Multifunctional slot intelligent charging pile remote control and dispatching system

Abstract

The invention discloses a multifunctional slot intelligent charging pile remote control and dispatching system, particularly relates to the field of charging piles, is used for solving the problems of low charging resource utilization rate and large power grid load impact, and provides an accurate basis for power distribution by collecting and predicting user charging requirements and vehicle charging capacity in real time based on federal learning. A prediction result is combined with real-time load data of the power grid through a multi-objective optimization algorithm, a power distribution baseline value considering user requirements and peak clipping and valley filling of the power grid is generated, and charging resources and the stability of the power grid are effectively balanced; a reinforcement learning algorithm is adopted to construct a dynamic game model of a user and a power grid, the output power of the charging pile is adjusted in real time, user behaviors are guided through an integral excitation strategy, conflicts between demands and loads are dynamically coped, and the system adaptability is improved; and finally, performing full-link simulation verification on the charging pile group by using a digital twinning technology, comparing a continuous optimization scheduling strategy based on actual operation data, and ensuring continuous improvement of system performance.

Description

Technical Field

[0001] The present invention relates to the field of charging piles, and more specifically, to a remote control and scheduling system for multi-functional slot intelligent charging piles. Background Art

[0002] In the actual operation of smart charging pile systems, guiding users to charge during off-peak hours through time-of-use electricity pricing strategies is a key means of achieving peak load shifting. However, users' actual charging behavior is often affected by factors such as travel schedule changes and temporary vehicle needs, resulting in discrepancies between scheduled charging times and actual conditions. Furthermore, the battery status of different vehicles (such as remaining charge and charging rate limits) changes dynamically during the charging process, making the system's pre-set fixed power allocation strategy difficult to adapt to real-time demand. Existing scheduling systems often design peak load shifting solutions based on static electricity price time periods and idealized assumptions about user behavior, lacking the ability to coordinate dynamic behavior and battery status. When actual user charging needs misalign with pre-set time periods, or when vehicle batteries cannot charge at the expected power, the system is unable to dynamically adjust charging task priorities and struggles to balance grid load regulation objectives with personalized user needs. This rigid scheduling logic not only wastes charging resources during off-peak hours but can also exacerbate grid fluctuations due to inaccurate power allocation during peak hours, ultimately undermining the effectiveness of peak load shifting strategies.

[0003] In order to solve the above problems, a technical solution is now provided. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides a multi-functional slot smart charging pile remote control and scheduling system. First, based on the federated learning technology, the user charging demand and vehicle charging capacity are collected and predicted in real time, ensuring data privacy while providing an accurate basis for power allocation; the prediction results are combined with the real-time load data of the power grid through a multi-objective optimization algorithm to generate a power allocation baseline value that takes into account user demand and peak shaving and valley filling of the power grid, effectively balancing charging resources and power grid stability; a reinforcement learning algorithm is used to construct a dynamic game model between users and power grids, and the output power of the charging pile is adjusted in real time. The user behavior is guided by the integral incentive strategy, dynamically responding to the conflict between demand and load, and improving the adaptability of the system; finally, the digital twin technology is used to perform full-link simulation verification of the charging pile group, and the scheduling strategy is continuously optimized based on the actual operation data comparison to ensure the continuous improvement of system performance, so as to solve the problems raised in the above background technology.

[0005] To achieve the above object, the present invention provides the following technical solutions: Federated Learning Module: This module collects vehicle arrival time deviations and battery status parameters in real time, and builds user behavior prediction models and battery charging capacity models based on federated learning. Optimization and scheduling module: This module combines the output of the user behavior prediction model with real-time grid load data to generate a power allocation baseline value through a multi-objective optimization algorithm. Dynamic Game Module: Based on a reinforcement learning algorithm, this module builds a dynamic game model between users and the grid. It integrates user behavior deviation information and grid load relief information, adjusts the output power of charging piles in real time according to the power allocation baseline value, and drives user charging behavior toward low-load periods based on a point incentive strategy. Simulation verification module: Use digital twin technology to perform full-link simulation verification of the charging pile group, compare simulation data with actual operation data, and output optimization prompt signals.

[0006] Furthermore, the federated learning module collects the vehicle arrival time deviation and battery status information of the charging pile in real time and performs preprocessing; based on the federated learning technology, the local model is trained on each charging pile node, and the global model is generated through aggregation through the central server.

[0007] Furthermore, the federated learning module uses support vector regression to build a battery charging capacity model to predict the vehicle's charging capacity; finally, the trained model is deployed to each charging pile node to generate charging demand prediction values ​​and charging capacity prediction values ​​in real time.

[0008] Furthermore, the optimization scheduling module obtains the charging demand forecast value and the charging capacity forecast value, and collects the grid load data and the grid load target curve in real time; smoothes the real-time grid load and calculates the grid load deviation; constructs a multi-objective optimization model, using the power allocation baseline value of each charging pile at each time point as the decision variable, and simultaneously realizes the optimization of grid peak shaving and valley filling and user charging demand satisfaction; sets constraints to ensure that the power allocation baseline value is within the capacity of the charging pile, meets the charging demand and controls the grid load deviation; designs the objective function, uses the exponential decay penalty form to evaluate the grid peak shaving and valley filling effect, and uses the margin index form to measure the user charging demand satisfaction; uses the improved particle swarm optimization algorithm to obtain the optimal solution; and finally generates the power allocation baseline value of each charging pile at each time point.

[0009] Furthermore, the user behavior deviation information includes a user behavior deviation index. The user behavior deviation index is obtained by calculating the relative deviation between the actual charging power and the predicted charging power within a time window and performing nonlinear mapping, and its value range is 0 to 1.

[0010] Furthermore, the grid load relief information includes a grid load relief index, which is obtained by evaluating the average ratio of the difference between the real-time load and the load target in the historical window and converting it through a Sigmoid function, and has a value range of 0 to 1.

[0011] Furthermore, the dynamic game module calculates the average value of the user behavior deviation index of all vehicles at the selected time point; the geometric mean method is used to integrate the average value and the inverse value of the grid load relief index to obtain the user-grid response synergy index.

[0012] Furthermore, the dynamic game module adopts a multi-agent reinforcement learning framework, treating each charging pile as an agent and the power grid as part of the environment; the state space includes the grid load at the current time point, the power allocation baseline value, the actual output power of each charging pile, and the user-grid response synergy index; the action space is the power increase or decrease relative to the power allocation baseline value, and the amplitude is limited by the adjustable range of a single pile; the reward function is designed to be the inverse of the weighted sum of the user-grid response synergy index and the grid load deviation; training adopts the deep Q network algorithm, and the experience replay mechanism is used to improve learning stability.

[0013] Furthermore, if the actual output power continues to deviate from the power allocation baseline value and the user-grid response synergy index exceeds the preset threshold, the dynamic game module guides users to adjust their charging behavior through point incentives.

[0014] Furthermore, the simulation verification module uses a dynamic time warping algorithm to align the simulation data with the actual data, and generates an optimization prompt signal if the power deviation or load deviation exceeds the corresponding preset threshold.

[0015] The technical effects and advantages of the multifunctional slot intelligent charging pile remote control and scheduling system of the present invention are as follows: The present invention significantly improves the utilization rate of charging resources, optimizes grid load management, enhances user experience and realizes intelligent remote management through the remote control and dispatching system of multi-functional slot intelligent charging piles. First, based on the federated learning technology, the user charging demand and vehicle charging capacity are collected and predicted in real time, ensuring data privacy while providing an accurate basis for power allocation; the prediction results are combined with the real-time load data of the grid through a multi-objective optimization algorithm to generate a power allocation baseline value that takes into account both user demand and grid peak shaving and valley filling, effectively balancing charging resources and grid stability; a reinforcement learning algorithm is used to construct a dynamic game model between users and the grid, and the output power of the charging pile is adjusted in real time and user behavior is guided by an integral incentive strategy to dynamically respond to conflicts between demand and load, thereby improving the adaptability of the system; finally, the digital twin technology is used to perform full-link simulation verification of the charging pile group, and the scheduling strategy is continuously optimized based on the actual operation data comparison to ensure the continuous improvement of system performance. In summary, the above effectively solves the problems of waste of resources, management lag and grid impact of traditional charging piles, promotes the intelligent and sustainable development of charging infrastructure, and has significant practical value and promotion potential. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1This is a structural diagram of the multi-functional slot smart charging pile remote control and scheduling system of the present invention.

[0017] Figure 2 This is a schematic diagram of the operation flow of the federated learning module of the multi-functional slot smart charging pile remote control and scheduling system of the present invention. DETAILED DESCRIPTION

[0018] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0019] Example 1: Figure 1 The present invention provides a multifunctional slot intelligent charging pile remote control and scheduling system, including: Federated Learning Module: This module collects vehicle arrival time deviations and battery status parameters in real time, and builds user behavior prediction models and battery charging capacity models based on federated learning. Optimization and scheduling module: This module combines the output of the user behavior prediction model with real-time grid load data to generate a power allocation baseline value through a multi-objective optimization algorithm. Dynamic Game Module: Based on a reinforcement learning algorithm, this module builds a dynamic game model between users and the grid. It integrates user behavior deviation information and grid load relief information, adjusts the output power of charging piles in real time according to the power allocation baseline value, and drives user charging behavior toward low-load periods based on a point incentive strategy. Simulation verification module: Use digital twin technology to perform full-link simulation verification of the charging pile group, compare simulation data with actual operation data, and output optimization prompt signals.

[0020] In the remote control and scheduling method for smart charging piles, the federated learning module aims to utilize federated learning technology to build user behavior prediction models and battery charging capacity models based on vehicle arrival time deviations and battery status parameters collected in real time by charging piles. This generates accurate charging demand and charging capacity predictions. These predictions are directly used in the optimized scheduling module to calculate power allocation baselines in conjunction with grid load data. This enables charging piles to dynamically respond to user demand and vehicle status, supporting grid load balancing.

[0021] like Figure 2 As shown in Figure 2, the federated learning module includes the following: S1-1, data acquisition and preprocessing: Each charging station records relevant data in real time and performs preliminary processing. Each charging station is referred to as a node, and nodes are numbered sequentially from 1 to the total number. For each vehicle, the difference between its actual arrival time and expected arrival time is recorded, referred to as the vehicle arrival time deviation (VAT), in minutes. Battery status information is also collected for each vehicle, including the remaining battery charge (expressed as a percentage), total battery capacity (in kilowatt-hours), and the battery's maximum allowed charging power (in kilowatts). During the preprocessing phase, outliers are first removed: if a vehicle's arrival time deviation exceeds 120 minutes, or its remaining battery charge is less than 0% or greater than 100%, the vehicle's data is considered an outlier and removed. Next, the remaining data is normalized. For each vehicle's arrival time deviation, remaining battery charge, total battery capacity, and maximum charging power, a processed value is calculated. This is achieved by subtracting the minimum value of that parameter from the charging station's historical value, and then dividing the result by the difference between the maximum and minimum values ​​of that parameter at the charging station. This yields a value between 0 and 1. After processing, each charging pile obtains a set of normalized data, including the normalized values ​​of each vehicle’s arrival time deviation, remaining battery power, total battery capacity, and maximum charging power.

[0022] S1-2, construction of federated learning framework: In this step, a federated learning framework is constructed based on the local data of each charging station. First, each charging station trains a local model using locally collected and normalized data. This local model uses a multi-layer perceptron architecture, consisting of an input layer that receives four normalized features (i.e., normalized values ​​of arrival time deviation, remaining battery charge, total battery capacity, and maximum charging power); two hidden layers, each containing 32 neurons and using the Reluctant Unit (ReLU) activation function; and an output layer that adjusts the output based on subsequent needs. Model parameters are updated using gradient descent with a fixed learning rate of 0.001, processing 32 data points at a time. After training, each charging station calculates the parameter updates for the local model and uploads them to the central server. To protect data privacy, the parameter updates are noised before upload by first calculating the Euclidean norm of the parameter updates, multiplying it by 0.01, then multiplying it by a random number drawn from a standard normal distribution, and adding the resulting value to the original parameter updates. After receiving the noisy parameter updates uploaded by all charging stations, the central server aggregates them by averaging. Specifically, the current model parameters of each charging station are added to the noisy parameter update, and the average of these results across all charging stations is taken to obtain the global model parameters. The aggregated global model parameters are then distributed back to each charging station. Ultimately, each charging station obtains the global model parameters for subsequent training.

[0023] S1-3, establishment of user behavior prediction model: A user behavior prediction model is constructed to predict the amount of electricity each vehicle will need to replenish in the future, referred to as the charging demand forecast, expressed in kilowatt-hours. The model inputs are each vehicle's normalized arrival time deviation, remaining battery charge, total battery capacity, and maximum charging power, and outputs the charging demand forecast for each vehicle. The model utilizes a long short-term memory (LSTM) network architecture, consisting of an input layer that receives four normalized parameters, a bidirectional LSTM layer with 64 hidden units, and an output layer that outputs the charging demand forecast using a single neuron. During training, the loss function is defined as the sum of the negative exponential functions of the absolute value of the difference between the predicted and actual charging demand. This loss function is calculated as follows: for each vehicle, the absolute value of the difference between the predicted and actual values ​​is calculated, the negative exponential of the difference is taken, and this value is summed across all vehicles. The model is initialized using the global model parameters aggregated through federated learning. Training is terminated using an early stopping strategy: if the validation set loss does not decrease within five consecutive training epochs, training is terminated. After training is complete, each charging station receives the user behavior prediction model and outputs the charging demand forecast for each vehicle.

[0024] S1-4, battery charging capability model establishment: A battery charging capability model is constructed to predict the maximum acceptable charging power for each vehicle in its current state. This is called the predicted charging capability value, expressed in kilowatts. The model inputs are each vehicle's normalized remaining battery charge, total battery capacity, and maximum charging power, and the output is the predicted charging capability value for each vehicle. The model employs support vector regression with a radial basis kernel function, kernel parameter set to 0.1, and regularization parameter set to 1.0. The loss function is defined as the sum of the squared reciprocals of the difference between the predicted and actual charging capabilities. The specific calculation method is: for each vehicle, the square of the difference between the predicted and actual values ​​is calculated, a constant of 0.01 is added to prevent the denominator from being zero, the reciprocal of this difference is taken, and the resulting value is summed across all vehicles. During training, the relevant parameters are initialized using the global model parameters aggregated in federated learning. The kernel and regularization parameters are optimized using a grid search method. After training, each charging station obtains a battery charging capability model and outputs a predicted charging capability value for each vehicle.

[0025] S1-5, model deployment and real-time prediction: The user behavior prediction model and battery charging capacity model are deployed at each charging station for real-time prediction. The real-time inputs are the collected and normalized arrival time deviation, remaining battery charge, total battery capacity, and maximum charging power. The model outputs a predicted charging demand and charging capacity for each vehicle, respectively. Subsequently, a summary is performed for each charging station: the total charging demand is calculated as the sum of the predicted charging demand values ​​for all vehicles at that station; and the total charging capacity is calculated as the sum of the predicted charging capacity values ​​for all vehicles at that station, subject to the maximum output power limit of the station. Finally, each charging station generates the total charging demand and total charging capacity, which serve as input data for subsequent steps.

[0026] The federated learning module uses federated learning technology to build user behavior prediction models and battery charging capacity models using vehicle arrival time deviations and battery status parameters collected by charging stations. This generates charging demand and charging capacity predictions for each vehicle, and aggregates them into total charging demand and charging capacity for each charging station. This data will be used in subsequent steps to calculate power allocation baselines in conjunction with grid load data, ensuring efficient charging station scheduling based on dynamic user demand and vehicle status.

[0027] The optimization scheduling module includes the following: S2-1, data preparation: During the data preparation phase, input data is first collected and preprocessed. This input data consists of two categories: one is the charging demand and charging capacity forecasts obtained for each charging station, reflecting the expected user demand for electricity in the future time period and the maximum power output capacity of the charging station, respectively; the other is the real-time grid load data and the grid load target curve, representing the total grid load at the current moment and the expected future load level, respectively. The preprocessing process first smoothes the real-time grid load using an exponential smoothing method. This is achieved by taking the weighted sum of the current real-time load and the smoothed load at the previous moment. The weights are designed based on the smoothing factor and its complement to reduce the interference of short-term fluctuations. The initial smoothed load is based on the initial real-time load. Subsequently, the grid load deviation is calculated as the difference between the smoothed real-time load and the target load.

[0028] S2-2, multi-objective optimization model construction: When constructing a multi-objective optimization model, the goal is to simultaneously maximize the peak-shaving and valley-filling effect of the power grid and optimize the satisfaction of user charging needs. The decision variable is defined as the power allocation baseline value of each charging pile at each time point. The model needs to meet several constraints: First, the power allocation baseline value of each charging pile at any time point must be between zero and its charging capacity forecast value; second, the total power provided by each charging pile during the forecast time period is calculated by accumulating its power allocation baseline values ​​at each time point and must not be lower than its charging demand forecast value; third, the total load of the power grid at any time point is composed of the sum of the smoothed real-time load and the power allocation baseline values ​​of all charging piles at that time point. The deviation from the target load should be controlled within the allowable range of the power grid scheduling requirements.

[0029] S2-3, optimization objective function: To achieve the above optimization goals, two objective functions are constructed. The first objective function is used to evaluate the peak-shaving and valley-filling effect of the power grid, using an exponential decay penalty: for each time point, the absolute value of the difference between the smoothed real-time load and the sum of the power distribution baseline values ​​of all charging piles relative to the target load is calculated, and after dividing it by a decay factor related to the average level of the power grid load, the exponential value with a natural constant as the base is taken, and then the exponential values ​​of all time points are accumulated. The goal is to minimize this sum. The second objective function is used to measure the degree to which the user's charging needs are met, using the margin index form: for each charging pile, the ratio of the total power provided by it during the forecast time period to the predicted value of charging demand is calculated, the natural logarithm is taken after adding one, and the logarithm of all charging piles is accumulated. The goal is to maximize this sum.

[0030] S2-4, solution and output: During the solution phase, an improved particle swarm optimization algorithm is employed, enhancing global search capabilities by introducing a dynamic inertia factor that decreases linearly with the number of iterations. The optimization process involves the following steps: First, a set of initial solutions that satisfy all constraints is randomly generated, namely, baseline power allocation values ​​for each charging pile at each time point. Next, multiple rounds of iterative optimization are performed, with the solution updated after each iteration and constraints verified. After the optimization is complete, a set of non-inferior solutions (the Pareto frontier) is obtained, from which the solution with the optimal comprehensive metric is selected as the final solution. The comprehensive metric is the weighted sum of two objective functions, with the weights determined based on the priority of grid dispatch requirements and user needs. The final optimal solution is the baseline power allocation value for each charging pile at each time point, which is used in subsequent steps.

[0031] Using this multi-objective optimization approach, the optimization scheduling module combines the charging demand and charging capacity forecasts from the previous step with real-time grid load data to generate baseline power allocation values ​​for each charging pile at each point in time. This result not only meets user charging needs but also optimizes the grid's peak-shaving and valley-filling effects, providing accurate data support for subsequent real-time adjustments.

[0032] In the smart charging pile remote control and scheduling method, the previous step uses a multi-objective optimization algorithm to generate a baseline power allocation value for each charging pile at each time point. This power allocation baseline value not only meets user charging needs but also optimizes the peak-shaving and valley-filling effect of the power grid. Building on this, the dynamic game module uses a reinforcement learning algorithm to construct a dynamic game model between users and the grid, adjusting the output power of the charging piles in real time. By calculating the user behavior deviation index and the grid load relief index, and integrating the synergistic effects of the two, it dynamically adjusts the point incentive strategy to guide user behavior deviation, thereby achieving a dynamic balance between users' actual charging needs and the grid load target.

[0033] The dynamic game module includes the following: S3-1, the user behavior deviation index is used to measure the degree of deviation between the actual charging behavior of each vehicle at a specific point in time and the predicted behavior. First, the actual charging power of each vehicle at that point in time is monitored in real time, and the predicted charging power is obtained from the user behavior prediction model. Then, the relative deviation between the actual charging power and the predicted charging power is calculated. The specific method is: divide the absolute value of the difference between the two by the sum of the predicted charging power and a small constant to avoid the denominator being zero. Next, in order to reduce the interference of short-term fluctuations, a time window is introduced to calculate the cumulative average of the offset within the window. Finally, the cumulative offset is converted into a user behavior deviation index through nonlinear mapping. The mapping method is: 1 minus the negative value with the natural constant as the base and the product of the offset and a control parameter as the exponent. The value range of this index is limited to between 0 and 1. The larger the value, the more significant the user behavior deviation. For example, the acquisition logic of the user behavior deviation index can be:

[0034] The user behavior deviation information includes the user behavior deviation index. To measure the Car in time The calculation process is as follows: First, real-time monitoring of the first Car in time Actual charging power and obtain the predicted charging power from the user behavior prediction model of the federated learning module Then, the relative deviation between the actual and predicted charging power is calculated using the formula:

[0035]

[0036] in, To prevent the denominator from being zero, a small constant is used. Then, to reduce the interference of short-term fluctuations, a time window is introduced. , calculate the cumulative average of the offset within the window:

[0037]

[0038] Finally, the cumulative offset is converted into the user behavior offset index through nonlinear mapping:

[0039] in, The exponent is a parameter that controls the sensitivity of the offset. The exponent ranges from [0, 1], with larger values ​​indicating more significant user behavior offsets. This method dynamically captures subtle changes in user behavior through time window smoothing and exponential transformation.

[0040] S3-2, the grid load relief index is used to measure the relief effect of the current charging pile power adjustment on the grid load target. First, the total load of the grid at that time point is monitored in real time, and the grid load target is obtained from the previous step. Then, the difference between the actual load and the target load is calculated. Next, a historical time window is introduced to calculate the average value of the grid load relief trend within the window, specifically: the difference between the target load and the actual load at each historical time point is divided by the target load, and then the average is calculated. Finally, the Sigmoid function is used to convert the average value into a grid load relief index. The parameters of the Sigmoid function include the steepness of a control curve and a relief threshold. The index ranges from 0 to 1. The larger the value, the better the relief effect. For example, the acquisition logic of the grid load relief index can be:

[0041] The grid load relief information includes the user behavior deviation index. It is used to measure the effect of current charging pile power adjustment on the grid load target. The calculation process is as follows: First, the real-time monitoring time Total grid load and obtain the grid load target from the optimization scheduling module Then, calculate the difference between the actual load and the target load:

[0042]

[0043] Next, the load relief factor is introduced , assess mitigation trends through historical data:

[0044] in, Is a time variable, indicating the time from the current point in time Moving backwards to a historical point in time, is the historical time window. Finally, the Sigmoid function is used to quantify the mitigation effect:

[0045]

[0046] in, Controls the steepness of the curve, is the mitigation threshold, and the index value range is [0,1]. The larger the value, the better the mitigation effect. This method combines historical trends and nonlinear mapping to improve the index's ability to respond to dynamic changes in the power grid.

[0047] S3-3, first, calculate the average value of the user behavior deviation index of all vehicles at that time point. Then, use the geometric mean method to combine the average value with the inverse value of the grid load relief index, that is, 1 minus the grid load relief index, to obtain the user-grid response synergy index. For example, the calculation of the user-grid response synergy index is:

[0048] User-grid response coordination index It is used to comprehensively evaluate the synergistic effect of user behavior deviation and grid load relief. The calculation process is as follows: First, calculate the time of all vehicles Average user behavior deviation index:

[0049]

[0050] in, is the total number of charging vehicles at present. Then, the geometric mean is used to integrate user behavior deviation and grid load relief:

[0051]

[0052] The geometric mean emphasizes the balance between the two. The smaller the value, the better the synergy effect. This method highlights the complementary relationship between user behavior and grid goals through geometric form.

[0053] S3-4 adopts a multi-agent reinforcement learning framework, treating each charging pile as an intelligent agent and the power grid as part of the environment. The state space includes the grid load at the current time point, the power distribution baseline value, the actual output power of each charging pile, and the user-grid response synergy index. The action space is the action amount, which is the power increase or decrease relative to the power distribution baseline value, and the amplitude is limited by the adjustable range of a single pile. The reward function is designed to be the inverse of the weighted sum of the user-grid response synergy index and the grid load deviation, and the weight is used to balance the importance of the two. The training adopts the deep Q network algorithm, and the learning stability is improved through the experience replay mechanism. For example, the reinforcement learning dynamic game model can be constructed as:

[0054] Using the multi-agent reinforcement learning (MARL) framework, each charging pile As an intelligent agent, the power grid is part of the environment. The state space includes time Grid load , power allocation baseline value , the actual output power of each charging pile and user-grid response synergy index The reward function is designed as:

[0055]

[0056] in, and The training adopts the Deep Q Network (DQN) algorithm, and the experience replay mechanism is used to improve learning stability.

[0057] S3-5: If the actual output power continues to deviate from the power allocation baseline value and the user-grid response synergy index exceeds the preset threshold, it indicates that there is a significant conflict between the user behavior and the grid load target, and it is necessary to use point incentives to guide users to adjust their charging behavior. The incentive intensity is dynamically adjusted based on the difference between the synergy index and the threshold. The specific method is: the basic incentive intensity is multiplied by an amplification factor based on the difference. The incentive is in the form of point rewards or penalties to encourage users to charge during off-peak hours of the grid. Dynamic adjustment of point incentives:

[0058] Incentive intensity Calculated as:

[0059] in, The incentive intensity is the basic incentive. Incentives take the form of point rewards or penalties, encouraging users to charge during off-peak hours. This method dynamically adjusts the incentive intensity through the synergy index to ensure targeted guidance.

[0060] The dynamic game model uses a reinforcement learning algorithm to build a dynamic game model between users and the grid. Based on the power allocation baseline value from the previous step, it calculates the user behavior deviation index and the grid load relief index, and then uses geometric averaging to obtain the user-grid response synergy index. Based on the comparison of this index with a preset threshold, the integration incentive strategy intensity is dynamically adjusted, and the charging pile output power is updated in real time. This achieves a dynamic balance between user demand and grid load, improving the adaptability and efficiency of the smart charging pile scheduling method.

[0061] The dynamic game module uses a reinforcement learning algorithm to construct a dynamic game model between users and the grid. It calculates the user-grid response synergy index by combining the user behavior deviation index and the grid load relief index. Based on the comparison of this index with a preset threshold, it dynamically adjusts the intensity of the point incentive strategy. Simultaneously, it adjusts the output power of charging piles in real time based on the power allocation baseline value generated by the optimization scheduling module to balance user charging demand with grid load targets. Building on this, the simulation verification module uses digital twin technology to conduct a full-link simulation verification of the charging pile cluster. The output power of the charging piles adjusted by the dynamic game module is compared with actual operating data, and optimization prompts are issued based on the comparison results to achieve continuous optimization and closed-loop control of the scheduling strategy.

[0062] The simulation verification module includes the following: S4-1, building a digital twin model: When building the digital twin model, a deep learning algorithm is first used to construct a digital twin model of the charging pile fleet using collected data on vehicle arrival time deviations, battery status parameters, and real-time grid load. The deep learning algorithm analyzes historical data and real-time inputs to predict the operational trends of the charging pile fleet. The model inputs include adjusted charging pile output power, battery status parameters, user charging demand forecasts, and grid load data. The output is a simulated charging pile operating status and grid load changes. During the processing, multi-source data fusion technology is used to integrate charging pile sensor data, user behavior data, and grid monitoring data into a unified input dataset to improve the accuracy of the model simulation.

[0063] S4-2, simulation verification: During the simulation verification phase, the constructed digital twin model was used to conduct a full-link simulation of the adjusted charging pile output power, predicting the operational status of the charging pile fleet and the grid load response over future time periods. To ensure that the scheduling strategy can adapt to various scenarios, random perturbations were introduced into the simulation to simulate the unpredictability of user behavior and the dynamic fluctuations of grid load. The simulation process generated simulated charging pile output power and simulated grid load through model calculations, which served as the basis for subsequent comparisons.

[0064] S4-3, data comparison: During the data comparison phase, actual operating data from the charging pile fleet is first collected in real time, including the actual output power of each charging pile and the actual load on the power grid. Then, a dynamic time warping algorithm is used to align the time series of the simulation results with the actual data, and the deviation between the two is calculated. The dynamic time warping algorithm analyzes the dynamic changes in the time series to find the optimal matching path between the simulation data and the actual data, quantifying the difference between the two in the time dimension and generating a deviation result. This process avoids comparison errors caused by time misalignment and ensures the accuracy of the deviation calculation.

[0065] S4-4, optimization tips: During the optimization prompt phase, power deviation thresholds and load deviation thresholds are pre-set as judgment criteria. When the power deviation or grid load deviation of any charging station exceeds the corresponding threshold, an optimization prompt is triggered. The generation of the optimization prompt is based on deviation analysis: if the charging station power deviation exceeds the threshold, the relevant parameters of the user behavior prediction model or battery charging capacity model are adjusted; if the grid load deviation exceeds the threshold, the constraints of the multi-objective optimization algorithm or the calculation logic of the reinforcement learning reward function are optimized. In addition, a Bayesian optimization algorithm is used to automatically search for the optimal parameter adjustment solution based on the deviation analysis results. The optimal parameter combination is determined through repeated iterative calculations, achieving adaptive optimization of the scheduling strategy.

[0066] The simulation verification module uses digital twin technology to perform full-link simulation verification of the charging pile fleet. It generates simulation results using adjusted charging pile output power and compares them with actual operating data. A dynamic time warping algorithm quantifies the deviation between charging pile output power and grid load. When the deviation exceeds a preset threshold, specific optimization prompts are generated for the preceding steps, and parameter adjustments are implemented using a Bayesian optimization algorithm. This process improves the accuracy of the smart charging pile scheduling strategy and the grid load balancing capability, forming a closed-loop control mechanism.

[0067] The above formulas are all dimensionless and numerical calculations. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters in the formulas are set by technicians in this field according to actual conditions.

[0068] It should be noted that the system of the present invention can be deployed on the device itself to realize embedded applications, and can also be run on a PC or other terminal with a user interface, thereby meeting a variety of hardware environments and usage requirements.

[0069] The above description is merely illustrative of certain exemplary embodiments of the present invention. It goes without saying that those skilled in the art will be able to modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the above drawings and description are illustrative in nature and should not be construed as limiting the scope of protection of the claims.

[0070] It should be noted that, in this document, if there are relational terms such as first and second, etc., they are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "comprises", "includes" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device that includes a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article or device. In the absence of further limitations, an element defined by the phrase "comprises a ..." does not exclude the presence of other identical elements in the process, method, article or device that includes the element.

[0071] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

Claims

1. Multifunctional slot intelligent charging pile remote control and dispatching system, characterized by: include: Federated Learning Module: This module collects vehicle arrival time deviations and battery status parameters in real time, and builds user behavior prediction models and battery charging capacity models based on federated learning. Optimization and scheduling module: This module combines the output of the user behavior prediction model with real-time grid load data to generate a power allocation baseline value through a multi-objective optimization algorithm. Dynamic Game Module: Based on a reinforcement learning algorithm, this module builds a dynamic game model between users and the grid. It integrates user behavior deviation information and grid load relief information, adjusts the output power of charging piles in real time according to the power allocation baseline value, and drives user charging behavior toward low-load periods based on a point incentive strategy. Simulation verification module: Use digital twin technology to perform full-link simulation verification of the charging pile group, compare simulation data with actual operation data, and output optimization prompt signals.

2. The multifunctional slot intelligent charging pile remote control and dispatching system according to claim 1 is characterized by: The federated learning module collects vehicle arrival time deviations and battery status information at charging piles in real time and performs preprocessing. Based on federated learning technology, a local model is trained on each charging pile node and aggregated through a central server to generate a global model.

3. The multifunctional slot intelligent charging pile remote control and dispatching system according to claim 2 is characterized by: The federated learning module uses support vector regression to build a battery charging capacity model to predict the vehicle's charging capacity; finally, the trained model is deployed to each charging pile node to generate charging demand prediction values ​​and charging capacity prediction values ​​in real time.

4. The multifunctional slot intelligent charging pile remote control and dispatching system according to claim 3 is characterized by: The optimization scheduling module obtains the charging demand forecast value and the charging capacity forecast value, and collects the grid load data and the grid load target curve in real time; smoothes the real-time grid load and calculates the grid load deviation; constructs a multi-objective optimization model, using the power allocation baseline value of each charging pile at each time point as the decision variable, and simultaneously optimizes the grid peak shaving and valley filling and the satisfaction of user charging needs; sets constraints to ensure that the power allocation baseline value is within the capacity of the charging pile, meets the charging demand and controls the grid load deviation; designs the objective function, uses the exponential decay penalty form to evaluate the grid peak shaving and valley filling effect, and uses the margin index form to measure the satisfaction of user charging needs; uses the improved particle swarm optimization algorithm to obtain the optimal solution; and finally generates the power allocation baseline value of each charging pile at each time point.

5. The multifunctional slot intelligent charging pile remote control and dispatching system according to claim 1 is characterized by: The user behavior deviation information includes the user behavior deviation index. The user behavior deviation index is obtained by calculating the relative deviation between the actual charging power and the predicted charging power within the time window and performing nonlinear mapping. The value range is 0 to 1.

6. The multifunctional slot intelligent charging pile remote control and dispatching system according to claim 5 is characterized by: The grid load relief information includes the grid load relief index. The grid load relief index is obtained by evaluating the average ratio of the difference between the real-time load and the load target in the historical window and converting it through the Sigmoid function. The value range is 0 to 1.

7. The multifunctional slot intelligent charging pile remote control and dispatching system according to claim 6 is characterized by: The dynamic game module calculates the average value of the user behavior deviation index of all vehicles at the selected time point; the geometric mean method is used to integrate the average value and the inverse value of the grid load relief index to obtain the user-grid response synergy index.

8. The multifunctional slot intelligent charging pile remote control and dispatching system according to claim 7 is characterized by: The dynamic game module uses a multi-agent reinforcement learning framework, treating each charging station as an agent and the power grid as part of the environment. The state space includes the grid load at the current time point, the power allocation baseline value, the actual output power of each charging station, and the user-grid response coordination index. The action space is the power increase or decrease relative to the power allocation baseline value, which is limited by the adjustable range of a single pile. The reward function is designed as the inverse of the weighted sum of the user-grid response synergy index and the grid load deviation. The training adopts the deep Q network algorithm, and the experience replay mechanism is used to improve learning stability.

9. The multifunctional slot intelligent charging pile remote control and dispatching system according to claim 8 is characterized by: If the actual output power continues to deviate from the power allocation baseline value and the user-grid response synergy index exceeds the preset threshold, the dynamic game module guides users to adjust their charging behavior through point incentives.

10. The multifunctional slot intelligent charging pile remote control and dispatching system according to claim 1 is characterized by: The simulation verification module uses a dynamic time warping algorithm to align the simulation data with the actual data. If the power deviation or load deviation exceeds the corresponding preset threshold, an optimization prompt signal is generated.

Images