An intelligent charging scheduling method based on multi-dimensional feature fusion and federated learning

Abstract

This invention belongs to the field of specific computer model technology and discloses an intelligent charging scheduling method based on multi-dimensional feature fusion and federated learning, including the following steps: S1, collecting dynamic data from the user side, charging facility side, and power grid side respectively, and performing standardization and spatiotemporal alignment processing in sequence; S2, generating a joint feature vector by weighted fusion after weight calculation and scenario-based adjustment through a spatiotemporally coupled multimodal attention mechanism; S3, constructing a server-side deep reinforcement learning global scheduling model and embedding a charging scenario-specific digital watermark during the initialization phase; S4, inputting the joint feature vector generated in S2 into the trained global scheduling model, and outputting a discretized charging scheduling strategy based on a multi-objective optimization reward function; S5, detecting the model parameters of each federated learning client, performing hierarchical anomaly processing and cleaning abnormal data, receiving user feedback data and converting it into training samples for online fine-tuning.

Description

Technical Field

[0001] This invention relates to the field of specific computer model technology, specifically to an intelligent charging scheduling method based on multidimensional feature fusion and federated learning. Background Technology

[0002] With the explosive growth of new energy vehicle ownership, the contradiction between large-scale charging demand and grid carrying capacity, charging facility layout and user personalized needs is becoming increasingly prominent. Existing charging scheduling technology is difficult to adapt to complex actual operation scenarios. Its core pain point is the lack of a global optimization scheduling scheme that can deeply integrate multi-source data and take into account privacy protection.

[0003] Traditional scheduling methods either rely on single-dimensional data or empirical weights, leading to biased decision-making and poor adaptability; or they use centralized model training, which requires aggregating user privacy data and operational data from various charging operators, causing serious privacy leakage risks; while decentralized training cannot achieve globally optimal scheduling results, making it difficult to achieve deep integration of multi-source data, global scheduling optimization, and data privacy protection in a coordinated manner, becoming the core bottleneck restricting the implementation of intelligent charging scheduling technology. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent charging scheduling method based on multi-dimensional feature fusion and federated learning, so as to solve the problems of insufficient data fusion and conflict between global optimization and privacy protection in the charging scheduling of new energy vehicles, and improve the level of intelligent scheduling and safety controllability in large-scale charging scenarios.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: A smart charging scheduling method based on multi-dimensional feature fusion and federated learning includes the following steps: S1. Collect dynamic data from the user side, charging facility side, and power grid side respectively, and perform standardization and spatiotemporal alignment processing in sequence. S2. Based on the data processed in S1, construct user feature data group, power grid feature data group and facility feature data group. After weight calculation and scenario-based adjustment through a spatiotemporally coupled multimodal attention mechanism, weighted fusion is used to generate a joint feature vector. S3. Construct a server-side deep reinforcement learning global scheduling model, with charging operators as federated clients, train the model through a federated training process, and embed a digital watermark specific to the charging scenario during the initialization phase. S4. Input the joint feature vector generated in S2 into the trained global scheduling model, and output a discretized charging scheduling strategy based on the multi-objective optimization reward function. S5. Detect the model parameters of each federated learning client, identify abnormal clients, perform hierarchical anomaly handling and clean the abnormal data, then receive user feedback data and convert it into training samples, and fine-tune the deep reinforcement learning global scheduling model online.

[0006] Preferably, the standardization process in step S1 includes: For numerical data, normalization is applied to map it to the [0,1] interval. The normalization formula is: ,in, For normalized data, For data before normalization, This represents the global maximum value of the data. This represents the global minimum value of the data. One-Hot encoding is used to convert categorical data into numerical values; the charging time window is converted into minutes in 24-hour format, and after conversion, it is not involved in standardization processing, but directly enters the spatiotemporal alignment process.

[0007] Preferably, the spatiotemporal alignment process in step S1 includes temporal alignment and spatial alignment. Time alignment: Extract charging request time The charging pile status, real-time grid power, real-time grid voltage, and real-time grid frequency are used to combine the timestamp of the grid load forecast data with the charging request time. Matching and calculating the charging time window with the charging request time. The time difference, all non-time dimension data are bound to the charging request time. Timestamp; Spatial alignment: Match the vehicle's GPS coordinates at the time of the charging request to the corresponding power grid topology node, and extract the charging request time of that node. Real-time grid load forecast data, based on charging request time The spatial distance between the vehicle's GPS coordinates and the charging station's WGS-84 coordinates at any given time is calculated using the spherical distance formula: ,in, Spatial distance between vehicle GPS coordinates and charging station WGS-84 coordinates. Earth's radius , : Latitude and radian values ​​of the vehicle and the charging station. Longitude difference in radians; Once the hardware parameters of the charging pile are standardized, they can be directly bound to the corresponding WGS-84 coordinates of the charging pile without needing to participate in spatiotemporal alignment.

[0008] Preferably, in step S2, the general basic weights of each feature data group are calculated through a spatiotemporally coupled multimodal attention mechanism, and the final dynamic weight adjustment is achieved by combining scene quantification indicators. Then, a joint feature vector is generated by calculating attention scores and weighted fusion. The user characteristic data set includes dynamic behavioral characteristics and static preference characteristics. Dynamic behavioral characteristics include real-time battery SOC and charging urgency. Static preference characteristics include historical charging time period preferences and historical charging station preferences based on historical data collection. The formula for calculating charging urgency is as follows: , urgency of charging :Target SOC , Real-time battery SOC : The duration of the charging time window; The power grid characteristic data set includes basic extracted features and calculated derived features. The basic extracted features are real-time load data and load forecast data. The calculated derived features include power grid vulnerability indicators, the proportion of renewable energy output, and power grid load margin. The calculation formula for the power grid vulnerability indicators is as follows: , Power grid vulnerability indicators Node voltage sensitivity Line capacity margin Rated voltage of the power grid Real-time voltage of the power grid Rated capacity of the line, Real-time operating capacity of the line; The facility feature data set includes basic extracted features and statistical calculation features. The basic extracted features are the real-time status of the charging pile, real-time output power, hardware parameters, and WGS-84 coordinates of the charging pile. The statistical calculation features include the predicted queuing time, the charging pile health status score, and the charging pile utilization rate. The formula for calculating the charging pile health status score is as follows: , Charging station health status score Historical failure rate : Timeliness of maintenance, 0.6: Historical failure rate correction item The weighting coefficient, 0.4: maintenance timeliness The weighting coefficients.

[0009] Preferably, the formula for calculating the general basic weight in step S2 is: , , Calculate and normalize to ensure that the user feature weights are normalized. Power grid characteristic weights Facility feature weights ,satisfy ; in, Sigmoid activation function 、 、 ： The spatiotemporally coupled multimodal attention mechanism uses adaptive coefficients learned in real time through an LSTM network. : Common basic weights for user feature data groups : General basic weights for power grid characteristic data groups : General basic weights for facility feature data groups.

[0010] Preferably, the joint feature vector generation method in step S2 is as follows: Calculate the attention matrix for each feature data group separately. , , , in, Feature data group identifier, Corresponding user feature data group Corresponding power grid characteristic data group Corresponding facility feature data group : Identified as The feature data set, namely For user feature data groups, For power grid characteristic data groups, For facility characteristic data group, Query matrix Key matrix Value matrix : Identified as The query matrix corresponding to the feature data group. : Identified as The key matrix corresponding to the feature data set. : Identified as The matrix of value values ​​corresponding to the feature data groups; Then calculate the attention score for each feature data group: ,in, : Identified as The attention score of the feature data set is denoted as ,Right now User feature attention score, Attention score for power grid features Attention score for facility features : softmax activation function Key matrix The transpose of the matrix, : Key vector dimension, The square root of the dimension of the key vector; Based on the final dynamic weights determined in step S2, the attention scores of each feature data group are weighted and fused to generate a joint feature vector: ,in, The final generated joint feature vector, , , The final weights are taken from the dynamic weight adjustment in step S2 and satisfy the normalization constraint. .

[0011] Preferably, the deep reinforcement learning global scheduling model described in step S3 is constructed based on the PPO algorithm, and the model optimization adopts the PPO overall objective function with KL divergence constraints, the expression of which is: ,in, The overall objective function of PPO with KL divergence constraints. For the trainable parameter set of the Actor policy network, : Regarding the first The mathematical expectation of a sample of each charging scheduling decision time step PPO pruning objective function, used to limit the magnitude of policy updates. : KL divergence penalty coefficient, used to adjust the constraint strength of the difference between the old and new strategies. The KL divergence between historical and current strategies measures the difference in probability distributions between the two strategies. Historical strategy probability distribution obtained from historical network parameters; : Based on current network parameters The obtained probability distribution of the current policy.

[0012] Preferably, the federated training process in step S3 specifically includes: The server generates the initial network parameters for the global scheduling model and sends them to the clients of each charging operator as the initial parameters for the client's local model. The client first performs geolocation blurring on the location information in the local historical charging data, uses the local data to complete model training, calculates and updates the gradient, and adds Gaussian noise to achieve differential privacy protection. The noise superposition method is as follows: ,in, : No. The first client The original gradients obtained from the first round of training, Gradient after adding noise The mean is 0 and the variance is 0. Gaussian noise, with noise variance determined based on a preset privacy budget; The federated client uploads the noisy gradient to the server using homomorphic encryption. The server shuffles the mixed encrypted gradient, aggregates the gradient using a weighted federated averaging algorithm, and updates the global scheduling model parameters. The server then distributes the updated global model parameters to each federated client. The client overwrites its local parameters and iterates until the model converges or reaches the preset maximum number of training rounds.

[0013] Preferably, the multi-objective optimization reward function expression in step S4 is: , , , ,in, The total reward value at time step t, used to measure the current state. Next action The scheduling effect; , , These are the reward weights for the power grid dimension, user dimension, and facility dimension, respectively. Rewards based on the power grid dimension Real-time power of the power grid For the target load of the power grid, The proportion of renewable energy output; User-based rewards : Facilities-related rewards.

[0014] Preferably, the anomaly detection in step S5 specifically involves: performing anomaly detection on the model parameters uploaded by each federated learning client based on statistical analysis, and calculating the average value of the parameters uploaded by all clients for each dimension of the model parameters. and standard deviation , will satisfy The parameters are determined to be abnormal, among which For the first A client uploads model parameters for a specific dimension, and the corresponding client is identified as an abnormal client.

[0015] By adopting the above technical solution, the present invention has the following advantages compared with the prior art: 1. This invention provides an intelligent charging scheduling method based on multi-dimensional feature fusion and federated learning. Through the collaborative optimization architecture of multi-dimensional feature fusion and federated learning, it achieves deep fusion of multi-source data and globally optimal charging scheduling without leaking user privacy and operator data assets.

[0016] 2. This invention provides an intelligent charging scheduling method based on multi-dimensional feature fusion and federated learning. Through unified standardized processing and spatiotemporal alignment mechanism, it eliminates the dimensional differences and spatiotemporal misalignment of multi-source data and breaks down data silos. At the same time, it constructs three major feature data groups of users, power grid and facilities to fully explore the value of data and provide complete and consistent data support for scheduling decisions.

[0017] 3. This invention provides an intelligent charging scheduling method based on multi-dimensional feature fusion and federated learning. It adopts a spatiotemporally coupled multimodal attention mechanism, learns adaptive coefficients in real time through an LSTM network, and combines scenario-based quantitative indicators such as grid load rate and renewable energy output ratio to achieve an organic combination of general basic weights and dynamic adjustment across all scenarios, covering all operating scenarios including normal, peak, off-peak, and emergency events. By calculating attention scores, it highlights the key dimensions within each feature data group. The joint feature vector generated by weighted fusion can simultaneously take into account the spatiotemporal coupling information and the recognizability of core features, greatly improving the accuracy of the scheduling strategy.

[0018] 4. This invention provides an intelligent charging scheduling method based on multi-dimensional feature fusion and federated learning. Based on the federated learning architecture, each charging operator, as a client, completes model training locally and only encrypts and uploads gradient parameters. The server-side uses a weighted federated averaging algorithm to aggregate and update the global model, avoiding the transmission of raw data across entities and protecting user privacy and operator data assets from the source. At the same time, the global scheduling model is built based on the PPO algorithm and achieves the global optimum of the scheduling strategy through objective function optimization with KL divergence constraints.

[0019] 5. This invention provides an intelligent charging scheduling method based on multi-dimensional feature fusion and federated learning. It embeds a digital watermark specific to the charging scenario and achieves unique identification of property rights without affecting model performance by dynamically adjusting the embedding strength. Combined with a periodically triggered sample verification mechanism, it effectively prevents the model from being stolen and tampered with. It establishes anomaly detection rules and hierarchical processing procedures based on statistical analysis to accurately identify malicious clients and abnormal parameters. Data backtracking and cleaning ensure the purity of model training and the stability of operation.

[0020] 6. This invention provides an intelligent charging scheduling method based on multi-dimensional feature fusion and federated learning. It collects feedback data such as satisfaction scores and rejection reasons through user terminal APP, converts them into reward correction items and supplementary training samples, and performs small-batch online fine-tuning of the global scheduling model to dynamically adapt to users' personalized needs and scenario changes. The discretized action space is adapted to the actual charging equipment specifications, and the multi-objective optimization reward function balances the interests of the power grid, users, and facilities, making the scheduling strategy more practical and executable. Attached Figure Description

[0021] Figure 1This is a flowchart of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0023] Example Please refer to Figure 1 As shown, this invention discloses an intelligent charging scheduling method based on multi-dimensional feature fusion and federated learning, characterized by the following steps: S1. Dynamic data from the user side, charging facility side, and power grid side are collected through the vehicle terminal, charging pile controller, and power grid monitoring equipment, respectively. The server performs standardization and spatiotemporal alignment processing on the collected data in sequence to form a consistent data foundation. The spatiotemporal alignment uses the charging request time collected by the user side as the unique time reference and the vehicle's GPS coordinates at the charging request time as the unique spatial reference. S2. Based on the data processed in S1, construct user feature data group, power grid feature data group and facility feature data group. Calculate the general basic weights of each feature data group through a spatiotemporally coupled multimodal attention mechanism. Combine the scenario quantification index to achieve the final dynamic weight adjustment. Then, generate a joint feature vector through attention score calculation and weighted fusion. S3. Construct a global scheduling model of deep reinforcement learning (DRL) managed by the server, with each charging operator as a federated learning client. The model training is completed through a federated training process of local training and encrypted parameter aggregation. At the same time, during the initialization phase of federated training of the global scheduling model, a digital watermark for charging scenarios is embedded into the model parameters. S4. Input the joint feature vector generated in S2 into the trained global scheduling model, and output a discretized charging scheduling strategy based on the multi-objective optimization reward function of grid load balancing, user cost minimization and charging efficiency maximization. S5. Perform anomaly detection on the model parameters uploaded by each federated learning client, identify abnormal clients based on the parameter statistical distribution, perform hierarchical anomaly handling and clean the abnormal data, then receive user feedback data and convert it into training samples, and fine-tune the deep reinforcement learning global scheduling model online.

[0024] The user-side data collected includes: real-time battery SOC and real-time battery data collected through the on-board diagnostic system (OBD). SOH Vehicle GPS coordinates (latitude and longitude), and charging request time, target SOC, and charging time window (start time and end time) collected via mobile APP.

[0025] The data collected on the charging facility side includes: charging pile status (idle / occupied / faulty), real-time output power, charging pile WGS-84 coordinates (latitude, longitude), maximum power, charging protocol type, and number of charging interfaces collected by the charging pile controller.

[0026] The data collected from the power grid side includes: real-time power, real-time voltage, and real-time frequency of the power grid collected by smart meters; power grid load forecast data (power sequence for the next 2 hours) and related data (renewable energy output and total output of the regional power grid) collected by the SCADA system; and preset inherent basic parameters of the power grid (regional power grid node voltage sensitivity, line capacity margin, rated load, rated voltage, rated line capacity, and power factor). ).

[0027] The standardization process in step S1 includes normalization of numerical data and One-Hot encoding of categorical data, and the charging time window is not included in the standardization process.

[0028] For numerical data, normalization is applied to map it to the [0,1] interval. The normalization formula is: ,in, For normalized data, For data before normalization, This represents the global maximum value of the data. This represents the global minimum value of the data. One-Hot encoding is used to convert categorical data into numerical values; the charging time window is converted into minutes in 24-hour format, and after conversion, it is not involved in standardization processing, but directly enters the spatiotemporal alignment process.

[0029] Step S1's spatiotemporal alignment process includes temporal alignment and spatial alignment: Time alignment: Extract charging request time The charging pile status, real-time grid power, real-time grid voltage, and real-time grid frequency are used to combine the timestamp of the grid load forecast data with the charging request time. Matching and calculating the charging time window with the charging request time. The time difference, all non-time dimension data are bound to the charging request time. Timestamp; Spatial alignment: Match the vehicle's GPS coordinates at the time of the charging request to the corresponding power grid topology node, and extract the charging request time of that node. Real-time grid load forecast data, based on charging request time The spatial distance between the vehicle's GPS coordinates and the charging station's WGS-84 coordinates at any given time is calculated using the spherical distance formula: ,in, Spatial distance between vehicle GPS coordinates and charging station WGS-84 coordinates. Earth's radius , : Latitude and radian values ​​of the vehicle and the charging station. Longitude difference in radians; Once the hardware parameters of the charging pile are standardized, they can be directly bound to the corresponding WGS-84 coordinates of the charging pile without needing to participate in spatiotemporal alignment.

[0030] The user feature data set mentioned in step S2 includes dynamic behavioral features and static preference features. The dynamic behavioral features include real-time battery SOC and charging urgency. The formula for calculating charging urgency is: ,in, urgency of charging :Target SOC , Real-time battery SOC : The duration of the charging time window; Static preference features include historical charging time period preferences and historical charging station preferences based on historical data collection statistics; The power grid characteristic data set includes basic extracted features and calculated derived features. The basic extracted features are real-time load data and load forecast data. The calculated derived features include power grid vulnerability indicators, the proportion of renewable energy output, and power grid load margin. The calculation formula for the power grid vulnerability indicators is as follows: ,in, Power grid vulnerability indicators Node voltage sensitivity Line capacity margin Rated voltage of the power grid Real-time voltage of the power grid Rated capacity of the line, Real-time operating capacity of the line; The facility feature data set includes basic extracted features and statistical calculation features. The basic extracted features are the real-time status of the charging pile, real-time output power, hardware parameters, and WGS-84 coordinates of the charging pile. The statistical calculation features include the predicted queuing time, the charging pile health status score, and the charging pile utilization rate. The formula for calculating the charging pile health status score is as follows: ,in Charging station health status score Historical failure rate : Timeliness of maintenance, 0.6: Historical failure rate correction item The weighting coefficient is higher because the historical failure rate of charging piles is a core indicator that objectively reflects the health level of the equipment. It directly determines the operational stability and failure risk of charging piles and has a greater impact on the priority of facility selection for charging scheduling. 0.4: Timeliness of maintenance The weighting coefficient, as an auxiliary indicator, reflects the equipment's operation and maintenance support capabilities in terms of timeliness of maintenance. It can make up for the static defects of historical failure rates (such as potential risks due to the lack of timely maintenance of some low-failure-rate equipment). The combination of the two can achieve a comprehensive static and dynamic assessment of the health status of charging piles.

[0031] The dynamic weight adjustment in step S2 is as follows: The formula for calculating the general basic weights is as follows: , , Calculate and normalize to ensure that the user feature weights are normalized. Power grid characteristic weights Facility feature weights ,satisfy ; in, Sigmoid activation function Map the initial weighting factors to the interval [0, 1]; 、 、 ： These are all adaptive coefficients specific to the spatiotemporally coupled multimodal attention mechanism, which is learned in real time by the attention mechanism through an LSTM network. The input to the LSTM network is spatiotemporally coupled feature data such as grid load rate, renewable energy output ratio, and charging request density over the past hour. : General basic weights of user feature data groups (initial weights without contextual adjustments). : General basic weights of power grid characteristic data groups (initial weights without scenario-specific adjustments). : General basic weights for facility feature data groups (initial weights without contextual adjustments). urgency of charging Power grid vulnerability indicators Charging station health status score.

[0032] Step S2, the joint feature vector generation, specifically involves the following steps, with all parameters traceable back to the previous steps and no new or sourced parameters: First, attention matrices are calculated for the user feature data group, the power grid feature data group, and the facility feature data group, respectively. The attention matrix includes a query matrix, a key matrix, and a value matrix, calculated as follows: , , , in, Feature data group identifier, Corresponding user feature data group Corresponding power grid characteristic data group Corresponding facility feature data group : Identified as The feature data set, namely For user feature data groups, For power grid characteristic data groups, For facility characteristic data group, Query matrix, trainable parameter matrix for spatiotemporally coupled multimodal attention mechanism. The key matrix represents the trainable parameter matrix for the spatiotemporally coupled multimodal attention mechanism. The Value matrix represents the trainable parameter matrix of the spatiotemporally coupled multimodal attention mechanism. : Identified as The query matrix corresponding to the feature data set has the same dimensions as the output dimension of the trainable parameter matrix. : Identified as The key matrix corresponding to the feature data group has the same dimensions as the query matrix. : Identified as The Value matrix corresponding to the feature data group has the same dimensions as the Query matrix. Then calculate the attention score for each feature data group: ,in, : Identified as The attention score of the feature data set is denoted as ,Right now User feature attention score, Attention score for power grid features The facility feature attention score is used to highlight the key dimensions within each feature data group. The output dimension is 64-dimensional, which is the preset output dimension of the spatiotemporally coupled multimodal attention mechanism. It can be adaptively adjusted according to the feature complexity of the actual charging scenario.

[0033] The softmax activation function normalizes the attention calculation results, ensuring they fall within the [0, 1] interval, thus guaranteeing that the attention score can be used for weighted fusion. Key matrix The transpose of the matrix is ​​used in conjunction with the query matrix. Perform matrix multiplication; : Key vector dimension, i.e., Key matrix The column dimensions are uniformly 64 across the entire text. The square root of the key vector dimension is used to scale the matrix multiplication result, avoiding training instability caused by excessively large calculation results. Based on the final dynamic weights determined in step S2, the attention scores of each feature data group are weighted and fused to generate a joint feature vector: ,in, The final generated joint feature vector serves as the sole input to the global scheduling model in step S3, with the same dimension as the attention score dimension (64 dimensions by default). , , The final weights are taken from the dynamic weight adjustment in step S2 and satisfy the normalization constraint. .

[0034] The deep reinforcement learning global scheduling model described in step S3 is built based on the PPO algorithm, including a hierarchical Actor policy network and a scene-adaptive Critic value network, and the overall input of both networks is a joint feature vector. The server uses Xavier to initialize network parameters; A hierarchical Actor policy network is used to output the probability distribution of charging scheduling actions. , The current charging scheduling state (represented by the joint feature vector). The charging scheduling process involves three layers of parallel decision-making: the first layer determines the charging time delay based on the grid load forecast data and the user's charging time window; the second layer allocates charging power levels based on the real-time battery SOC and the maximum power of the charging pile; and the third layer outputs charging station recommendations based on the spherical distance between the vehicle and the charging pile, the charging pile status, and the predicted queuing time.

[0035] The scenario-adaptive Critic value network is used to output a value estimate for the current state. The value assessment weights are dynamically adjusted based on the current operating scenario: during peak hours, priority is given to grid load balancing; during off-peak hours, priority is given to user charging costs; and in normal scenarios, the interests of the grid, users, and charging facilities are balanced.

[0036] The optimization uses the PPO total objective function with KL divergence constraints, expressed as: ,in, The overall objective function of PPO with KL divergence constraints. For the trainable parameter set of the Actor policy network, : Regarding the first The mathematical expectation of a sample of each charging scheduling decision time step PPO pruning objective function, used to limit the magnitude of policy updates. : KL divergence penalty coefficient, used to adjust the constraint strength of the difference between the old and new strategies. The KL divergence between historical and current strategies measures the difference in probability distributions between the two strategies. Historical strategy probability distribution obtained from historical network parameters; : Based on current network parameters The obtained probability distribution of the current policy.

[0037] The federated training process in step S3 specifically includes: The server generates the initial network parameters for the global scheduling model and sends them to the clients of each charging operator as the initial parameters for the client's local model. The client first performs geolocation blurring on the location information in the local historical charging data, then uses the local historical data to train a local model. After training, the model update gradient is calculated, and Gaussian noise is added to the gradient to achieve differential privacy protection. The noise superposition method is as follows: ,in, : No. The first client The original gradients obtained from the first round of training, Gradient after adding noise The mean is 0 and the variance is 0. Gaussian noise, with noise variance determined based on a preset privacy budget; The client uses homomorphic encryption to upload the noisy gradient to the server. The server scrambles and mixes the received encrypted gradients, then aggregates the gradients using a weighted federated average algorithm, and updates the global model parameters based on the aggregated gradients. The server distributes the updated global model parameters to each client. The client uses the global parameters to override the local parameters and enters the next iteration until the model converges or reaches the preset maximum number of training rounds, at which point training stops.

[0038] The process of embedding a digital watermark specifically for the charging scenario in step S3 is as follows: A dedicated trigger sample matching the charging scheduling scenario is constructed. The trigger sample is composed of charging request-related features. During global model training, the watermark information is embedded into the model parameters. The embedding method forces the model to output a preset fixed action for the trigger sample. The watermark embedding strength is dynamically adjusted with each training round to avoid affecting the model's convergence effect. After the model is deployed online, the server periodically inputs the trigger sample into the model for verification. If the deviation between the model output and the preset fixed action exceeds a preset threshold (e.g., 10%), it is determined that the model has a risk of misuse, triggering a copyright warning and restricting the model's usage rights.

[0039] During global model training, the watermark information is embedded into the model parameters using the following formula: ,in, The parameter set of the global scheduling model. Watermark embedding strength coefficient; : The sign function, which outputs ±1, is used to control the direction of parameter updates; The watermark is a fixed hash sequence obtained by performing a SHA-256 hash operation on the copyright information.

[0040] To avoid the watermark affecting model convergence and scheduling accuracy, the watermark strength is dynamically changed with each training epoch: ,in, The current training epoch of the model, with an initial strength of . As the training rounds progress, the watermark is gradually embedded to ensure stable embedding without compromising model performance.

[0041] The action space of the global scheduling model described in step S4 is designed with discretization to adapt to the device specifications and user needs of actual charging scenarios, specifically including three dimensions: Charging time delay: The discrete action set is {0,15,30,45,60,75,90,105,120} minutes, with a time interval of 15 minutes, covering charging delay requirements of 0-2 hours; Charging power allocation: The tiered adjustment set is {20kW, 40kW, 60kW}, which is compatible with the maximum power specifications of mainstream charging piles, taking into account both charging speed and grid load pressure; Charging station recommendation: Output a list of Top-3 charging station IDs, sorted by priority based on the spherical distance between the vehicle and the charging pile (taken from the spatial alignment result in step S1), the health status score of the charging pile (taken from the facility feature data group in step S2), and the real-time idle status of the charging pile (taken from the data collected in step S1). The multi-objective optimization reward function is used to quantitatively evaluate the merits of scheduling strategies and provide optimization guidance for model training. Its core expression is: , , , ,in, The total reward value at time step t, used to measure the current state. Next action The scheduling effect; , , These represent the reward weights for the power grid, user, and infrastructure dimensions, respectively, with default values ​​of [value]. =0.5、 =0.3、 =0.2, and can be adaptively adjusted according to the final dynamic weighting rule in step S2 (increase during peak hours). Improve during off-peak hours Emergency escalation ); Rewards based on the power grid dimension, focusing on grid load balancing and renewable energy utilization. The real-time power of the power grid is taken from the power grid side data collected in step S1. The target load of the power grid is determined based on the rated load of the power grid and the output of renewable energy in step S1. The proportion of renewable energy output (taken from the grid characteristic data group in step S2). User-centric rewards focus on minimizing user charging costs. Charging cost = charging power × charging time × electricity price during the time period (charging power is taken from the automatic operation space, and charging time is determined by the battery's real-time SOC and target). SOC (Calculations are made based on the queuing time predictions in the facility characteristic data set, with electricity prices related to real-time grid power). The facility-level rewards focus on maximizing charging efficiency and balancing the load of charging piles. The load variance of charging piles is calculated based on the real-time output power of the charging piles.

[0042] 0.2: Percentage of renewable energy output The reward coefficient is designed based on: The core optimization objective in the power grid dimension is load balancing (in the formula) Therefore, the main weight is assigned to load balancing, and the proportion of renewable energy output is used as an auxiliary optimization target to guide the dispatch strategy to prioritize the use of clean energy, which is in line with the low-carbon operation requirements of the power grid. Extensive power grid simulation experiments have verified that 0.2 is the optimal coefficient: if the coefficient is too high, the dispatch strategy will overemphasize the use of renewable energy and disrupt the load balance of the power grid; if the coefficient is too low, it will not effectively reflect the dispatch orientation of clean energy. 0.2 can achieve a balance between the dual objectives of load balance and renewable energy consumption.

[0043] 0.6: The weighting coefficient for charging costs is a core dimension of user charging experience. New energy vehicle users are far more sensitive to charging costs than waiting time, so it is given a higher weight to align with users' actual consumption preferences. 0.4: The weighting coefficient for waiting time, serving as a secondary indicator of user experience, takes into account charging efficiency needs, avoiding excessively long waiting times for users due to the pursuit of low cost, and balancing cost control and efficiency experience.

[0044] The final output charging scheduling strategy is the total reward value in the above action space. The largest action combination, specifically output includes: charging start time, charging end time (calculated from the charging request time collected in step S1 + time delay + charging duration), graded charging power, Top-3 recommended charging station IDs and corresponding navigation information (generated based on the spherical distance calculated based on spatial alignment in step S1).

[0045] In step S5, the online fine-tuning of the model based on user feedback and the detection and handling of client-side anomalies are as follows: Online model fine-tuning based on user feedback Feedback collection: Display relevant information about the scheduling strategy (recommended charging station locations, charging costs, and waiting times) through the user terminal APP, and provide interactive functions, including allowing users to set the maximum acceptable charging delay via a slider, confirm or reject the execution of the scheduling strategy, and give a 1-5 star satisfaction rating for the executed scheduling strategy.

[0046] Feedback data processing: Record the specific reasons why users refuse the dispatch strategy (including but not limited to excessively long waiting time, excessively high charging costs, and excessively long distances), and convert user satisfaction scores into reward adjustment items. The formula for calculating reward adjustment items is as follows: .

[0047] Online fine-tuning: The server merges user feedback data (rejection reasons, reward correction items, satisfaction scores) with the original data from steps S1 and S2 as supplementary training samples to perform small-batch iterative fine-tuning of the global scheduling model; after fine-tuning, the updated global model parameters are sent to each federated learning client (the local model training node for each charging operator) to realize dynamic model iteration and improve the model's adaptability to users' personalized needs.

[0048] (II) Client-side anomaly detection and handling Anomaly detection rules: Anomaly detection is performed on the model parameters uploaded by each federated learning client based on statistical analysis. The average value of the model parameters uploaded by all clients is calculated for each dimension of the model parameters. and standard deviation , will satisfy The parameters are determined to be abnormal, among which For the first A client uploads model parameters for a specific dimension, and the corresponding client is identified as an abnormal client.

[0049] Graded anomaly handling process: Upon first detection of an anomaly: The server sends a parameter verification request to the client, requiring the client to recollect data, retrain the model, and upload the parameters.

[0050] If an anomaly is detected a second time: mark the client as a suspicious client, suspend its participation in federated learning for one training cycle, and restore its participation after the client completes parameter verification and passes the verification.

[0051] If anomalies are detected three or more times: the client will be classified as a malicious client, permanently banned from participating in federated learning, and an audit log (including client ID, abnormal parameters, abnormal time, training rounds, etc.) will be recorded for subsequent traceability.

[0052] Abnormal data cleaning: The server performs backtracking checks on historical training data uploaded by suspicious or malicious clients to investigate traces of data tampering or forgery. If traces of data tampering are found, all historical training data contributed by that client is immediately deleted. During retraining, the server retains the valid historical training parameters of non-abnormal clients and only removes data from abnormal clients to improve retraining efficiency. After data cleaning, the model's accuracy is verified to ensure that the model scheduling performance is not affected by abnormal data.

[0053] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A smart charging scheduling method based on multi-dimensional feature fusion and federated learning, characterized in that: Includes the following steps: S1. Collect dynamic data from the user side, charging facility side, and power grid side respectively, and perform standardization and spatiotemporal alignment processing in sequence. S2. Based on the data processed in S1, construct user feature data group, power grid feature data group and facility feature data group. After weight calculation and scenario-based adjustment through a spatiotemporally coupled multimodal attention mechanism, weighted fusion is used to generate a joint feature vector. S3. Construct a server-side deep reinforcement learning global scheduling model, with charging operators as federated clients, train the model through a federated training process, and embed a digital watermark specific to the charging scenario during the initialization phase. S4. Input the joint feature vector generated in S2 into the trained global scheduling model, and output a discretized charging scheduling strategy based on the multi-objective optimization reward function. S5. Detect the model parameters of each federated learning client, identify abnormal clients, perform hierarchical anomaly handling and clean the abnormal data, then receive user feedback data and convert it into training samples, and fine-tune the deep reinforcement learning global scheduling model online.

2. The intelligent charging scheduling method based on multi-dimensional feature fusion and federated learning as described in claim 1, characterized in that, The standardization process in step S1 includes: For numerical data, normalization is applied to map it to the [0,1] interval. The normalization formula is: ,in, For normalized data, For data before normalization, This represents the global maximum value of the data. This represents the global minimum value of the data. One-Hot encoding is used to convert categorical data into numerical values; the charging time window is converted into minutes in 24-hour format, and after conversion, it is not involved in standardization processing, but directly enters the spatiotemporal alignment process.

3. The intelligent charging scheduling method based on multi-dimensional feature fusion and federated learning as described in claim 1, characterized in that: Step S1, the spatiotemporal alignment process, includes temporal alignment and spatial alignment. Time alignment: Extract charging request time The charging pile status, real-time grid power, real-time grid voltage, and real-time grid frequency are used to combine the timestamp of the grid load forecast data with the charging request time. Matching and calculating the charging time window with the charging request time. The time difference, all non-time dimension data are bound to the charging request time. Timestamp; Spatial alignment: Match the vehicle's GPS coordinates at the time of the charging request to the corresponding power grid topology node, and extract the charging request time of that node. Real-time grid load forecast data, based on charging request time The spatial distance between the vehicle's GPS coordinates and the charging station's WGS-84 coordinates at any given time is calculated using the spherical distance formula: ,in, Spatial distance between vehicle GPS coordinates and charging station WGS-84 coordinates. Earth's radius , : Latitude and radian values ​​of the vehicle and the charging station. Longitude difference in radians; Once the hardware parameters of the charging pile are standardized, they can be directly bound to the corresponding WGS-84 coordinates of the charging pile without needing to participate in spatiotemporal alignment.

4. The intelligent charging scheduling method based on multi-dimensional feature fusion and federated learning as described in claim 1, characterized in that: In step S2, the general basic weights of each feature data group are calculated through a spatiotemporally coupled multimodal attention mechanism, and the final dynamic weight adjustment is achieved by combining scene quantification indicators. Then, a joint feature vector is generated by calculating attention scores and weighted fusion. The user characteristic data set includes dynamic behavioral characteristics and static preference characteristics. Dynamic behavioral characteristics include real-time battery SOC and charging urgency. Static preference characteristics include historical charging time period preferences and historical charging station preferences based on historical data collection. The formula for calculating charging urgency is as follows: , urgency of charging :Target SOC , Real-time battery SOC : The duration of the charging time window; The power grid characteristic data set includes basic extracted features and calculated derived features. The basic extracted features are real-time load data and load forecast data. The calculated derived features include power grid vulnerability indicators, the proportion of renewable energy output, and power grid load margin. The calculation formula for the power grid vulnerability indicators is as follows: , Power grid vulnerability indicators Node voltage sensitivity Line capacity margin Rated voltage of the power grid Real-time voltage of the power grid Rated capacity of the line, Real-time operating capacity of the line; The facility feature data set includes basic extracted features and statistical calculation features. The basic extracted features are the real-time status of the charging pile, real-time output power, hardware parameters, and WGS-84 coordinates of the charging pile. The statistical calculation features include the predicted queuing time, the charging pile health status score, and the charging pile utilization rate. The formula for calculating the charging pile health status score is as follows: , Charging station health status score Historical failure rate : Timeliness of maintenance, 0.6: Historical failure rate correction item The weighting coefficient, 0.4: maintenance timeliness The weighting coefficients.

5. The intelligent charging scheduling method based on multi-dimensional feature fusion and federated learning as described in claim 4, characterized in that: The formula for calculating the general basic weights mentioned in step S2 is as follows: , , Calculate and normalize to ensure that the user feature weights are normalized. Power grid characteristic weights Facility feature weights ,satisfy ; in, Sigmoid activation function 、 、 ： The spatiotemporally coupled multimodal attention mechanism uses adaptive coefficients learned in real time through an LSTM network. : Common basic weights for user feature data groups : General basic weights for power grid characteristic data groups : General basic weights for facility feature data groups.

6. The intelligent charging scheduling method based on multi-dimensional feature fusion and federated learning as described in claim 1, characterized in that, The joint feature vector generation method in step S2 is as follows: Calculate the attention matrix for each feature data group separately. , , , in, Feature data group identifier, Corresponding user feature data group Corresponding power grid characteristic data group Corresponding facility feature data group : Identified as The feature data set, namely For user feature data groups, For power grid characteristic data groups, For facility characteristic data group, Query matrix Key matrix Value matrix : Identified as The query matrix corresponding to the feature data group. : Identified as The key matrix corresponding to the feature data set. : Identified as The matrix of value values ​​corresponding to the feature data groups; Then calculate the attention score for each feature data group: ,in, : Identified as The attention score of the feature data set is denoted as ,Right now User feature attention score, Attention score for power grid features Attention score for facility features : softmax activation function Key matrix The transpose of the matrix, : Key vector dimension, The square root of the dimension of the key vector; Based on the final dynamic weights determined in step S2, the attention scores of each feature data group are weighted and fused to generate a joint feature vector: ,in, The final generated joint feature vector, , , The final weights are taken from the dynamic weight adjustment in step S2 and satisfy the normalization constraint. .

7. The intelligent charging scheduling method based on multi-dimensional feature fusion and federated learning as described in claim 1, characterized in that, The deep reinforcement learning global scheduling model described in step S3 is constructed based on the PPO algorithm, and the model optimization adopts the PPO overall objective function with KL divergence constraints, the expression of which is: ,in, The overall objective function of PPO with KL divergence constraints. For the trainable parameter set of the Actor policy network, : Regarding the first The mathematical expectation of a sample of each charging scheduling decision time step PPO pruning objective function, used to limit the magnitude of policy updates. : KL divergence penalty coefficient, used to adjust the constraint strength of the difference between the old and new strategies. The KL divergence between historical and current strategies measures the difference in probability distributions between the two strategies. Historical strategy probability distribution obtained from historical network parameters; : Based on current network parameters The obtained probability distribution of the current policy.

8. The intelligent charging scheduling method based on multi-dimensional feature fusion and federated learning as described in claim 1, characterized in that, The federated training process in step S3 specifically includes: The server generates the initial network parameters for the global scheduling model and sends them to the clients of each charging operator as the initial parameters for the client's local model. The client first performs geolocation blurring on the location information in the local historical charging data, uses the local data to complete model training, calculates and updates the gradient, and adds Gaussian noise to achieve differential privacy protection. The noise superposition method is as follows: ,in, : No. The first client The original gradients obtained from the first round of training, Gradient after adding noise The mean is 0 and the variance is 0. Gaussian noise, with noise variance determined based on a preset privacy budget; The federated client uploads the noisy gradient to the server using homomorphic encryption. The server shuffles the mixed encrypted gradient, aggregates the gradient using a weighted federated averaging algorithm, and updates the global scheduling model parameters. The server then distributes the updated global model parameters to each federated client. The client overwrites its local parameters and iterates until the model converges or reaches the preset maximum number of training rounds.

9. The intelligent charging scheduling method based on multi-dimensional feature fusion and federated learning as described in claim 1, characterized in that, The multi-objective optimization reward function expression in step S4 is: , , , ,in, The total reward value at time step t, used to measure the current state. Next action The scheduling effect; , , These are the reward weights for the power grid, user, and infrastructure dimensions, respectively. Rewards based on the power grid dimension Real-time power of the power grid For the target load of the power grid, The proportion of renewable energy output; User-based rewards : Facilities-related rewards.

10. The intelligent charging scheduling method based on multi-dimensional feature fusion and federated learning as described in claim 1, characterized in that, The anomaly detection in step S5 specifically involves: performing anomaly detection on the model parameters uploaded by each federated learning client based on statistical analysis, and calculating the average value of the parameters uploaded by all clients for each dimension of the model parameters. and standard deviation , will satisfy The parameters are determined to be abnormal, among which For the first A client uploads model parameters for a specific dimension, and the corresponding client is identified as an abnormal client.

Images