Multiport hybrid charging pile power pool reinforcement learning scheduling method and system

Abstract

The invention discloses a multi-port hybrid charging pile power pool reinforcement learning scheduling method and system, and relates to the technical field of computer data processing. The reinforcement learning scheduling method and system for the multi-port hybrid charging pile power pool comprises the following steps: S1, periodically collecting port scheduling data, and preprocessing the port scheduling data; s2, constructing and training a vehicle target SOC prediction model, outputting a vehicle target SOC, evaluating charging urgency of each accessed vehicle, and dividing scheduling priority levels; s3, constructing a port power distribution model, outputting a power distribution value of each charging port, and issuing the power distribution value to a power control interface for execution; and S4, evaluating power scheduling execution deviation, and entering a protection mode when evaluation is abnormal. The problems that the total power of the charging pile cannot be flexibly scheduled due to fixed-proportion power distribution and topological isolation of alternating-current and direct-current ports, high-power resources are occupied by high-electric-quantity vehicles, and charging of low-electric-quantity vehicles is limited are solved.

Description

Technical Field

[0001] This invention relates to the field of computer data processing technology, specifically to a reinforcement learning scheduling method and system for multi-port hybrid charging pile power pools. Background Technology

[0002] With the rapid popularization of new energy vehicles, charging piles, as the core infrastructure for electric vehicle energy replenishment, directly affect users' charging experience and the utilization rate of power grid resources through their operating efficiency and power allocation strategies. Existing technologies have proposed various methods for monitoring the operation and power allocation of charging piles, but in complex scenarios, problems such as insufficient power utilization, rigid scheduling, and a lack of adaptive optimization still exist.

[0003] For example, the invention with announcement number CN118885703B proposes a method for detecting and analyzing the operating status of high-power devices in charging piles, belonging to the field of charging pile technology. It includes the following steps: Step S1: Obtaining detection data of charging piles under different scenarios based on charging pile detection equipment, and establishing a data model library; Step S2: Classifying the detection data according to charging gun power and type, preprocessing the classified detection data, and handling missing values ​​and noise data; Step S3: Embedding the preprocessed detection data into the data model, using the least squares fitting method in a high-order polynomial fitting algorithm to process the preprocessed detection data, and finally calculating the optimal fitting curve; Step S4: Deploying the optimal fitting curve, evaluating the fitting result, obtaining the current detection data of the charging pile, comparing and analyzing the fitting result of the optimal fitting curve with the current detection data, and determining existing faults and potential faults of the charging pile.

[0004] For example, the invention with announcement number CN119416651B provides a power allocation method, medium, and system for multi-gun charging piles, belonging to the field of charging pile technology. It includes: firstly, acquiring historical operating data of the charging piles and establishing a mathematical model containing multiple error terms, such as equations for charging power, time, and battery capacity. Then, constructing a hybrid model, including a multi-branch neural network and a mathematical model, to adapt to charging piles of different sizes. The power allocation model is obtained by training with historical data. During actual charging, real-time data is input and the power allocation is adjusted, while feedback data is collected. Based on real-time data and prediction errors, a lightweight LoRa model is used to continuously fine-tune the fusion sub-network, and the main model is updated periodically.

[0005] However, although the above-mentioned technical solutions have made breakthroughs in operation status detection, mathematical modeling and multi-head power allocation, they still have the following shortcomings: First, the power allocation strategy still tends to be static modeling and lacks dynamic perception of vehicle battery status, urgent needs and environmental conditions; second, the scheduling algorithm lacks flexibility in multi-port mixed scenarios and it is difficult to achieve the overall optimal utilization of the power pool; third, there is a lack of feedback loop at the execution level, which makes it impossible for the model to learn and adapt to abnormal execution situations.

[0006] Therefore, in order to address the above issues, there is an urgent need for a reinforcement learning scheduling method and system for multi-port hybrid charging pile power pools. Summary of the Invention

[0007] Technical problems to be solved

[0008] To address the shortcomings of existing technologies, this invention provides a reinforcement learning scheduling method and system for the power pool of multi-port hybrid charging piles. This method solves the problem that fixed power allocation and AC / DC port topology isolation prevent flexible scheduling of the total power of charging piles, resulting in high-battery vehicles occupying large power resources and low-battery vehicles having limited charging capacity.

[0009] Technical solution

[0010] To achieve the above objectives, this invention provides the following technical solution: a reinforcement learning scheduling method and system for multi-port hybrid charging pile power pools, comprising: S1, periodically collecting port scheduling data and performing synchronization alignment, anomaly removal, dimension elimination, and normalization processing on the port scheduling data to obtain preprocessed port scheduling data; S2, based on the preprocessed port scheduling data, constructing and training a vehicle target SOC prediction model, inputting real-time port scheduling data into the prediction model, outputting the vehicle target SOC, and evaluating the charging urgency of each connected vehicle and classifying scheduling priority levels; S3, calling the port scheduling data and sorting results of connected vehicles in the real-time scheduling queue, constructing a port power allocation model, outputting the power allocation value of each charging port, and sending it to the power control interface to execute real-time power scheduling; S4, after power allocation is executed, evaluating the deviation during the power scheduling execution process, continuing charging when the evaluation result is normal, and entering protection mode when the evaluation result is abnormal, and re-triggering the target SOC prediction and power allocation process to achieve closed-loop optimization of power scheduling.

[0011] Further, the specific steps for periodically collecting port scheduling data and performing synchronous alignment, anomaly elimination, dimension elimination, and normalization on the port scheduling data to obtain the preprocessed port scheduling data are as follows: Use a fixed-width sliding time window as a sampling period to periodically collect port scheduling data, which includes vehicle battery SOC, vehicle battery voltage, vehicle battery temperature, grid voltage, charging current, charging pile output power, and ambient temperature; Through constructing a synchronous sampling scheduling mechanism based on the access priority of multi-source data, perform timing alignment on asynchronous data in the port scheduling data; By adopting an exponential sliding average filtering algorithm, perform noise suppression and signal smoothing on the port scheduling data; Through the univariate outlier detection algorithm of the Z-Score method, eliminate outliers from the port scheduling data; Through the combination of quantile interval scaling and min-max normalization algorithm, perform cross-dimensional dimension elimination and data distribution unification on the port scheduling data.

[0012] Further, based on the preprocessed port scheduling data, the specific steps for constructing and training a vehicle target SOC prediction model and inputting real-time port scheduling data into the prediction model to output the vehicle target SOC are as follows: Extract the historical port scheduling data after preprocessing, and use the supervised learning method to construct a vehicle target SOC prediction model based on a multi-layer feedforward neural network; Use the port scheduling data as input features and the vehicle target SOC when the vehicle actually completes the charging task as the supervision label to form a prediction training sample set; And use the minimum mean square error as the loss function to train the vehicle target SOC prediction model; Input the real-time port scheduling data into the trained vehicle target SOC prediction model to output the vehicle target SOC.

[0013] Further, the specific steps for evaluating the charging urgency of each connected vehicle are as follows: Square the difference between the vehicle target SOC and the vehicle battery SOC, and then divide it by the vehicle battery temperature plus one to obtain the power difference temperature control factor; Divide the rated output current by the sum of the charging current and the minimum term, and take the power function value with the current suppression exponent as the exponent of the ratio to obtain the unit current utilization term; Take the natural logarithm of the sum of one and the ratio of the grid voltage to the vehicle battery voltage to obtain the voltage matching degree term; Multiply the power difference temperature control factor, the unit current utilization term, and the voltage matching degree term in sequence to obtain the scheduling urgency evaluation value.

[0014] Further, the specific steps for dividing the scheduling priority levels are as follows: Compare the scheduling urgency evaluation value D with the multi-level scheduling thresholds D1 and D2 in real time, and divide the connected vehicles into three priority queues: When D ≤ D1, allocate the vehicle to the scheduling waiting queue and postpone the power allocation for the current cycle; When D1 < D < D2, allocate the vehicle to the regular scheduling queue and enter the normal power scheduling process; When D ≥ D2, allocate the vehicle to the immediate scheduling queue and sort it from largest to smallest according to the scheduling urgency evaluation value.

[0015] Further, the specific steps for constructing a port power allocation model by calling the port scheduling data and ranking results of the connected vehicles in the real-time scheduling queue, outputting the power allocation value of each charging port, and sending it to the power control interface to execute real-time power scheduling are as follows: Call the port scheduling data, scheduling urgency evaluation value, and ranking results corresponding to the connected vehicles in the real-time scheduling queue; select the deep deterministic policy gradient algorithm as the training framework to construct a port power allocation model based on a policy network and a value network group; construct an allocation training sample set using historical port scheduling data and actual power allocation records; construct a composite reward function that integrates scheduling deviation, temperature control pressure, and voltage deviation during training, and obtain the function output value; use the function output value as the cumulative negative reward value to drive the port power allocation model to perform parameter optimization and policy convergence through the policy gradient update mechanism; input the port scheduling data and ranking results of the connected vehicles in the real-time scheduling queue into the trained port power allocation model, output the power allocation value corresponding to each charging port; and send the power allocation value as an execution instruction to each corresponding charging port for execution through the power control interface.

[0016] Furthermore, the specific steps for constructing a composite reward function that integrates scheduling deviation, temperature control pressure, and voltage deviation to obtain the function output value are as follows: Square the power allocation value of the previous cycle and multiply it by the charging current to obtain the energy load term; subtract the charging current from the port rated current, add a minimum term, and then multiply by the scheduling urgency assessment value to obtain the normalized scheduling benchmark term; divide the energy load term by the normalized scheduling benchmark term to obtain the scheduling deviation penalty term; divide the absolute value of the difference between the vehicle battery temperature and the ambient temperature by the sum of the vehicle battery temperature and the minimum term to obtain the temperature control pressure term; divide the absolute value of the difference between the grid voltage and the vehicle battery voltage by the sum of the grid voltage, the vehicle battery voltage, and the minimum term to obtain the voltage deviation term; and sum the scheduling deviation penalty term, the temperature control pressure term, and the voltage deviation term sequentially to obtain the function output value.

[0017] Furthermore, after power allocation is executed, the specific steps for evaluating deviations during power scheduling execution are as follows: After power allocation is executed, extract the real-time charging pile output power and corresponding port scheduling data. Multiply the absolute value of the difference between the grid voltage and the vehicle battery voltage by the charging current to obtain the voltage deviation current response term. Divide the voltage deviation current response term by the sum of the vehicle battery temperature and the minimum term to obtain the temperature control normalization term. Divide the power allocation value by the sum of the difference between the port rated current and the charging current and the minimum term to obtain the current load amplification term. Multiply the temperature control normalization term by the current load amplification term to obtain the scheduling load response term. Divide the absolute value of the difference between the vehicle battery SOC of the current cycle and the vehicle battery SOC of the previous cycle by the sum of the sampling time interval and the minimum term, and square it to obtain the SOC change rate term. Divide the difference between the charging pile output power and the power allocation value by the sum of the power allocation value and the minimum term, and square it to obtain the power execution deviation term. Add the scheduling load response term, the SOC change rate term, and the power execution deviation term to obtain the scheduling anomaly evaluation value.

[0018] Furthermore, when the evaluation result is normal, charging continues; when the evaluation result is abnormal, it enters protection mode and re-triggers the target SOC prediction and power allocation process. The specific steps for achieving closed-loop optimization of power scheduling are as follows: The scheduling anomaly evaluation value is compared with the anomaly threshold in real time. When the scheduling anomaly evaluation value is less than or equal to the anomaly threshold, the current scheduling behavior is determined to be normal, and the charging task continues. When the scheduling anomaly evaluation value is greater than the anomaly threshold, the current scheduling behavior is determined to be abnormal, the power output of the current charging port is suspended, and protection mode is entered. Real-time port scheduling data is then extracted to re-predict the vehicle's target SOC and re-allocate the power. The system calculates and monitors the battery voltage change rate and vehicle battery temperature in real time during protection mode. When the battery voltage change rate exceeds the voltage change threshold, the V2G power limiting mechanism is activated. When the vehicle battery temperature is below the battery temperature threshold, the system switches from the regular fast charging mode to the pulse heating mode, using the battery's internal resistance to generate heat for rapid temperature rise until the vehicle battery temperature exceeds the battery temperature threshold, at which point it switches back to the regular fast charging mode. Simultaneously, the port scheduling data and sorting results of the current cycle are added as new samples to the allocation training sample set for offline iterative training of the subsequent port power allocation model, forming a self-learning closed-loop mechanism based on scheduling performance feedback.

[0019] The second aspect of this invention provides a reinforcement learning scheduling system for a multi-port hybrid charging pile power pool, comprising: a port data acquisition and preprocessing module, a vehicle demand priority assessment module, a charging power intelligent allocation module, and a scheduling anomaly closed-loop optimization module. The port data acquisition and preprocessing module periodically acquires port scheduling data and performs synchronization alignment, anomaly removal, dimension elimination, and normalization processing on the port scheduling data to obtain preprocessed port scheduling data. The vehicle demand priority assessment module constructs and trains a vehicle target SOC prediction model based on the preprocessed port scheduling data, inputs real-time port scheduling data into the prediction model, and outputs the vehicle target SOC prediction model. The system calculates the target SOC of each vehicle and assesses the charging urgency of each connected vehicle, assigning scheduling priorities. The intelligent charging power allocation module retrieves port scheduling data and sorting results from the real-time scheduling queue, constructs a port power allocation model, outputs the power allocation value for each charging port, and sends it to the power control interface for real-time power scheduling. The scheduling anomaly closed-loop optimization module evaluates deviations during power scheduling execution. If the evaluation result is normal, charging continues; if the evaluation result is abnormal, it enters protection mode and re-triggers the target SOC prediction and power allocation process, achieving closed-loop optimization of power scheduling.

[0020] Beneficial effects

[0021] The present invention has the following beneficial effects:

[0022] (1) The reinforcement learning scheduling method and system for the power pool of the multi-port hybrid charging pile constructs a scheduling urgency evaluation value, takes multi-dimensional data such as vehicle target SOC, vehicle battery SOC, vehicle battery temperature, current, and voltage as input, calculates the charging urgency of each vehicle in real time, and divides the scheduling priority level accordingly. This makes the power allocation no longer dependent on a fixed ratio or static rules, but dynamically adjusted according to the vehicle status, environmental conditions and grid conditions, realizing differentiated allocation and effectively avoiding the problem of high-power vehicles occupying high power resources for a long time.

[0023] (2) The multi-port hybrid charging pile power pool reinforcement learning scheduling method and system constructs a power pooled port power allocation model, integrates AC ports and DC ports into the power pool management, and combines reinforcement learning strategy for global optimization, so that each port can be a dynamic unit in the power pool and can realize cross-port power scheduling under the premise of meeting the constraints, thereby breaking through the limitations brought by topology isolation, enabling the total power resources to flow reasonably between ports, and significantly improving the overall power utilization of the system.

[0024] (3) The reinforcement learning scheduling method and system for the power pool of the multi-port hybrid charging pile, through the urgency ranking queue mechanism, enables vehicles with low SOC or long-term access to automatically enter the real-time scheduling queue and obtain power allocation first. Combined with the real-time optimization capability of the power allocation model, when the total power is limited, priority is given to vehicles that urgently need to replenish energy, thereby minimizing the user's waiting time, alleviating the situation of limited charging for low-battery vehicles, and significantly improving charging efficiency and user experience.

[0025] (4) The reinforcement learning scheduling method and system for the power pool of the multi-port hybrid charging pile, by constructing a scheduling anomaly evaluation value, integrates indicators such as voltage deviation current response, SOC change rate, and power execution deviation into a multi-dimensional detection system, and quantifies the deviation in the scheduling execution process in real time; when the anomaly evaluation value exceeds the threshold, the system automatically triggers the protection mechanism, performs power limiting or pauses output, and recalculates the power allocation, thereby effectively preventing risks such as overcurrent, overheating and voltage mismatch, and ensuring the safe operation of the charging pile and vehicle battery. Attached Figure Description

[0026] Figure 1 Flowchart of reinforcement learning scheduling method for power pool of multi-port hybrid charging pile;

[0027] Figure 2 A structural diagram of a reinforcement learning scheduling system for a multi-port hybrid charging pile power pool.

[0028] Figure 3 This is a schematic diagram illustrating the determination of charging queues based on scheduling urgency assessment values.

[0029] Figure 4 This is a flowchart of fault response and temperature control recovery in protection mode. Detailed Implementation

[0030] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] Please see Figures 1-4This invention provides a technical solution: a reinforcement learning scheduling method and system for multi-port hybrid charging pile power pools, comprising: S1, periodically collecting port scheduling data and performing synchronization alignment, anomaly removal, dimension elimination, and normalization processing on the port scheduling data to obtain preprocessed port scheduling data; S2, based on the preprocessed port scheduling data, constructing and training a vehicle target SOC prediction model, inputting real-time port scheduling data into the prediction model, outputting the vehicle target SOC, and evaluating the charging urgency of each connected vehicle and classifying scheduling priority levels; S3, calling the port scheduling data and sorting results of connected vehicles in the real-time scheduling queue, constructing a port power allocation model, outputting the power allocation value of each charging port, and sending it to the power control interface to execute real-time power scheduling; S4, after power allocation is executed, evaluating the deviation during the power scheduling execution process, continuing charging when the evaluation result is normal, and entering protection mode when the evaluation result is abnormal, and re-triggering the target SOC prediction and power allocation process to achieve closed-loop optimization of power scheduling.

[0032] Specifically, port scheduling data is periodically collected, and synchronization alignment, anomaly removal, dimension elimination, and normalization are performed on the port scheduling data. The specific steps to obtain preprocessed port scheduling data are as follows: A fixed-width sliding time window is set as one sampling period, and port scheduling data is periodically collected according to the set sampling frequency and timestamp. The port scheduling data includes vehicle battery SOC, vehicle battery voltage, vehicle battery temperature, grid voltage, charging current, charging pile output power, and ambient temperature. Each piece of port scheduling data is bound to a unique timestamp and retains its original accuracy. By constructing a synchronous sampling and scheduling mechanism based on the priority of multi-source data access, asynchronous data is interpolated and time-series aligned according to the refresh rate and time lag characteristics of each port scheduling data to ensure that each... Port scheduling data is synchronized at the frame level within the sampling period. An exponential moving average filtering algorithm is used to fuse historical trends and current values, performing noise suppression and signal smoothing on the port scheduling data sequence to improve the stability and continuity of the original data. Furthermore, a univariate outlier detection algorithm based on the Z-Score method is used to mark and remove extreme data points in the port scheduling data that deviate from the standard deviation by more than a set threshold, reducing the impact of outliers on subsequent modeling performance. Finally, a quantile interval scaling combined with a min-max normalization algorithm is used to normalize and unify the scale of port scheduling data under different dimensions, achieving dimensionless representation of various types of port scheduling data in the feature space, providing high-quality standardized input for subsequent prediction models and decision-making modules.

[0033] In this implementation scheme, a synchronous sampling scheduling mechanism based on the access priority of multi-source data is constructed. Combined with the exponential moving average filtering algorithm, the Z-Score univariate outlier detection algorithm, and the quantile interval scaling and min-max normalization algorithm, high-precision preprocessing of port scheduling data is achieved. This significantly improves the temporal consistency and data distribution uniformity of asynchronous scheduling data, enhances the input data quality and modeling robustness of subsequent vehicle target SOC prediction models and power allocation models, and thus improves the real-time performance and reliability of the scheduling decision-making process.

[0034] Specifically, based on preprocessed port scheduling data, a vehicle target SOC prediction model is constructed and trained. The specific steps for inputting real-time port scheduling data into the prediction model and outputting the vehicle target SOC are as follows: Historical port scheduling data, processed through synchronization alignment, anomaly removal, dimension elimination, and normalization, is extracted. A supervised learning method is used to construct a vehicle target SOC prediction model based on a multi-layer feedforward neural network structure. The feedforward neural network structure consists of an input layer, multiple hidden layers, and an output layer. The input layer receives port scheduling data as input features to the model. The hidden layers use the ReLU activation function to achieve non-linear mapping. The number of nodes in each hidden layer is dynamically adjusted according to the training error to improve the model's accuracy in predicting target SOC. The model is designed to fit the relationship between the features and the target SOC of the connected vehicle. The output layer uses a linear activation function to output the target SOC prediction value of the connected vehicle. The parameters in the port scheduling data are used as input features, and the target SOC of the vehicle collected when the vehicle actually completes the charging task is used as the supervision label to form a prediction training sample set with a complete structure and sufficient sample distribution. The minimum mean square error is used as the loss function. The backpropagation algorithm and gradient descent optimization strategy are combined to train the vehicle target SOC prediction model by parameter iteration and error minimization. After the model training is completed, the port scheduling data collected in real time and preprocessed is input into the vehicle target SOC prediction model to output the target SOC prediction result of the connected vehicle at the current time.

[0035] In this implementation scheme, preprocessed port scheduling data is introduced as input features, and a multi-layer feedforward neural network structure is used to achieve nonlinear feature fitting. The minimum mean square error is used as the loss function, and the backpropagation algorithm and gradient descent optimization strategy are combined to effectively improve the accuracy and stability of the vehicle target SOC prediction results. Furthermore, it achieves high-accuracy identification of the charging demand of connected vehicles, providing highly reliable data support for subsequent scheduling urgency assessment and power allocation strategies, and enhancing the system's real-time perception capability and adaptive scheduling performance in multi-port hybrid charging scenarios.

[0036] Specifically, the steps for assessing the charging urgency of each connected vehicle are as follows: First, the square of the difference between the predicted target SOC of the vehicle and the real-time collected vehicle battery SOC is taken to characterize the deviation between the current expected charging level and the actual state of charge. Then, the squared difference is divided by the sum of the vehicle battery temperature and a constant, which is used to introduce the regulating effect of the vehicle battery thermal state on charging demand, resulting in a power difference temperature control factor. Second, the rated output current of the port is divided by the sum of the charging current and a minimum term. The minimum term is a non-zero, extremely small positive real number used to avoid numerical anomalies caused by the denominator approaching zero. The comparison value is then taken as a power function value with the current suppression exponent as the exponent, forming a unit current utilization term to quantify the current utilization efficiency control capability of the current port. The current suppression index is obtained by analyzing the deviation distribution between charging current and rated output current in historical port scheduling data, using a statistical fitting algorithm based on quantile intervals, with a value range of (0-1). Next, the ratio of real-time grid voltage to vehicle battery voltage is incremented by 1 and its natural logarithm is taken to obtain the voltage matching degree term, which characterizes the impact of charging system voltage adaptability on scheduling behavior. Finally, the power difference temperature control factor, unit current utilization term, and voltage matching degree term are multiplied sequentially to integrate the combined influence of multiple factors, outputting a scheduling urgency assessment value that comprehensively characterizes the urgency of charging access for vehicles.

[0037] The specific formula for calculating the dispatch urgency assessment value is as follows:

[0038] ;

[0039] In the formula, This represents the urgency assessment value for dispatching. Indicates the vehicle battery's SOC (State of Charge). Indicates the vehicle's target SOC. Indicates the vehicle battery temperature. Indicates the charging current. Indicates the rated output current. Indicates the vehicle battery voltage. Indicates the grid voltage. Indicates minterms, This indicates the current suppression index.

[0040] In this embodiment, Table 1 is a data table of scheduling urgency assessment values, listing the key characteristic variables and corresponding scheduling urgency assessment values ​​of the five access vehicles in the current scheduling cycle. The key variables used include: vehicle battery SOC, vehicle target SOC, vehicle battery temperature, charging current, rated output current, vehicle battery voltage, and grid voltage. The current suppression index is uniformly set to 0.1. Specific data are as follows: In vehicle 1, the vehicle battery SOC is 50%, the target SOC is 80%, the battery temperature is 28°C, the charging current is 40A, the rated output current is 120A, the battery voltage is 360V, the grid voltage is 400V, and the corresponding dispatch urgency assessment value is 25.88; In vehicle 2, the vehicle battery SOC is 45%, the target SOC is 90%, the battery temperature is 32°C, the charging current is 70A, the rated output current is 150A, the battery voltage is 380V, the grid voltage is 400V, and the corresponding dispatch urgency assessment value is 47.62; In vehicle 3, the vehicle battery SOC is 35%, the target SOC is 85%, and the battery temperature is... The evaluation value for vehicle 1 is 82.63, with a battery SOC of 25, a charging current of 30, a rated output current of 100, a vehicle battery voltage of 350, and a grid voltage of 400. For vehicle 2, the evaluation value is 82.63, with a battery SOC of 50, a target SOC of 95, a battery temperature of 35, a charging current of 80, a rated output current of 150, a battery voltage of 380, and a grid voltage of 400. For vehicle 3, the evaluation value is 43.07, with a battery SOC of 40, a target SOC of 85, a battery temperature of 30, a charging current of 50, a rated output current of 120, a battery voltage of 365, and a grid voltage of 400.

[0041] Table 1. Data Table of Dispatch Urgency Assessment Values

[0042]

[0043] like Figure 3 The figure shows the distribution of dispatch urgency assessment values ​​for five access vehicles within the same sampling period. The horizontal axis represents the access vehicle number, and the vertical axis represents the dispatch urgency assessment value D. Different colored line points in the figure indicate the dispatch queues to which the vehicles belong: green represents the dispatch waiting queue, blue represents the regular dispatch queue, and red represents the immediate dispatch queue. As can be seen from the figure, vehicle 1 has a dispatch urgency assessment value of 25.88, which is below the first-level dispatch threshold, placing it in the dispatch waiting queue; vehicles 2, 4, and 5 have dispatch urgency assessment values ​​of 47.62, 43.07, and 52.75, respectively, falling between the first-level and second-level thresholds, placing them in the regular dispatch queue; vehicle 3 has a dispatch urgency assessment value of 82.63, which is above the second-level dispatch threshold, classifying it as an immediate dispatch queue, and requiring priority allocation of power resources. Figure 3 This intuitively reflects the effectiveness of the present invention in classifying and determining the charging needs of multiple vehicles by using the scheduling urgency assessment value.

[0044] This implementation scheme achieves a quantitative assessment of the vehicle's current charging demand by constructing a scheduling urgency evaluation model that integrates the vehicle's target SOC, vehicle battery SOC, vehicle battery temperature, charging current, port rated output current, current suppression index, grid voltage, and vehicle battery voltage. The evaluation process incorporates three types of feature variables: energy difference temperature control factor, unit current utilization term, and voltage matching degree term, ensuring comprehensive coverage of the model across energy state, thermal state, and voltage adaptability dimensions. This method improves the accuracy of identifying the vehicle's actual charging demand, enhances the rationality of scheduling priority allocation, provides a more refined and dynamic scheduling basis for subsequent port power allocation, and significantly improves the overall system operating efficiency and resource allocation adaptability.

[0045] Specifically, the steps for classifying scheduling priorities are as follows: The scheduling urgency assessment value D is compared in real-time with the multi-level scheduling thresholds D1 and D2. Access vehicles are divided into three queues with different scheduling priorities based on their charging urgency level. When the scheduling urgency assessment value D is less than or equal to the first-level scheduling threshold D1, the corresponding access vehicle's current charging demand is deemed not urgent. The vehicle is marked as low priority and assigned to the scheduling waiting queue, with power allocation instructions for the current sampling period temporarily suspended. When the scheduling urgency assessment value D is greater than the first-level scheduling threshold D1 and less than the second-level scheduling threshold D2, the corresponding access vehicle is deemed to have medium charging urgency. It is then placed in the regular scheduling queue and enters the normal power scheduling process to receive regular power allocation. The normal power scheduling process involves caching and queuing the port scheduling data of access vehicles in the regular scheduling queue, using sequential polling to schedule vehicles one by one, inputting the vehicle's port scheduling data into the trained port power allocation model, calculating the power allocation value based on the current system load status, and executing charging tasks in batches after the real-time scheduling queue is completed. This ensures that medium-priority vehicles can complete energy replenishment without interfering with high-priority scheduling. When the dispatch urgency assessment value D is greater than or equal to the secondary dispatch threshold D2, the corresponding access vehicle is determined to have significant charging urgency and is included in the real-time dispatch queue. All access vehicles in the real-time dispatch queue are sorted in real time according to the dispatch urgency assessment value from largest to smallest to prioritize the charging resource allocation needs of vehicles with high urgency and improve the responsiveness and accuracy of dispatch.

[0046] In this implementation scheme, by using a real-time comparison mechanism based on the scheduling urgency assessment value D and the multi-level scheduling thresholds D1 and D2, the connected vehicles are divided into scheduling waiting queues, regular scheduling queues, and immediate scheduling queues. This ensures that vehicles with different charging urgency receive differentiated resource scheduling strategies. Specifically, low-urgency vehicles are temporarily deferred to avoid resource waste, medium-urgency vehicles obtain orderly energy replenishment through sequential polling and system load coordination, and high-urgency vehicles achieve priority guarantee of charging resources through the immediate response of the reinforcement learning model. This significantly improves the accuracy and responsiveness of power scheduling and the dynamic adaptability of the overall system load.

[0047] Specifically, the steps for calling the port scheduling data and sorting results of the connected vehicles in the real-time scheduling queue, constructing a port power allocation model, outputting the power allocation value of each charging port, and sending it to the power control interface to execute real-time power scheduling are as follows: Calling the port scheduling data, scheduling urgency assessment value, and sorting results corresponding to the connected vehicles in the real-time scheduling queue; selecting the deep deterministic policy gradient algorithm as the training framework; constructing a port power allocation model composed of a policy network and a value network; the policy network is used to generate power allocation strategies, and the value network is used to estimate the expected return of the current strategy; constructing an allocation training sample set using historical port scheduling data after synchronization alignment, anomaly removal, dimension elimination, and normalization processing, and corresponding actual power allocation records; during training, constructing a system that integrates scheduling deviation, temperature control pressure, and voltage... The composite reward function of the deviation quantifies the impact of the power allocation strategy on the system's operating state. The output value of the generator function serves as the cumulative negative reward value, driving the port power allocation model to continuously optimize the parameters of the port power allocation model and achieve asymptotic convergence of the final strategy through a policy gradient update mechanism. The port scheduling data and sorting results of the vehicles connected in the real-time scheduling queue of the current period are input into the trained port power allocation model. Through forward propagation, the power allocation value corresponding to each charging port is generated, achieving dynamic optimal allocation of resources under the constraints of safety, grid capacity, and vehicle charging capability. The power allocation value is then used as a power control command and sent to each charging port through the power control interface of the charging controller, driving the charging pile to perform real-time power scheduling and completing the implementation of the scheduling target within the current period.

[0048] In this implementation scheme, a port power allocation model based on a deep deterministic policy gradient algorithm is constructed by calling the port scheduling data and sorting results of vehicles accessing the real-time scheduling queue. A composite reward function incorporating scheduling deviation, temperature control pressure, and voltage deviation is introduced to achieve efficient training and adaptive optimization of the port power allocation strategy. This method not only improves the accuracy of response to scheduling urgency assessment values ​​but also dynamically outputs the power allocation values ​​corresponding to each charging port based on the system load status. This effectively ensures priority allocation of charging resources for vehicles with high urgency, enhances the real-time performance and precision of scheduling decisions, and significantly improves the resource allocation efficiency and charging service quality of multi-port hybrid charging piles under complex operating conditions.

[0049] Specifically, the steps for constructing a composite reward function that integrates scheduling deviation, temperature control pressure, and voltage deviation, and obtaining the function output value are as follows: First, the power allocation value allocated to the charging port in the previous cycle is squared and multiplied with the real-time charging current to construct an energy load term reflecting the current power allocation energy output load, used to measure the system power consumption intensity; Second, the current remaining capacity factor is constructed by subtracting the real-time charging current from the port rated current and adding a minimum term, and then multiplied with the current current remaining capacity factor and the current scheduling urgency assessment value to generate a normalized scheduling benchmark term under a unified standard, to characterize the reasonable power proportion of each access vehicle under the current system load conditions; Third, the energy load... The term is divided by the normalized scheduling baseline term to form a scheduling deviation penalty term used to measure the degree of deviation in allocation accuracy. Subsequently, the absolute value of the difference between the vehicle battery temperature and the ambient temperature is extracted and divided by the sum of the vehicle battery temperature and the minimum term to obtain a temperature control pressure term that measures the thermal stress level of the system. Further, the absolute value of the difference between the grid voltage and the vehicle battery voltage is calculated and divided by the sum of the grid voltage, the vehicle battery voltage and the minimum term to form a voltage deviation term that reflects the voltage adaptation state. Finally, the scheduling deviation penalty term, the temperature control pressure term and the voltage deviation term are accumulated to obtain the function output value, which serves as a comprehensive feedback index of the composite reward function in the port power allocation model, used to drive gradient updates and policy iterations in the policy optimization process.

[0050] The specific formula for calculating the function's output value is as follows:

[0051] ;

[0052] In the formula, Indicates the function output value. This represents the power allocation value from the previous cycle. This represents the urgency assessment value for dispatching. Indicates the charging current. Indicates the rated current of the port. Indicates the vehicle battery temperature. Indicates ambient temperature. Indicates the grid voltage. Indicates the vehicle battery voltage. Indicates a minus term.

[0053] In this implementation scheme, a composite reward function is constructed that integrates the scheduling deviation penalty term, temperature control pressure term, and voltage deviation term. This function fully integrates the power allocation value of the previous cycle, charging current, port rated current, scheduling urgency assessment value, vehicle battery temperature, ambient temperature, grid voltage, and vehicle battery voltage. It comprehensively quantifies the power scheduling accuracy, thermal safety margin, and voltage adaptability, thereby achieving precise guidance for the gradient optimization direction and strategy convergence path during the training process of the port power allocation model. This effectively improves the adaptability, robustness, and execution efficiency of the power scheduling strategy in multi-port high-load operation scenarios.

[0054] Specifically, after power allocation is executed, the specific steps for evaluating the deviation during the power scheduling process are as follows: After power allocation is completed, extract the real-time charging pile output power, grid voltage, vehicle battery voltage, vehicle battery temperature, charging current, port rated current, vehicle battery SOC, and power allocation value; First, use the absolute value of the difference between the grid voltage and the vehicle battery voltage as the voltage offset, and multiply it with the charging current in the current sampling period to obtain the voltage deviation current response term, which reflects the actual impact of voltage deviation on charging behavior; Second, divide the voltage deviation current response term by the sum of the vehicle battery temperature and the minimum term to obtain the temperature control normalization term affected by temperature regulation; Further, divide the power allocation value by the difference between the port rated current and the current charging current, and add the minimum term to obtain the current load amplification term that measures the current load level of the port; Then, multiply the temperature control normalization term by the current load amplification term. Multiplying the load amplification terms yields the scheduling load response term, which characterizes the interaction between scheduling load pressure and thermal risk. Then, the absolute value of the difference between the vehicle battery SOC in the current sampling period and the vehicle battery SOC in the previous period is divided by the sum of the sampling time interval and the minimum term, and the ratio is squared to obtain the SOC change rate term, which describes the rate of change of charge state. Next, the difference between the actual output power of the charging pile and the corresponding power allocation value is divided by the sum of the power allocation value and the minimum term, and the ratio is squared to obtain the power execution deviation term, which reflects the degree of deviation in power scheduling execution. Finally, the scheduling load response term, the SOC change rate term, and the power execution deviation term are added sequentially to form a scheduling anomaly assessment value that comprehensively evaluates the accuracy of power scheduling execution, system thermal stability, and power transfer efficiency. This value supports the determination of the effectiveness of subsequent system scheduling behavior and the triggering strategy of protection mechanisms.

[0055] The specific formula for calculating the scheduling anomaly evaluation value is as follows:

[0056] ;

[0057] In the formula, This represents the evaluation value for scheduling anomalies. Indicates the charging current. Indicates the grid voltage. Indicates the vehicle battery voltage. Indicates the vehicle battery temperature. Indicates the power allocation value. Indicates the rated current of the port. Indicates the vehicle battery's SOC (State of Charge). This indicates the vehicle's battery SOC in the previous cycle. Indicates the sampling time interval. Indicates minterms, This indicates the output power of the charging station.

[0058] In this implementation plan, by constructing a scheduling anomaly assessment value and integrating three key factors—scheduled load response, SOC change rate, and power execution deviation—quantitative evaluation of the power allocation execution process in terms of load pressure, dynamic changes in power consumption, and power execution accuracy can be achieved. This can accurately identify potential abnormal operating conditions in the scheduling process, improve the robustness and reliability of power scheduling behavior, and provide a scientific basis and judgment basis for subsequent protection mode triggering and scheduling closed-loop optimization, thereby ensuring the thermal safety of system operation, the stability of power transmission, and the consistency of the charging process.

[0059] Specifically, when the evaluation result is normal, charging continues; when the evaluation result is abnormal, protection mode is entered, and the target SOC prediction and power allocation process is re-triggered. The specific steps for achieving closed-loop optimization of power scheduling are as follows: The scheduling anomaly evaluation value is compared with the anomaly threshold in real time. When the scheduling anomaly evaluation value is less than or equal to the anomaly threshold, the current scheduling behavior is determined to be in normal operation, the controller maintains the predetermined power output of the charging port, and continues to execute the current charging task; when the scheduling anomaly evaluation value is greater than the anomaly threshold, an abnormal risk is identified in the current scheduling behavior, the power output command of the current charging port is immediately suspended, the control logic switches to protection mode, and the port scheduling data collected at the current moment is extracted. The vehicle target SOC prediction model inference is re-executed to obtain the updated vehicle target SOC result, and based on the real-time scheduling queue and sorting results, the power allocation value of the corresponding port is recalculated to correct the scheduling deviation; in protection mode, such as Figure 4As shown, the system monitors two safety indicators of the vehicle battery: the rate of change of battery terminal voltage and the vehicle battery temperature, through high-speed sampling. When the rate of change of battery voltage exceeds the voltage change threshold, the V2G power limiting mechanism is activated. The upper limit of the discharge command is dynamically lowered through the power control interface to limit the output power, thereby effectively suppressing the risk of lithium dendrite formation and preventing abnormal battery overcurrent. When the vehicle battery temperature is lower than the battery temperature threshold, the charging control strategy is switched from the conventional fast charging mode to the pulse heating mode. An intermittent current excitation strategy is used to rapidly increase the cell temperature by utilizing the in-situ heating mechanism of the battery internal resistance until the vehicle battery temperature exceeds the battery temperature threshold. Then, the system automatically switches back to the conventional fast charging process and restores the standard charging path. At the same time, the port scheduling data processed in this cycle and its priority in the scheduling ranking are added as structured new samples to the allocation training sample set of the port power allocation model to participate in subsequent offline reinforcement learning iterative training tasks. This continuously optimizes the strategy parameters of the port power allocation model, ultimately forming a power scheduling adaptive optimization mechanism that integrates scheduling performance feedback, autonomous learning, and closed-loop control.

[0060] In this implementation scheme, a judgment mechanism based on the comparison of scheduling anomaly evaluation values ​​and anomaly thresholds is constructed to achieve high-precision identification of power scheduling behavior status. This ensures that the charging port maintains stable power output when scheduling behavior is normal and switches to protection mode immediately when scheduling behavior is abnormal, effectively preventing overload risks and strategy failures. In protection mode, high-frequency monitoring indicators of battery voltage change rate and vehicle battery temperature are introduced to further refine the safety response strategy. The battery operation safety and low-temperature adaptability are improved through V2G power limiting mechanism and pulse heating mode. At the same time, new samples are constructed and the training sample set is updated using the current period port scheduling data and sorting results, so that the port power allocation model can be continuously optimized in offline iteration. Finally, a power scheduling adaptive optimization mechanism with scheduling anomaly identification capability, safety response capability, and data-driven update capability is formed, comprehensively improving the intelligence, robustness, and adaptability of the system of this invention.

[0061] like Figure 2As shown, the second aspect of this invention provides a reinforcement learning scheduling system for a multi-port hybrid charging pile power pool, comprising: a port data acquisition and preprocessing module, a vehicle demand priority assessment module, a charging power intelligent allocation module, and a scheduling anomaly closed-loop optimization module. The port data acquisition and preprocessing module is used to periodically acquire port scheduling data and perform synchronization alignment, anomaly removal, dimension elimination, and normalization processing on the port scheduling data to obtain preprocessed port scheduling data. The vehicle demand priority assessment module is used to construct and train a vehicle target SOC prediction model based on the preprocessed port scheduling data, inputting real-time port scheduling data into the prediction model. The system calculates the target SOC of each vehicle and assesses the charging urgency of each connected vehicle, assigning scheduling priorities. The intelligent charging power allocation module retrieves port scheduling data and sorting results from the real-time scheduling queue, constructs a port power allocation model, outputs the power allocation value for each charging port, and sends it to the power control interface for real-time power scheduling. The scheduling anomaly closed-loop optimization module evaluates deviations during power scheduling execution. If the evaluation result is normal, charging continues; if the evaluation result is abnormal, it enters protection mode and re-triggers the target SOC prediction and power allocation process, achieving closed-loop optimization of power scheduling.

[0062] like Figure 4 As shown, this invention enters a protection mode in case of scheduling anomalies and adjusts the charging control strategy through dual safety indicators. First, the system continuously monitors two core state parameters: the vehicle battery voltage change rate and the vehicle battery temperature, and triggers corresponding control logic branches accordingly: if the battery voltage change rate exceeds the set voltage change threshold, the V2G power limiting mechanism is immediately activated to effectively suppress the risk of current surges caused by drastic voltage fluctuations; if the vehicle battery temperature is detected to be below the battery temperature threshold, a pulse heating mode is triggered to raise the cell temperature through intermittent current excitation; after the battery temperature rises above the safety threshold, the system automatically switches back to the normal fast charging process and restores the normal charging path.

[0063] In this implementation scheme, a multi-port hybrid charging pile power pool reinforcement learning scheduling system is constructed, comprising a port data acquisition and preprocessing module, a vehicle demand priority assessment module, a charging power intelligent allocation module, and a scheduling anomaly closed-loop optimization module, forming an end-to-end intelligent charging scheduling process. Based on periodically collected port scheduling data, the system completes data synchronization and alignment, anomaly removal, dimension elimination, and normalization processing to ensure input data quality and consistency. By constructing a vehicle target SOC prediction model and combining it with a scheduling urgency assessment mechanism, the system dynamically determines the charging demand of connected vehicles and classifies their priority levels, improving the accuracy of resource allocation. A port power allocation model is constructed based on a strategy gradient algorithm to achieve intelligent adjustment of the power values ​​of each charging port, improving the system's scheduling flexibility and response speed. Through a scheduling anomaly assessment mechanism and protection mode switching strategy, the system ensures the stability and security of the power scheduling process. Combined with a data-driven model self-learning mechanism, the system continuously optimizes and updates scheduling strategy parameters in a closed loop, comprehensively improving the intelligent scheduling capability and system adaptability of this invention under complex operating conditions.

[0064] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0065] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A reinforcement learning scheduling method for the power pool of a multi-port hybrid charging pile, characterized in that, Includes the following steps: S1 periodically collects port scheduling data and performs synchronization alignment, anomaly removal, dimension elimination and normalization processing on the port scheduling data to obtain preprocessed port scheduling data; S2. Based on the preprocessed port scheduling data, construct and train the vehicle target SOC prediction model, input the real-time port scheduling data into the prediction model, output the vehicle target SOC, evaluate the charging urgency of each access vehicle, and classify the scheduling priority level. S3 calls the port scheduling data and sorting results of the connected vehicles in the real-time scheduling queue, constructs the port power allocation model, outputs the power allocation value of each charging port, and sends it to the power control interface to execute real-time power scheduling. S4. After power allocation is executed, the deviation during the power scheduling process is evaluated. If the evaluation result is normal, charging continues. If the evaluation result is abnormal, the protection mode is entered and the target SOC prediction and power allocation process is re-triggered to achieve closed-loop optimization of power scheduling.

2. The reinforcement learning scheduling method for the power pool of a multi-port hybrid charging pile according to claim 1, characterized in that: The specific steps for periodically collecting port scheduling data and performing synchronization alignment, anomaly removal, dimension elimination, and normalization processing on the port scheduling data to obtain preprocessed port scheduling data are as follows: A fixed-width sliding time window is used as a sampling period to periodically collect port scheduling data, which includes vehicle battery SOC, vehicle battery voltage, vehicle battery temperature, grid voltage, charging current, charging pile output power, and ambient temperature. By constructing a synchronous sampling and scheduling mechanism based on the access priority of multi-source data, asynchronous data in port scheduling data is time-aligned; by employing an exponential moving average filtering algorithm, noise suppression and signal smoothing are performed on port scheduling data; an outlier removal is performed on port scheduling data using a univariate outlier detection algorithm based on the Z-Score method; and cross-dimensional dimension elimination and data distribution unification are achieved on port scheduling data by using a quantile interval scaling combined with a minimum-maximum normalization algorithm.

3. The reinforcement learning scheduling method for the power pool of a multi-port hybrid charging pile according to claim 1, characterized in that: The specific steps for constructing and training a vehicle target SOC prediction model based on preprocessed port scheduling data, inputting real-time port scheduling data into the prediction model, and outputting the vehicle target SOC are as follows: Preprocessed historical port scheduling data is extracted, and a vehicle target SOC prediction model based on a multi-layer feedforward neural network is constructed using a supervised learning method. The port scheduling data is used as the input feature, and the vehicle target SOC when the vehicle actually completes the charging task is used as the supervision label to form a prediction training sample set. The minimum mean square error is used as the loss function to train the vehicle target SOC prediction model; real-time port scheduling data is input into the trained vehicle target SOC prediction model to output the vehicle target SOC.

4. The reinforcement learning scheduling method for the power pool of a multi-port hybrid charging pile according to claim 1, characterized in that: The specific steps for assessing the charging urgency of each connected vehicle are as follows: After squaring the difference between the target SOC of the vehicle and the SOC of the vehicle battery and dividing it by the vehicle battery temperature plus one, the power difference temperature control factor is obtained; dividing the rated output current by the sum of the charging current and the minterm and taking the power function value with the current suppression index as the exponent of the ratio, the unit current utilization term is obtained; taking the natural logarithm after adding one to the ratio of the grid voltage to the vehicle battery voltage, the voltage matching degree term is obtained. Multiplying the power difference temperature control factor, the unit current utilization term, and the voltage matching degree term in sequence, the scheduling urgency evaluation value is obtained.

5. The reinforcement learning scheduling method for the power pool of a multi-port hybrid charging pile according to claim 1, characterized in that: The specific steps for dividing the scheduling priority levels are as follows: Compare the scheduling urgency evaluation value D with the multi-level scheduling thresholds D1 and D2 in real time, and divide the connected vehicles into three priority queues: When D ≤ D1, allocate the vehicle to the scheduling waiting queue and suspend the power allocation for the current cycle; When D1 < D < D2, allocate the vehicle to the regular scheduling queue and enter the normal power scheduling process; When D ≥ D2, allocate the vehicle to the immediate scheduling queue and sort it from largest to smallest according to the scheduling urgency evaluation value.

6. The reinforcement learning scheduling method for the power pool of a multi-port hybrid charging pile according to claim 1, characterized in that: The specific steps for calling the port scheduling data and sorting results of the connected vehicles in the immediate scheduling queue, constructing a port power allocation model, outputting the power allocation values of each charging port, and sending them to the power control interface to execute real-time power scheduling are as follows: Call the port scheduling data, scheduling urgency evaluation value, and sorting results corresponding to the connected vehicles in the immediate scheduling queue, select the deep deterministic policy gradient algorithm as the training framework, and construct a port power allocation model based on the policy network and value network group; Construct an allocation training sample set using historical port scheduling data and actual power allocation records. During the training process, construct a composite reward function that combines scheduling deviation, temperature control pressure, and voltage deviation to obtain the function output value; Use the function output value as the negative reward cumulative value to drive the port power allocation model to perform parameter optimization and policy convergence through the policy gradient update mechanism; Input the port scheduling data and sorting results of the connected vehicles in the immediate scheduling queue into the trained port power allocation model, and output the power allocation values corresponding to each charging port; and use the power allocation values as execution instructions and send them to the corresponding charging ports through the power control interface for execution.

7. The reinforcement learning scheduling method for the power pool of a multi-port hybrid charging pile according to claim 6, characterized in that: The specific steps for constructing a composite reward function that combines scheduling deviation, temperature control pressure, and voltage deviation to obtain the function output value are as follows: Square the power allocation value of the previous cycle and multiply it by the charging current to obtain the energy load term; subtract the charging current from the rated current of the port, add the minterm, and then multiply by the scheduling urgency evaluation value to obtain the normalized scheduling reference term; divide the energy load term by the normalized scheduling reference term to obtain the scheduling deviation penalty term; divide the absolute value of the difference between the vehicle battery temperature and the ambient temperature by the sum of the vehicle battery temperature and the minterm to obtain the temperature control pressure term; divide the absolute value of the difference between the grid voltage and the vehicle battery voltage by the sum of the grid voltage, the vehicle battery voltage, and the minterm to obtain the voltage deviation term; add the scheduling deviation penalty term, the temperature control pressure term, and the voltage deviation term in sequence to obtain the function output value.

8. The reinforcement learning scheduling method for the power pool of a multi-port hybrid charging pile according to claim 1, characterized in that: The specific steps for evaluating the deviation situation during the execution of power scheduling after the power allocation is performed are as follows: After power allocation is executed, the real-time charging pile output power and corresponding port scheduling data are extracted. The absolute value of the difference between the grid voltage and the vehicle battery voltage is multiplied by the charging current to obtain the voltage deviation current response term. The voltage deviation current response term is divided by the sum of the vehicle battery temperature and the minimum term to obtain the temperature control normalization term. The power allocation value is divided by the sum of the difference between the port rated current and the charging current and the minimum term to obtain the current load amplification term. The temperature control normalization term is multiplied by the current load amplification term to obtain the scheduling load response term. The absolute value of the difference between the vehicle battery SOC of the current cycle and the vehicle battery SOC of the previous cycle is divided by the sum of the sampling time interval and the minimum term, and the square is taken to obtain the SOC change rate term. The difference between the charging pile output power and the power allocation value is divided by the sum of the power allocation value and the minimum term, and the square is taken to obtain the power execution deviation term. The scheduling load response term, the SOC change rate term, and the power execution deviation term are added together to obtain the scheduling anomaly evaluation value.

9. The reinforcement learning scheduling method for the power pool of a multi-port hybrid charging pile according to claim 1, characterized in that: The specific steps for achieving closed-loop optimization of power scheduling—continuing charging when the evaluation result is normal, entering protection mode when the evaluation result is abnormal, and re-triggering the target SOC prediction and power allocation process—are as follows: The scheduling anomaly assessment value is compared with the anomaly threshold in real time. When the scheduling anomaly assessment value is less than or equal to the anomaly threshold, the current scheduling behavior is determined to be normal and the charging task continues. When the scheduling anomaly assessment value is greater than the anomaly threshold, the current scheduling behavior is determined to be abnormal, the power output of the current charging port is suspended, the system enters protection mode, and real-time port scheduling data is extracted to recalculate the vehicle target SOC prediction and power allocation value. In protection mode, the battery voltage change rate and vehicle battery temperature are monitored in real time. When the battery voltage change rate is greater than the voltage change threshold, the V2G power limiting mechanism is activated. When the vehicle battery temperature is less than the battery temperature threshold, the regular fast charging mode is switched to pulse heating mode, which uses the battery internal resistance to generate heat to quickly raise the temperature until the vehicle battery temperature is greater than the battery temperature threshold, at which point the mode is switched back to regular fast charging mode. Meanwhile, the port scheduling data and sorting results of the current period are added as new samples to the allocation training sample set for offline iterative training of the subsequent port power allocation model, forming a self-learning closed-loop mechanism based on scheduling performance feedback.

10. A reinforcement learning scheduling system for the power pool of a multi-port hybrid charging pile, characterized in that: include: The system includes a port data acquisition and preprocessing module, a vehicle demand priority assessment module, a charging power intelligent allocation module, and a scheduling anomaly closed-loop optimization module, among which: The port data acquisition and preprocessing module is used to periodically acquire port scheduling data and perform synchronization alignment, anomaly removal, dimension elimination and normalization processing on the port scheduling data to obtain preprocessed port scheduling data. The vehicle demand priority assessment module is used to construct and train a vehicle target SOC prediction model based on preprocessed port scheduling data, input real-time port scheduling data into the prediction model, output the vehicle target SOC, assess the charging urgency of each connected vehicle, and classify scheduling priority levels. The intelligent charging power allocation module is used to call the port scheduling data and sorting results of the connected vehicles in the real-time scheduling queue, construct the port power allocation model, output the power allocation value of each charging port, and send it to the power control interface to execute real-time power scheduling. The scheduling anomaly closed-loop optimization module is used to evaluate the deviation in the power scheduling process after power allocation is executed. When the evaluation result is normal, charging continues. When the evaluation result is abnormal, it enters protection mode and re-triggers the target SOC prediction and power allocation process to achieve closed-loop optimization of power scheduling.

Images