Intelligent scheduling system and method for high-efficiency charging platform

Abstract

The invention relates to the technical field of computers, and discloses an intelligent scheduling system and method for a high-efficiency charging platform, and the method comprises the steps: constructing a four-dimensional constraint model integrating the power grid load, the user demand, the equipment health degree and electricity price prediction, and employing a depth deterministic strategy gradient algorithm to drive a multi-target dynamic scheduling decision maker, the global optimization distribution of the charging resources in the space-time power dimension is realized, and strategy reconstruction is completed within 10s when a power grid emergency instruction or a device fault occurs. The system comprises a power grid sensing module, a user acquisition module, a health assessment module, an electricity price response module, a scheduling decision module, an instruction execution module and an emergency reconstruction module, and supports millisecond-level adaptive evolution. According to the method, the weighted reward function is constructed by quantifying the four indexes of power grid stability, user satisfaction, equipment loss and platform income, and the instruction verification and steady-state adaptive mechanism is combined, so that the user experience is synchronously improved, the service life of the equipment is prolonged, and the operation energy consumption is reduced on the premise of ensuring the safety.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of computers, and particularly relates to an intelligent scheduling system and method for a high-efficiency charging platform. BACKGROUND

[0002] With the rapid development of the electric vehicle industry and the large-scale deployment of charging infrastructure, the scheduling management of the charging platform is evolving from single-device control to multi-objective collaborative optimization. Current mainstream charging scheduling systems are mostly based on static rules or single-objective optimization models, and their core logic revolves around minimizing user waiting time or maximizing operator revenue, relying on preset thresholds and fixed strategies for task allocation. However, the charging scenario is essentially a highly dynamic and multi-dimensional coupled complex system: user behavior is random and diverse, power grid load shows peak-valley fluctuations and regional differences, charging pile devices have individual differences in aging rate and load tolerance, and external environment such as electricity price policy, V2G instruction, device failure, etc. sudden variables further exacerbate the uncertainty of scheduling decisions. The single-target-oriented scheduling mechanism is easily trapped in a local optimal trap when facing the above multi-dimensional variables, leading to intensified power grid impact, device overload loss, and decreased user satisfaction, etc. chain negative effects, which is difficult to support the operation needs of the new generation of charging network with high density, high concurrency, and high reliability.

[0003] Among them, the intelligent scheduling system of the high-efficiency charging platform is the core hub connecting user demand, power grid constraints, and device capacity, and its core task is to realize the optimal containerization allocation of charging requests in a dynamic environment. The system needs to comprehensively evaluate task priority, resource availability, cost constraints, and system stability within a millisecond response window, complete the end-to-end closed-loop scheduling from request access to physical execution. However, the existing scheduling architecture generally uses centralized rule engines or heuristic algorithms, and its decision logic is fixed in the code layer, lacking the ability to perceive the state of the environment and the mechanism of strategy self-evolution. When encountering real-time price jumps, regional power grid emergency power cuts, or sudden offline of key charging piles, etc. non-steady-state scenarios, the system often cannot dynamically reconstruct the scheduling target weight, nor can it autonomously learn the mapping relationship between historical decisions and environmental feedback, ultimately leading to rigid scheduling strategies, response lag, and global performance degradation.

[0004] The prior art has structural defects in multi-target cooperation and dynamic adaptability: firstly, the scheduling target dimension is fragmented, and there is a lack of ability to uniformly model heterogeneous indicators such as user waiting cost, power grid regulation pressure, equipment balanced loss, and operator marginal benefit, resulting in a compromise in the optimization process; secondly, the algorithm architecture is closed, and the charging task cannot be packaged as a migratable, measurable, and traceable independent computing unit, resulting in coarse scheduling granularity and low resource utilization; thirdly, the decision-making mechanism lacks online learning ability, and in the face of high-frequency disturbances such as price fluctuations, V2G instruction insertion, and device state mutations, it cannot iteratively update strategy parameters through continuous interaction with the environment, making it difficult to approach the Pareto optimal solution under complex constraints. The above defects are dramatically amplified during high-concurrency charging peak periods or in regional power grid vulnerability scenarios, forming a technical dilemma of scheduling efficiency collapse and system stability risk coexistence, which urgently needs a next-generation intelligent scheduling system with containerized task abstraction ability, multi-target constraint modeling ability, and reinforcement learning self-optimization ability to break through. SUMMARY

[0005] The present application provides an intelligent scheduling system and method for a high-performance charging platform, which builds a multi-target dynamic trade-off decision-making architecture, integrates four-dimensional constraint conditions of real-time power grid load state, user individualized charging demand, charging pile health degree evaluation, and price fluctuation prediction, establishes a self-adaptive scheduling engine based on reinforcement learning, and realizes global optimal allocation of charging resources in time, space, and power dimensions. When facing sudden power grid instructions, device failures, or user behavior changes, the system can autonomously reconstruct the scheduling strategy to ensure that the user waiting experience is optimized, the device service life is extended, and the platform's overall operating income is improved while ensuring the safe and stable operation of the power grid.

[0006] The present application provides an intelligent scheduling method for a high-performance charging platform, which comprises:

[0007] The load curve, frequency deviation value, and voltage stability margin index of the current regional power grid are obtained in real time through the power grid data interface to build a power grid safe operation boundary constraint model;

[0008] The battery remaining capacity of the vehicle to be charged, the expected full charging time set by the user, the acceptable maximum charging cost threshold, and the permission identifier of whether to allow participation in vehicle-to-grid energy feedback are collected through the user terminal interaction interface;

[0009] The cumulative running time, historical failure times, current output power fluctuation amplitude, and radiator temperature value of each charging pile are collected through the charging pile state monitoring unit to calculate the real-time health score of each charging pile;

[0010] The time-of-use price prediction sequence within the next 2 hours is obtained through the power market data interface, and a price sensitivity response model is established in combination with the historical price fluctuation rule;

[0011] The grid safe operation boundary constraint model, user individualized charging demand parameters, real-time health score of charging piles and price sensitivity response model are jointly input into a multi-objective dynamic scheduling decision maker, the multi-objective dynamic scheduling decision maker is constructed by using a deep deterministic policy gradient algorithm, a state space of the multi-objective dynamic scheduling decision maker is constituted by a grid load rate, a user waiting queue length, a charging pile health degree distribution and a price gradient, an action space of the multi-objective dynamic scheduling decision maker is a matching relationship between each vehicle to be allocated and an available charging pile and a corresponding power output gear, and a reward function of the multi-objective dynamic scheduling decision maker is constituted by weighted summation of a grid stability gain term, a user satisfaction improvement term, a device loss suppression term and a platform revenue growth term;

[0012] The multi-objective dynamic scheduling decision maker outputs optimal charging pile allocation instructions and corresponding power setting values, and sends the optimal charging pile allocation instructions and the corresponding power setting values to a charging pile control unit for execution.

[0013] When an emergency vehicle-to-grid instruction is issued by a grid dispatching center or any charging pile has a hardware-level fault, a strategy reset mechanism is triggered immediately, all current non-emergency charging tasks are frozen, a global scheduling scheme satisfying new constraint conditions is recalculated, and the instruction issuing and execution unit switching is completed within 10s.

[0014] According to another aspect of the present application, an intelligent scheduling system of a high-efficiency charging platform is provided, which comprises:

[0015] A grid state perception module is configured to acquire a load curve, a frequency deviation value and a voltage stability margin index of a current regional grid in real time through a grid data interface, and construct a grid safe operation boundary constraint model.

[0016] A user demand acquisition module is configured to acquire, through a user terminal interactive interface, a remaining capacity of a battery of a vehicle to be charged, an expected full-charge time set by a user, an acceptable maximum charging cost threshold and a permission identification of whether to allow participation in vehicle-to-grid energy feedback.

[0017] A device health evaluation module is configured to acquire, through a charging pile state monitoring unit, an accumulated operation time length, a historical fault frequency, a current output power fluctuation amplitude and a radiator temperature value of each charging pile, and calculate a real-time health score of each charging pile.

[0018] A price prediction response module is configured to acquire, through a power market data interface, a time-of-use price prediction sequence within the next 2h, and establish a price sensitivity response model in combination with a historical price fluctuation rule.

[0019] A multi-objective dynamic scheduling decision maker is configured to receive the power grid safe operation boundary constraint model, user individualized charging demand parameters, real-time health score of charging piles, and price sensitivity response model, and is configured to adopt a deep deterministic policy gradient algorithm to construct a state space composed of power grid load rate, user waiting queue length, charging pile health degree distribution, and price gradient, an action space composed of matching relationship between each vehicle to be allocated and available charging pile and corresponding power output gear, and a reward function composed of weighted sum of power grid stability gain term, user satisfaction improvement term, device loss suppression term, and platform revenue growth term;

[0020] An instruction execution and feedback module is configured to receive optimal charging pile allocation instructions and corresponding power setting values output by the multi-objective dynamic scheduling decision maker, send the instructions to a charging pile control unit for execution, and return execution state and deviation information in real time.

[0021] An emergency strategy reconstruction module is configured to trigger a strategy reset mechanism immediately when an emergency vehicle-to-grid instruction is issued by a power grid dispatching center or any charging pile has a hardware-level fault, freeze all current non-emergency charging tasks, recalculate a global scheduling scheme that meets new constraint conditions, and complete instruction issuance and execution unit switching within 10 seconds.

[0022] The power grid safe operation boundary constraint model is specifically constructed as follows: 80% of the rated capacity of a regional transformer is taken as the upper limit threshold of power, an absolute value of frequency deviation greater than 0.2 Hz is taken as the emergency regulation trigger condition, and a voltage fluctuation amplitude exceeding the nominal value by ±5% is taken as the power regulation prohibited interval, all of which together constitute a three-dimensional safe operation hyperplane, and any scheduling instruction must ensure that the system state point is located within the safe region surrounded by the hyperplane after execution.

[0023] The user satisfaction improvement term is calculated as follows: for each user, a negative exponential function of the difference between actual waiting time and historical average waiting time is taken as the basic satisfaction value, the product of the charging cost saving ratio and the expected full charge time achievement rate is superimposed, and then multiplied by the user level weight coefficient, and finally the arithmetic mean value of all users is taken to obtain the global user satisfaction index.

[0024] The device loss suppression term is calculated as follows: for each charging pile, the square of the difference between its current health score and the ideal health score benchmark value is taken as the loss penalty factor, multiplied by the total charging time of the charging pile allocated in this round of scheduling, and then summed for all charging piles to obtain the negative value of the device loss suppression term.

[0025] The platform revenue growth item is calculated as follows: for each charging order, the actual fee collected is subtracted from the grid purchase cost in the period to obtain the gross profit of each order. After accumulating the gross profits of all orders, the energy feedback compensation paid to users for participating in the vehicle-to-grid response is deducted to obtain the net platform revenue value.

[0026] The neural network structure in the deep deterministic policy gradient algorithm includes three fully connected hidden layers, each layer having 256, 512, and 256 neurons, respectively, and using a rectified linear unit as the activation function. The output layer uses a hyperbolic tangent function to compress the action value into a preset power gear interval. The experience replay buffer capacity is set to 100,000 state-action-reward-next state four-tuple. Each time of training randomly samples 2560 groups of samples from the buffer for gradient descent update. The initial learning rate is set to 0.001, and is attenuated by 10% after completing 5000 iterations.

[0027] The emergency strategy reconstruction module prioritizes vehicles with a battery state of charge higher than 80% and user authorization to participate in energy feedback, sorts them in descending order of dischargeable power capacity, and allocates them to the charging pile closest to the substation and with a health score higher than 0.8 to perform reverse discharge operation, with the upper limit of discharge power set to 70% of the rated power of the charging pile and the discharge duration less than 30 minutes.

[0028] The emergency strategy reconstruction module removes the charging pile from the available resource pool upon detecting a hardware-level failure of the charging pile, sorts all vehicles being served by the charging pile in ascending order of remaining charge, and reallocates them to the idle charging pile with the highest health score within a 3km range. The new allocation scheme ensures that the estimated completion time of each vehicle does not exceed 120% of the original scheme.

[0029] The instruction execution and feedback module includes an instruction verification submodule that verifies whether the instruction violates the grid safety operation boundary constraint model, exceeds the user-set fee threshold, or is allocated to a charging pile with a health score lower than 0.5 before issuing the dispatch instruction. If any of the violation conditions exist, the instruction is returned to the multi-objective dynamic scheduling decision maker for local modification, with a maximum of three modification times. If the constraints cannot be met after three modifications, the user task is marked as unschedulable and the user is notified to reset the parameters.

[0030] The multi-objective dynamic scheduling decision maker performs global strategy update every 30 seconds, and introduces 5% random exploration actions to avoid falling into a local optimal solution. In a non-emergency state, if the reward function value fluctuation range of 10 consecutive strategy updates is less than 1‰, the system is determined to be in a steady state, and the update interval is automatically extended to two minutes to reduce the consumption of computing resources.

[0031] The electricity price sensitivity response model is constructed by adopting a long short-term memory network, input features include a past 24h time-of-use electricity price sequence, a daily weather forecast temperature value, a holiday identifier and a regional electric vehicle ownership growth rate, and output is a predicted electricity price value and a confidence interval every 15 minutes within the next 2h, when the predicted electricity price exceeds a user set threshold value, the period is automatically marked as a low priority charging window, and the power allocation weight of a non-emergency charging task in the period is reduced in the scheduling decision.

[0032] The charging pile state monitoring unit includes a current transformer, a voltage sensor, a temperature sensor and a vibration sensor, the sampling frequency is 100 times per second, and the original data is input to the equipment health score calculation model after sliding average filtering; the equipment health score calculation model adopts a support vector regression algorithm, input features are average output power fluctuation rate, maximum radiator temperature, cumulative vibration energy value and fault interval days of the charging pile within the past 30 days, and output is a continuous score value between 0 and 1, and the charging pile with a score lower than 0.3 is prohibited from participating in any scheduling task.

[0033] Compared with the prior art, the advantages and positive effects of the present application are that:

[0034] 1. By constructing a four-dimensional constraint model integrating power grid safety, user demand, equipment health and electricity price response, the problems of power grid impact risk and equipment overload caused by single target of traditional scheduling system are completely solved;

[0035] 2. Deep reinforcement learning architecture is adopted to realize online adaptive evolution of scheduling strategy, so that the system has millisecond-level strategy reconstruction capability when facing price mutation, emergency instruction or equipment failure, and service continuity and system robustness are ensured;

[0036] 3. By quantitatively defining four core indexes of power grid stability gain, user satisfaction improvement, equipment loss suppression and platform revenue growth and constructing a weighted reward function, dynamic balance and collaborative optimization among multiple targets are realized;

[0037] 4. The instruction checking and local correction mechanism is introduced to ensure that each scheduling instruction meets all the hard constraint conditions and eliminates the risk of illegal operation;

[0038] 5. The steady-state detection and adaptive adjustment of computing resources function is set to reduce system operation energy consumption and hardware cost on the premise of ensuring scheduling accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 is the overall technical scheme architecture schematic diagram of the intelligent scheduling system and method of the high-efficiency charging platform proposed by the present application;

[0040] Figure 2is the core principle framework diagram of the multi-objective dynamic scheduling decision maker based on reinforcement learning in the application;

[0041] Figure 3 is the logical flow framework diagram of the fusion modeling of the power grid safety boundary constraint and the four-dimensional scheduling factor in the application;

[0042] Figure 4 is the logical flow framework diagram of user demand collection and satisfaction quantitative evaluation in the application;

[0043] Figure 5 is the logical flow framework diagram of charging pile health degree evaluation and emergency resource redistribution in the application;

[0044] Figure 6 is the multi-level interaction relationship and data flow diagram of the price prediction response model and the dynamic optimization of the scheduling strategy in the application; DETAILED DESCRIPTION

[0045] Please refer to Figures 1-6 The application provides an intelligent scheduling system and method of a high-performance charging platform, and the core is to construct a multi-objective dynamic trade-off decision architecture. The architecture fuses four-dimensional constraint conditions of real-time power grid load state, user individualized charging demand, charging pile health degree evaluation and price fluctuation prediction, establishes an adaptive scheduling engine based on reinforcement learning, and realizes global optimal allocation of charging resources in time, space and power dimensions. When facing sudden power grid instructions, equipment failures or user behavior changes, the system can autonomously reconfigure the scheduling strategy to ensure that the user waiting experience is optimized, the equipment service life is extended and the platform comprehensive operation income is improved on the premise of ensuring the safe and stable operation of the power grid. The specific embodiments of the application will be described in detail below.

[0046] The intelligent scheduling method of the high-performance charging platform strictly follows the preset step S1 to S7 sequence in the execution process.

[0047] S1, the load curve, frequency deviation value and voltage stability margin index of the current regional power grid are obtained in real time through the power grid data interface, and a power grid safe operation boundary constraint model is constructed.

[0048] The specific construction method of the model is as follows:

[0049] 80% of the rated capacity of the regional transformer is taken as the upper threshold of the power, the absolute value of the frequency deviation is greater than 0.2Hz as the emergency regulation trigger condition, and the voltage fluctuation amplitude exceeds the nominal value ±5% as the power regulation prohibited interval. The three together constitute a three-dimensional safe operation hyperplane, and any scheduling instruction must ensure that the system state point is located in the safe region surrounded by the hyperplane after execution.

[0050] The grid data interface polls the data server of the grid dispatch center every five seconds to obtain the latest grid operation parameters.

[0051] The obtained data is first subjected to a data cleaning module to eliminate abnormal values and missing values, and then sent to a boundary constraint model calculation unit.

[0052] The calculation unit converts the original data into a three-dimensional vector according to the preset physical rules, and the three components of the vector respectively represent the safety distance of the current grid state with respect to the power upper limit, frequency stability and voltage stability.

[0053] This three-dimensional vector is updated in real time and stored in shared memory for subsequent dispatch decision maker calls.

[0054] S2, through the user terminal interactive interface, the battery remaining capacity of the vehicle to be charged, the user set expected full charge time, the acceptable maximum charging cost threshold and the permission identifier whether to allow participation in vehicle-to-grid energy feedback are collected.

[0055] The user terminal interactive interface is deployed in a mobile application or a vehicle information entertainment system. When the user initiates a charging request, the above four parameters must be filled in or confirmed.

[0056] The battery remaining capacity is reported by the vehicle's own battery management system through a standard communication protocol to ensure data accuracy. The expected full charge time is manually input by the user. The system will recommend an economic and efficient default time according to the current grid load and

[0057] The electricity price prediction recommends an economic and efficient default time to the user, but the user has the final decision.

[0058] The acceptable maximum charging cost threshold is also set by the user. The system will provide a reference value based on historical data and current prediction. The permission identifier whether to allow participation in vehicle-to-grid energy feedback is a binary switch, which the user can freely choose to turn on or off.

[0059] All collected user data is packaged into a structured data packet, including the user's unique identifier, vehicle identification code, four core parameters and timestamp, and then transmitted to the central dispatch server through an encrypted channel.

[0060] After the server receives the data packet, it decrypts and formats the data packet. The qualified data packet is stored in the user demand database and waits for the dispatch engine to process.

[0061] S3, through the charging pile state monitoring unit, the cumulative running time, historical fault times, current output power fluctuation amplitude and radiator temperature value of each charging pile are collected to calculate the real-time health score of each charging pile.

[0062] The charging pile state monitoring unit includes a current transformer, a voltage sensor, a temperature sensor, and a vibration sensor, and the sampling frequency is 100 times per second.

[0063] The original data is input to the device health score calculation model after being filtered by a moving average filter. The device health score calculation model uses a support vector regression algorithm, the input features are the average output power fluctuation rate of the charging pile in the past 30 days, the maximum radiator temperature, the cumulative vibration energy value, and the fault interval days, and the output is a continuous score value between 0 and 1. Charging piles with a score below 0.3 are prohibited from participating in any scheduling tasks.

[0064] The specific calculation process is as follows:

[0065] First, the original data collected by each sensor is preprocessed, including denoising, normalization, and feature extraction. For example, the power fluctuation rate is obtained by calculating the ratio of the standard deviation to the mean of the instantaneous power value; the maximum radiator temperature takes the peak value in the past hour; the cumulative vibration energy value is obtained by integrating the vibration acceleration signal; and the fault interval days are queried from the maintenance log database.

[0066] These features are combined into a four-dimensional feature vector, which is input into a pre-trained support vector regression model.

[0067] The model is trained on a large amount of historical data and can accurately map the feature vector to the health score. The calculated score is updated in real time to the charging pile state database and marked for availability.

[0068] The health score is not only used for scheduling decisions, but also for predictive maintenance. When the score is below 0.5 for three consecutive times, the system will automatically generate a maintenance work order.

[0069] S4, through the power market data interface, obtain the time-of-use price prediction sequence for the next 2h, and establish a price sensitivity response model combining the historical price fluctuation rules.

[0070] The price sensitivity response model is built using a long short-term memory network, and the input features include the time-of-use price sequence for the past 24h, the daily weather forecast temperature value, the holiday identifier, and the regional electric vehicle ownership growth rate. The output is the price prediction value and its confidence interval for every 15min interval in the next 2h.

[0071] When the predicted price exceeds the user's set threshold, automatically mark this period as a low-priority charging window, and reduce the power allocation weight of non-emergency charging tasks in this period in the scheduling decision.

[0072] The power market data interface obtains the latest price announcement and prediction data from the power trading center every 10 minutes.

[0073] The acquired data is fused with locally stored historical data, weather data, and holiday data to form a sequence dataset containing hundreds of time steps.

[0074] The dataset is fed into a long short-term memory network for processing.

[0075] The network structure contains two hidden layers, each with 128 memory units, which can effectively capture long-term dependencies in the price sequence.

[0076] The model outputs not only point estimates of future prices, but also a confidence interval to quantify the uncertainty of the prediction.

[0077] When formulating strategies, the scheduling engine takes into account both the predicted values and the confidence intervals. For periods of high uncertainty, more conservative scheduling strategies are adopted.

[0078] In addition, the model dynamically adjusts the priority of the user's charging window based on the prediction results, guiding users to charge during the price trough period, thereby reducing the overall charging cost.

[0079] S5, the power grid safe operation boundary constraint model, user individualized charging demand parameters, real-time health score of charging piles and price sensitivity response model are jointly input into a multi-objective dynamic scheduling decision maker.

[0080] The multi-objective dynamic scheduling decision maker is constructed using the deep deterministic policy gradient algorithm, whose state space is composed of grid load rate, user waiting queue length, charging pile health degree distribution and price gradient, its action space is the matching relationship between each vehicle to be allocated and available charging pile and the corresponding power output gear, and its reward function is composed of the weighted sum of grid stability gain term, user satisfaction improvement term, equipment loss suppression term and platform revenue growth term.

[0081] The neural network structure in the deep deterministic policy gradient algorithm contains three fully connected hidden layers, with 256, 512 and 256 neurons respectively in each layer. The activation function uses the rectified linear unit, and the output layer uses the hyperbolic tangent function to compress the action value into the preset power gear interval. The experience replay buffer capacity is set to 100,000 state-action-reward-next state four-tuple, 2560 samples are randomly sampled from the buffer for gradient descent update each time, the initial learning rate is set to 0.001, and the learning rate is attenuated by 10% every 5000 iterations.

[0082] The workflow of the scheduling decision maker is as follows:

[0083] First, read the latest four-dimensional constraint data from shared memory and database, and splice it into a high-dimensional state vector.

[0084] Then, the state vector is input into the policy network, which is a three-layer fully connected layer with nonlinear transformation, and an action vector is output.

[0085] The action vector is processed by the post-processing module and parsed into a specific charging pile allocation scheme and power setting value.

[0086] The post-processing module ensures that the allocation scheme meets all the hard constraints, such as not allocating vehicles to charging piles with low health scores, not exceeding the user's cost threshold, etc.

[0087] The parsed scheme is sent to the reward function calculation module to calculate the comprehensive reward value of this scheduling. The calculation of the reward function is one of the core innovations of the invention, and its specific formula is as follows:

[0088] Reward function R = a·R grid + β·R user + γ·R device + δ·R profit ;

[0089] Wherein, R grid is the power grid stability gain term, which is positively related to the reduction of the power grid load rate, the reduction of the frequency deviation, and the improvement of the voltage stability margin after scheduling;

[0090] R user is the user satisfaction improvement term, which is calculated as follows: for each user, the negative exponential function of the difference between the actual waiting time and the historical average waiting time is taken as the basic satisfaction value, the product of the charging cost saving ratio and the expected full charge time achievement rate is added, and then multiplied by the user level weight coefficient. Finally, the arithmetic mean of all users is taken to obtain the global user satisfaction index;

[0091] R device is the device loss suppression term, which is calculated as follows: for each charging pile, the square of the difference between its current health score and the ideal health score is taken as the loss penalty factor, multiplied by the total charging time of the charging pile in this round of scheduling, and then summed up to obtain the device loss suppression term;

[0092] R profit is the platform revenue growth term, which is calculated as follows: for each charging order, the actual charging fee minus the power grid purchase cost in this period is taken as the gross profit of each order, and the gross profit of all orders is accumulated to obtain the net platform revenue value after deducting the energy feedback compensation paid to users for participating in the vehicle-to-grid response.

[0093] The weight coefficients a, β, γ, δ are dynamically adjusted according to the macro target of system operation, for example, when the power grid is under pressure, a will be adjusted to a higher value to prioritize the stability of the power grid.

[0094] S6, the multi-objective dynamic scheduling decision maker outputs optimal charging pile allocation instructions and corresponding power set values, and sends them to the charging pile control unit for execution.

[0095] The instruction execution and feedback module includes an instruction verification submodule, which is used to verify whether the instruction violates the power grid safe operation boundary constraint model, exceeds the user-set cost threshold, or is allocated to a charging pile with a health score lower than 0.5 before the scheduling instruction is issued. If any of the violation conditions exists, the instruction is returned to the multi-objective dynamic scheduling decision maker for local correction. The upper limit of the number of correction times is three. If the constraint conditions cannot be met after three times of correction, the user task is marked as unschedulable and the user is notified to reset the parameters.

[0096] The instruction issuing process is asynchronous. After the scheduling decision maker generates an instruction, it is placed in a high-priority message queue. The instruction execution and feedback module takes the instruction from the queue and first performs verification.

[0097] The verification process includes three steps:

[0098] First, check whether the total power demand in the instruction exceeds the upper limit set by the power grid safe operation boundary constraint model.

[0099] Second, calculate the expected charging cost of each user to ensure that it does not exceed the user-set threshold.

[0100] Third, check the health score of the allocated charging pile to ensure that it is not lower than 0.5.

[0101] Only instructions that pass all three verifications will be sent to the corresponding charging pile control unit. After receiving the instruction, the control unit immediately adjusts the output power and sends an acknowledgment signal to the feedback module. After receiving the acknowledgment signal, the feedback module updates the system state and records the execution results in the log database. If the instruction fails the verification, the local correction mechanism is triggered. The correction mechanism locks the part of the instruction that violates the constraints, such as the power allocation of a certain user being too high, and then re-optimizes only that part while keeping the other parts unchanged. The corrected instruction is verified again, and this process is repeated a maximum of three times. If three times of correction fail, the system determines that the task cannot be completed under the current constraint conditions, sends a notification to the user, and suggests that they adjust the desired full charge time or cost threshold.

[0102] S7, when an emergency vehicle-to-grid instruction is issued by the power grid dispatching center or any charging pile experiences a hardware-level failure, the strategy reset mechanism is triggered immediately, all non-emergency charging tasks are frozen, a global scheduling scheme that meets the new constraint conditions is recalculated, and the instruction issuing and execution unit switching is completed within 10 seconds.

[0103] The emergency strategy reconstruction module prioritizes vehicles with battery state of charge higher than 80% and user authorization for energy feedback, sorts them by dischargeable power capacity from large to small, and assigns them to the charging pile closest to the substation and with a health score higher than 0.8, executes reverse discharge operation, sets the upper limit of discharge power to 70% of the charging pile rated power, and the discharge duration is less than 30 minutes.

[0104] The emergency strategy reconstruction module removes the charging pile from the available resource pool upon detecting a hardware-level failure, sorts all vehicles being served by the charging pile by remaining charge from small to large, and reassigns them to the idle charging pile with the highest health score within a 3km range, ensuring that the estimated completion time of each vehicle does not exceed 120% of the original scheme. The emergency strategy reconstruction module continuously monitors the system state through a separate monitoring thread.

[0105] When the monitoring thread captures the emergency instruction from the grid dispatch center or the failure alarm signal from the charging pile, it immediately sends an interrupt request to the main scheduling engine. Upon receiving the interrupt request, the main scheduling engine pauses the current scheduling period, saves the current state snapshot, and then starts the emergency scheduling subroutine.

[0106] In the vehicle-to-grid scenario, the subroutine first filters out all vehicles authorized to participate in energy feedback and with battery state of charge higher than 80% from the user database.

[0107] Then, according to the battery capacity of the vehicle and the health of the charging pile, the maximum dischargeable power of each vehicle is calculated.

[0108] Next, these vehicles are sorted by dischargeable power from large to small, and they are assigned to the charging pile closest to the substation and with the highest health score.

[0109] After the assignment is complete, the system sends discharge instructions to these vehicles and charging piles, and sets the upper limit of discharge power and duration.

[0110] In the charging pile failure scenario, the subroutine first removes the failed pile from the resource pool and obtains the list of all vehicles being served.

[0111] Then, according to the remaining charge of the vehicle, the vehicles are sorted by remaining charge from small to large, and the vehicle with the least remaining charge is prioritized for reassignment.

[0112] The system searches for the idle charging pile with the highest health score within a 3km range and calculates the estimated completion time after transferring the vehicle to the pile. Only when the estimated completion time does not exceed 120% of the original scheme, the assignment is confirmed.

[0113] Once the new allocation scheme is generated, the system immediately sends instructions to the relevant vehicles and charging stations to complete the switch.

[0114] The entire emergency reconfiguration process must be completed within 10 seconds to ensure the stability of the power grid and the continuity of user services.

[0115] The multi-objective dynamic scheduling decision-maker performs a global policy update every 30 seconds, introducing 5% random exploration during each update to avoid getting trapped in local optima. In non-urgent situations, if the reward function value fluctuates by less than 1‰ over 10 consecutive policy updates, the system is considered to have entered a steady state, and the update interval is automatically extended to two minutes to reduce computational resource consumption. This policy update mechanism is crucial for the system to maintain long-term optimal performance.

[0116] Every 30 seconds, the scheduling decision-maker randomly samples 2,560 sets of historical data from the experience replay buffer to train its internal neural network.

[0117] During training, in addition to deterministic actions based on historical best strategies, random actions are injected with a 5% probability to explore the unknown state space and prevent the algorithm from converging to a local optimum too early.

[0118] After training, the network parameters are updated, and a new policy is activated. The system also includes a steady-state detector that continuously monitors the reward function values ​​for the last 10 policy updates.

[0119] If the fluctuation range of these 10 values ​​is less than 1‰, it indicates that the system is already operating in a stable high-reward region and does not require frequent updates. At this point, the steady-state detector will send a signal to the scheduling engine to extend the update interval from 30 seconds to two minutes.

[0120] This mechanism reduces the system's computational load and energy consumption, especially at night or during off-peak hours when charging demand and grid conditions are relatively stable.

[0121] Once the system detects a drastic change in the external environment, such as a sudden surge in electricity prices or a large influx of users, causing increased fluctuations in the reward function value, the steady-state detector will immediately resume its 30-second update frequency to ensure the system can respond quickly to changes.

[0122] The intelligent scheduling system of the high-efficiency charging platform has a physical architecture composed of multiple functional modules working together.

[0123] The power grid status sensing module is deployed on a server cluster in the data center and is connected to the data interface of the power grid dispatch center through a high-speed fiber optic network.

[0124] This module includes a data acquisition agent, a data cleaning engine, and a boundary constraint model calculation unit.

[0125] The data collection agent is responsible for timing polling and pulling grid data; the data cleaning engine is responsible for processing noise and missing in data; and the boundary constraint model calculation unit is responsible for converting raw data into a three-dimensional safety vector.

[0126] The front end of the user demand collection module runs on the user's mobile device or vehicle-mounted system, and the back end is deployed on the application server in the cloud. The front end is responsible for providing a friendly interactive interface to guide the user to input necessary parameters; the back end is responsible for receiving, decrypting, verifying and storing user data. The core of the device health assessment module is the edge computing node distributed on each charging pile.

[0127] Each node is equipped with a microprocessor responsible for collecting sensor data, performing local filtering and preliminary feature extraction, and then uploading the processed data to the health score calculation server in the cloud through a wireless network.

[0128] The score calculation server runs a support vector regression model to generate real-time health scores for each charging pile.

[0129] The electricity price prediction response module is deployed on a separate prediction server, which has powerful computing power for running a long short-term memory network model.

[0130] It regularly obtains external data from the electricity market, weather bureau and traffic management department, and generates future electricity price predictions.

[0131] The multi-objective dynamic scheduling decision maker is the core of the entire system, which runs on a high-performance computing cluster equipped with multiple graphics processors to accelerate neural network training and inference.

[0132] The decision maker communicates with other modules through a high-speed internal bus, obtains real-time state data and outputs scheduling instructions.

[0133] The instruction execution and feedback module includes a distributed message middleware system for reliably delivering scheduling instructions and executing feedback.

[0134] It also includes an instruction verification sub-module to ensure the safety and compliance of all instructions. The emergency strategy reconstruction module runs independently on a backup server as a high-priority daemon process to ensure timely response to emergency events when the main system is busy or fails.

[0135] All modules communicate through a unified interface protocol, and the data format uses standardized structured data packets to ensure the scalability and maintainability of the system.

[0136] Compared with the prior art, the advantages and positive effects of the present application are:

[0137] 1. By constructing a four-dimensional constraint model that integrates power grid safety, user demand, equipment health and price response, the problem of power grid impact risk and equipment overload caused by the single objective of traditional dispatching systems is completely solved;

[0138] 2. The online adaptive evolution of the dispatching strategy is realized by using a deep reinforcement learning architecture, so that the system has the ability to reconstruct the strategy in milliseconds when facing price mutations, emergency instructions or equipment failures, thereby ensuring service continuity and system robustness;

[0139] 3. By quantitatively defining the four core indicators of power grid stability gain, user satisfaction improvement, equipment loss suppression and platform revenue growth and constructing a weighted reward function, dynamic balance and collaborative optimization among multiple objectives are realized;

[0140] 4. The instruction verification and local correction mechanism is introduced to ensure that every dispatching instruction meets all the hard constraint conditions and eliminates the risk of illegal operation;

[0141] 5. The steady-state detection and adaptive adjustment of computing resources function is set to reduce the system operation energy consumption and hardware cost while ensuring the accuracy of dispatching.

[0142] 6. The implementation of the present application not only improves the operating efficiency and economic benefits of the charging platform, but also provides strong technical support for the stable operation of the power grid and the popularization and promotion of electric vehicles.

[0143] It should be noted that in this paper, relationship terms such as first and second are used only 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 the entities or operations. Moreover, the term "comprises" "includes" or any other variant thereof is intended to cover non-exclusive inclusion, so that the process, method, article or equipment including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or equipment.

[0144] Although embodiments of the present application have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent scheduling of high-performance charging platforms, characterized in that, The method comprises the following steps: S1, acquiring the load curve, frequency deviation value and voltage stability margin index of the current regional power grid in real time through the power grid data interface, and constructing a power grid safe operation boundary constraint model; S2, collecting the battery remaining capacity of the vehicle to be charged, the expected full charging time set by the user, the acceptable maximum charging cost threshold and the permission identification of whether to allow participation in vehicle-to-grid energy feedback through the user terminal interaction interface; S3, collecting the cumulative running time, historical fault times, current output power fluctuation amplitude and radiator temperature value of each charging pile through the charging pile state monitoring unit, and calculating the real-time health score of each charging pile; S4, acquiring the time-of-use electricity price prediction sequence within the next 2 hours through the power market data interface, and establishing a price sensitivity response model combined with the historical price fluctuation rule; S5, inputting the power grid safe operation boundary constraint model, user individualized charging demand parameters, charging pile real-time health score and price sensitivity response model into a multi-objective dynamic scheduling decision maker, wherein the multi-objective dynamic scheduling decision maker is constructed by using a deep deterministic policy gradient algorithm, the state space of the multi-objective dynamic scheduling decision maker is composed of power grid load rate, user waiting queue length, charging pile health degree distribution and price gradient, the action space of the multi-objective dynamic scheduling decision maker is the matching relationship between each vehicle to be distributed and the available charging pile and the corresponding power output gear, and the reward function of the multi-objective dynamic scheduling decision maker is composed of the weighted sum of power grid stability gain term, user satisfaction improvement term, equipment loss suppression term and platform revenue growth term; S6, the multi-objective dynamic scheduling decision maker outputs the optimal charging pile distribution instruction and corresponding power setting value, and sends them to the charging pile control unit for execution; S7, when an emergency vehicle-to-grid instruction is issued by the power dispatching center or any charging pile has a hardware-level fault, a strategy reset mechanism is triggered immediately, all current non-emergency charging tasks are frozen, a global scheduling scheme satisfying the new constraint condition is recalculated, and the instruction issuing and execution unit switching is completed within 10 seconds.

2. The intelligent scheduling method of the high-performance charging platform according to claim 1, characterized in that, The power grid safe operation boundary constraint model is constructed, comprising: Taking 80% of the rated capacity of the regional transformer as the upper threshold of power, taking the absolute value of frequency deviation greater than 0.2 Hz as the emergency regulation trigger condition, and taking the voltage fluctuation amplitude exceeding the nominal value by ±5% as the power regulation prohibited interval, the three together constitute a three-dimensional safe operation hyperplane, and any scheduling instruction must ensure that the system state point is located in the safe region surrounded by the hyperplane after execution.

3. The intelligent scheduling method of the high-performance charging platform according to claim 2, characterized in that, The real-time health score of each charging pile is calculated, comprising: Collecting raw data through current transformers, voltage sensors, temperature sensors and vibration sensors at a frequency of 100 times per second, and inputting the data into the device health score calculation model after sliding average filtering; The device health score calculation model adopts a support vector regression algorithm, the input features are the average output power fluctuation rate, the maximum radiator temperature, the cumulative vibration energy value and the fault interval days of the charging pile in the past 30 days, and the output is a continuous score value between 0 and 1, and the charging pile with a score lower than 0.3 is prohibited from participating in any scheduling task.

4. The intelligent scheduling method of the high-performance charging platform according to claim 3, characterized in that, The price sensitivity response model is established, comprising: The model is constructed by using a long short-term memory network, input features include 24-hour historical time-of-use (TOU) price sequence, daily weather forecast temperature value, holiday identifier and regional electric vehicle (EV) growth rate, and output is the price prediction value and its confidence interval every 15 minutes in the next 2 hours; When the predicted price exceeds the user-set threshold, the time period is automatically marked as a low-priority charging window, and the power allocation weight of non-urgent charging tasks in the time period is reduced in the scheduling decision.

5. The intelligent scheduling method of the high-performance charging platform according to claim 4, characterized in that, The calculation method of the user satisfaction improvement term in the reward function is: For each user, the negative exponential function of the difference between the actual waiting time and the historical average waiting time is taken as the basic satisfaction value, the product of the charging cost saving ratio and the expected full charge time achievement rate is superimposed, and then multiplied by the user level weight coefficient, and finally the arithmetic mean of all users is taken to obtain the global user satisfaction index.

6. The intelligent scheduling method of the high-performance charging platform according to claim 5, characterized in that, The calculation method of the device loss suppression term in the reward function is: For each charging pile, the square of the difference between its current health score and the ideal health score is taken as the loss penalty factor, multiplied by the total charging time allocated to the charging pile in this round of scheduling, and then summed for all charging piles to obtain the device loss suppression term.

7. The intelligent scheduling method of the high-performance charging platform according to claim 6, characterized in that, The calculation method of the platform revenue growth term in the reward function is: For each charging order, the actual collection fee minus the grid purchase cost in the period is taken as the single order gross profit, all order gross profits are accumulated, and the energy feedback compensation paid to users for participating in the vehicle-to-grid response is deducted to obtain the net platform revenue value.

8. The intelligent scheduling method of the high-performance charging platform according to claim 7, characterized in that, After triggering the strategy reset mechanism, the emergency scheduling is executed, including: When an emergency vehicle-to-grid instruction is detected, vehicles with a battery state of charge higher than 80% and user authorization to participate in energy feedback are preferentially selected, sorted in descending order of dischargeable power capacity, and sequentially allocated to charging piles closest to the substation and with a health score higher than 0.8 to perform reverse discharge operation, with a discharge power upper limit set to 70% of the charging pile rated power and a discharge duration less than 30 minutes; When a charging pile hardware-level failure is detected, the charging pile is immediately removed from the available resource pool, all vehicles being served by the charging pile are sorted in ascending order of remaining charging capacity, and are sequentially reassigned to idle charging piles within a 3km range with the highest health score, ensuring that the expected completion time of each vehicle in the new allocation scheme does not exceed 120% of the original scheme.

9. An intelligent scheduling system for a high-performance charging platform, characterized in that, It includes: A power grid state perception module for acquiring the load curve, frequency deviation value and voltage stability margin index of the current regional power grid in real time through a power grid data interface, and constructing a power grid safe operation boundary constraint model; A user demand collection module for collecting the battery remaining capacity of the vehicle to be charged, the user-set expected full charge time, the acceptable maximum charging cost threshold and the permission identifier of whether to allow participation in vehicle-to-grid energy feedback through a user terminal interaction interface; A device health assessment module for collecting the cumulative running time, historical fault times, current output power fluctuation amplitude and radiator temperature value of each charging pile through a charging pile state monitoring unit, and calculating the real-time health score of each charging pile; The electricity price prediction response module is configured to obtain a time-of-use electricity price prediction sequence within the next 2 hours through the electricity market data interface, and to establish an electricity price sensitivity response model in combination with historical electricity price fluctuation rules; The multi-objective dynamic scheduling decision maker is configured to receive the power grid safe operation boundary constraint model, the user individualized charging demand parameter, the real-time health score of the charging pile, and the electricity price sensitivity response model, and to be constructed using a deep deterministic policy gradient algorithm, wherein a state space of the deep deterministic policy gradient algorithm is constituted by a power grid load rate, a user waiting queue length, a charging pile health degree distribution, and an electricity price gradient, an action space of the deep deterministic policy gradient algorithm is constituted by a matching relationship between each to-be-assigned vehicle and an available charging pile and a corresponding power output gear, and a reward function of the deep deterministic policy gradient algorithm is constituted by a weighted sum of a power grid stability gain term, a user satisfaction improvement term, a device loss suppression term, and a platform revenue growth term. The instruction execution and feedback module is configured to receive optimal charging pile allocation instructions and corresponding power setting values output by the multi-objective dynamic scheduling decision maker, to send the optimal charging pile allocation instructions and the corresponding power setting values to the charging pile control unit for execution, and to return execution states and deviation information in real time. The emergency strategy reconstruction module is configured to trigger a strategy reset mechanism immediately when an emergency vehicle-to-grid instruction is issued by a power grid dispatching center or any charging pile has a hardware-level fault, to freeze all current non-emergency charging tasks, to recalculate a global scheduling scheme that satisfies new constraint conditions, and to complete switching of the instruction issuing and execution unit within 10 seconds.

10. The intelligent scheduling system of high-performance charging platforms according to claim 9, characterized in that, The instruction execution and feedback module includes an instruction verification submodule configured to verify, before issuing a scheduling instruction, whether the scheduling instruction violates the power grid safe operation boundary constraint model, whether the scheduling instruction exceeds a user-set cost threshold, and whether the scheduling instruction is allocated to a charging pile with a health score lower than 0.5, and to return the scheduling instruction to the multi-objective dynamic scheduling decision maker for local correction if any of the violation conditions exists, wherein the number of correction times is limited to three, and if the scheduling instruction still cannot satisfy the constraint conditions after three times of correction, the user task is marked as unschedulable and the user is notified to reset parameters.

Images