficient: From the long-term contribution profile of the federated learning module. It quantifies the marginal contribution of a user's historical data to improving the performance of the global model. It is a standardized value (ranging from [0,1]). β is an amplification factor used to adjust the weight of data contributions in the total revenue.

[0047] Response Quality Coefficient: This coefficient is greater than 1. It is determined by the user's real-time response reputation score. And the timeliness of this response A joint decision.

[0048] Battery loss compensation: The cost of battery loss during this charge-discharge cycle, calculated based on the battery stress model, is determined by a coefficient. (Can be set to 1) to provide compensation, ensuring that the user's battery life is "zero loss" or "minor profit".

[0049] The second step is a state-based real-time protection mechanism. This involves real-time monitoring of the user's vehicle SOC, battery temperature, and the user-set minimum charge level. During the discharge process, if the real-time SOC drops to a safe buffer margin or the battery temperature becomes abnormal, the system will automatically and silently remove the user from the current discharge queue and switch to charging or silent mode. This significantly enhances the user's sense of security and willingness to participate.

[0050] Feedback and Model Self-Optimization Module: Drives the entire system to form a continuous self-optimizing closed loop. It continuously collects strategy execution results, including actual user response, grid-side effects, user-side effects, and battery loss data, to perform closed-loop optimization, including: 1) Model update: User response data is used as new label data and fed back to the federated learning modeling module to start a new round of model training, so that the profile and prediction model can continue to evolve and become more and more accurate; 2) Strategy optimization: Evaluate the effects on the grid side and the user side, dynamically adjust the multi-objective weights in the strategy generation module, and find the optimal solution for the system; 3) Profile iteration: A user's long-term response behavior will in turn update their own profile tags.

[0051] The resulting optimization loop is: Data Acquisition → Federated Learning (Model Optimization) → User Profiling → Policy Generation → Dynamic Execution and Incentives → Performance Evaluation and Feedback → Data Acquisition... The aforementioned technical solution uses personalized modeling based on user preset preferences and behavioral habits, and can utilize federated learning algorithms to design personalized strategies for guiding user behavior. However, this invention differs from the average aggregation model parameters used in traditional federated learning algorithms. In vehicle-to-grid interaction scenarios, user behavior data is extremely heterogeneous, with significant differences in travel mileage, charging time, discharge willingness, and sensitivity to electricity prices among different users. The simple averaging of traditional federated learning algorithms leads to slow global model convergence, decreased accuracy, and even an inability to fit the specific patterns of individual groups, resulting in coarse and unreliable user profiles. The improved algorithm provided by this invention ensures the accuracy and diversity of model optimization by adding a dynamic weighted aggregation mechanism, and establishes fair incentive feedback through a contribution proof mechanism to ensure system sustainability and data coverage. It performs in-depth analysis and profile construction on user charging and discharging behavior habits (including high-frequency charging time, frequently used charging locations, parking duration, typical charging amount, charging urgency), sensitivity to electricity price signals and incentive policies, battery health, etc.

[0052] Suppose a city experiences peak electricity consumption between 6:00 PM and 8:00 PM. From 100 electric vehicles, it is necessary to identify the most suitable vehicles for peak-shaving and formulate a strategy. Then, how would you proceed? Figure 1 As shown, the implementation of the above system includes the following steps: Step 1: Data Acquisition and Federated Learning Modeling Vehicle A local data: SOC=75%, average daily charging time: 8 pm; high frequency charging location: home; historical response record: 10 invitations received, 8 accepted due to high electricity prices.

[0053] Vehicle B local data: SOC=70%, historical response record: 5 invitations, 4 accepted under green electricity incentives; subjective preference: set "green energy priority".

[0054] A dynamic weighted federated learning model is established to train a global user profile model: Vehicle A's data is of high quality and exhibits strong regularity, with its dynamic aggregation weight calculated to be 0.12. Vehicle B's data has a unique distribution, reflecting a preference for green electricity, and its dynamic aggregation weight calculated to be 0.09. The model outputs a global user behavior profile model that accurately identifies user charging habits, price sensitivity, and green electricity preferences.

[0055] Step 2: User Behavior Profile Generation Feature projection: Input the local data of vehicles A and B into the global model and extract high-dimensional feature vectors.

[0056] Vehicle A feature vector: [Price sensitivity: 0.95, Green electricity preference: 0.35, Regularity: 0.90, Response potential: 0.88] Vehicle B feature vector: [Price sensitivity: 0.45, Green electricity preference: 0.92, Regularity: 0.70, Response potential: 0.75] Cluster analysis is performed on the feature vectors of all 100 users to generate core clusters. Based on the mean feature value of each cluster, a business label is automatically assigned. An example is shown below: Cluster 1: Economically Rational Type: Price Sensitivity > 0.8, Green Electricity Preference < 0.5; Vehicle A is classified into this cluster and labeled "Economically Rational Type".

[0057] Cluster 2: Environmental Contribution Type: Price Sensitivity < 0.6, Green Electricity Preference > 0.8; Vehicle B is classified into this cluster and labeled as "Environmental Contribution Type".

[0058] Cluster 3: Conservative and cautious type: response potential <0.5, regular dispersion.

[0059] At this point, the understanding of the user is complete, and vehicles A and B are dynamically and automatically assigned precise behavioral labels.

[0060] Step 3: Personalized Strategy Generation and Execution Based on the newly generated labels, different optimization weights are applied: For vehicle A, a high-yield weight is applied, generating the strategy: "Discharge immediately for 1.5 hours, with an estimated profit of 25 yuan." For vehicle B, a high-green-energy weight is applied, generating the strategy: "It is recommended to charge during the midday solar PV surplus to contribute to the environment and earn green points." Step 4: Incentive Allocation and Model Optimization Vehicle A completes its discharge, and its data contribution and response quality are quantified, ultimately earning a reward of 51.97 yuan. Vehicle B's response behavior is fed back into the federated learning system as new data. Vehicle B's "green electricity preference" data is then amplified in the next round of improved federated training through a dynamic weighting mechanism. The global model is optimized, leading to more accurate identification of "environmentally contributing" users in the future.

Claims

1. A system for constructing high-precision user behavior profiles using an improved federated learning algorithm, characterized in that, include: The federated learning modeling module based on dynamic weighting and contribution proof is used to collaboratively train a global behavior profile model using distributed local data obtained through a multi-source privacy-preserving data acquisition module. The multi-dimensional user behavior profile module accurately classifies and extracts features from user behavior, and outputs prediction results through the generated high-quality global behavior profile model. The personalized vehicle-to-grid interaction strategy generation module transforms the output of the multi-dimensional user behavior profile module into charging and discharging guidance strategies, achieving multi-objective dynamic optimization. The dynamic execution and fair incentive module accurately executes guidance strategies while ensuring grid security and core user interests, and quantifies long-term, multi-dimensional contributions based on a fair incentive algorithm, building a sustainable vehicle-to-grid interaction ecosystem.

2. The system for constructing high-precision user behavior profiles using an improved federated learning algorithm as described in claim 1, characterized in that, The implementation of the dynamic weighting includes the following steps: The metadata obtained through the multi-source privacy-preserving data acquisition module is finely designed and evaluated by data volume, data quality score, local training loss reduction value, and data distribution difference. A modified linear unit and a threshold are introduced. When the data distribution difference is less than the threshold, no labeling is performed. Only when the data distribution difference is greater than the threshold is a suppression term obtained. Based on the suppression term, the modified linear unit is used to calculate dynamic weights to suppress abnormal clients that may introduce noise. The final weights are obtained by controlling the smoothness of the weight distribution through function normalization. The obtained final weights are used for dynamic weighted aggregation during the training of the global behavior profile model.

3. The system for constructing high-precision user behavior profiles using an improved federated learning algorithm as described in claim 1, characterized in that, Federated learning is employed to enhance contribution proof mechanisms and optimize system sustainability and fairness. This includes the following steps: Quantifying contributions: The server maintains a global validation set, with data distributed as evenly as possible to represent all user types. After each aggregation round, the performance improvement of the new global model is calculated using a gradient-based approximation algorithm to reduce computational overhead, thus eliminating the need for model retraining. Contribution is approximated by analyzing the marginal gain of performance during aggregation. Updating and utilizing long-term contribution profiles: A decay factor is used to emphasize recent contributions, outputting a long-term contribution profile for each client. Long-term contribution profiles provide a positive feedback loop for dynamic weighting: Long-term contributions incentivize users to provide higher quality and more stable data, thereby improving the data source quality in the dynamic weighting aggregation process, forming a closed loop of improved data quality -> more accurate model -> fairer contribution evaluation -> stronger incentives -> further improvement in data quality.

4. The system for constructing high-precision user behavior profiles using an improved federated learning algorithm as described in claim 1, characterized in that, The charging and discharging guidance strategy is based on a dynamic strategy generation engine that is based on profile awareness. It combines profile-driven dynamic reconstruction of objective functions, a rolling optimization framework, and interpretable output.

5. The system for constructing high-precision user behavior profiles using an improved federated learning algorithm as described in claim 1, characterized in that, It also includes a feedback and model self-optimization module: used to continuously collect policy execution results for closed-loop optimization, including model updates, policy tuning, and profile iteration.

6. The system for constructing high-precision user behavior profiles using an improved federated learning algorithm as described in claim 1, characterized in that, The features extracted by the multi-dimensional user behavior profiling module include daily average charging time distribution C1, average single charging amount C2, preferred charging power C3, charging location regularity C4, daily average driving mileage T1, commuting regularity T2, long-distance travel frequency T3, trip predictability T4, historical electricity price response rate R1, historical incentive response rate R2, average response time R3, subjective preference settings R4, battery health V1, and average energy consumption V2.

7. A system for constructing high-precision user behavior profiles using an improved federated learning algorithm as described in claim 6, characterized in that, The multi-dimensional user behavior profiling module learns from the extracted features through cluster analysis and defines and identifies six typical user groups as shown in the table below: 。 8. A system for constructing high-precision user behavior profiles using an improved federated learning algorithm as described in claim 5, characterized in that, In the dynamic strategy generation engine, a profile-driven dynamic multi-objective optimization framework is established, including the following steps: For economically rational users, the profile adaptive weights are automatically configured, and the objective function is approximately to maximize user benefits; for environmentally contributing users, the grid load demand item is refined into green electricity consumption matching degree, and a high-weight grid load demand is configured, with the objective function approximately to maximize the use of green electricity; for conservative and cautious users, a high weight is configured for the battery stress accumulation model, and the objective function focuses on minimizing battery loss; and the constraints can be personalized based on different user types.

9. A system for constructing high-precision user behavior profiles using an improved federated learning algorithm as described in claim 8, characterized in that, In the dynamic strategy generation engine, rolling optimization based on model predictive control is performed. The model predictive control framework is used to perform rolling optimization within a finite time domain, including the following steps: predicting the load baseline and schedulable time window for the next few hours based on the user's travel plan and historical behavior baseline; at the current moment, solving the personalized objective function based on the predicted information and real-time status to obtain the optimal charging and discharging power sequence for the future period; executing only the first control command at the current moment, and then re-predicting and optimizing based on the latest system status at the next moment, rolling forward.

10. A system for constructing high-precision user behavior profiles using an improved federated learning algorithm as described in claim 9, characterized in that, The dynamic execution and fair incentive module employs the following steps: Each user maintains a dynamic response reputation score; based on historical response records such as the number of successful responses and the total number of invitations received, a multi-dimensional comprehensive incentive formula is formed to comprehensively measure user value, as shown in the following formula: In the formula, This represents the total incentive reward ultimately given to user i. This refers to the total discharge amount that user i actually feeds back to the grid in a single vehicle-to-grid interaction event, or the equivalent adjusted amount of electricity generated by adjusting charging behavior according to guidance during a specific time period. For real-time node electricity prices, Contribution to long-term data For the response quality coefficient, For user i, a real-time response reputation score. To ensure the timeliness of this response for user i, The cost of battery degradation resulting from this charge-discharge cycle is represented by β, where β is the amplification factor. The compensation coefficient is used for the real-time protection mechanism based on the state: real-time monitoring of the user's vehicle SOC, battery temperature, and the minimum guaranteed charge set by the user; during the execution of the discharge command, if the real-time SOC drops to the safety buffer margin or the battery temperature is abnormal, the user will be automatically and silently removed from the current discharge queue and switched to charging or silent state.

Images