Intelligent scheduling system for traffic charging stations fused with machine learning
By integrating machine learning into the intelligent scheduling system, the problem of multi-source asynchronous data processing for electric vehicle charging stations is solved, enabling accurate description of the charging network and safe and feasible scheduling decisions, thus ensuring the stability and security of the power grid operation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- LINYI UNIVERSITY
- Filing Date
- 2026-03-12
- Publication Date
- 2026-05-19
AI Technical Summary
Existing electric vehicle charging station scheduling technologies fail to effectively handle multi-source asynchronous data, leading to equipment-level failures and uneven utilization. Ignoring the physical topology of the power grid may cause safety hazards and scheduling failures.
An intelligent scheduling system integrating machine learning is adopted. Through data acquisition and preprocessing modules, hierarchical decomposition and timestamp alignment are performed to identify potential conflict patterns, construct a scheduling relationship graph, generate scheduling task descriptions with topological constraints, and generate collaborative scheduling schemes using machine learning models.
It enables accurate and consistent description of the charging network, identifies supply and demand imbalances, ensures that the dispatching scheme conforms to the actual operating conditions of the power grid, and ensures safety and feasibility.
Smart Images
Figure CN121810005B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent scheduling technology for power grid charging, specifically to an intelligent scheduling system for transportation charging stations that integrates machine learning. Background Technology
[0002] Currently, the operation and scheduling of electric vehicle charging stations mainly rely on statistical analysis of historical data or monitoring and alarms based on simple thresholds. Existing technologies typically aggregate or average the operating parameters of charging piles, forming a single load curve at the station level for monitoring. When formulating scheduling strategies, optimization models that treat each charging station as an independent entity are commonly used, or only the matching relationship between charging demand and station static capacity and geographical distance is considered. These methods focus on describing the macroscopic state and independent responses.
[0003] These conventional technical solutions have shortcomings. The coarse aggregation of operating parameters masks details such as equipment-level faults and uneven utilization, and fails to consider the logical relationships between data at multiple levels, including equipment, stations, and regions. The timing asynchrony caused by different data sources due to varying collection frequencies and times is ignored, leading to biased situational assessments. Furthermore, neglecting the scheduling decisions based on the underlying power grid topology may render scheduling commands infeasible at the power transmission level, potentially causing safety hazards or scheduling failures.
[0004] Intelligent scheduling of charging stations faces two key challenges: how to construct a time-consistent state perception system from heterogeneous and asynchronous multi-source data that can truly reflect the overall picture from micro-devices to macro-regions; and how to deeply integrate charging demand conflicts with the physical connection relationship, power flow path, and capacity constraints of the power grid when generating scheduling instructions, so that scheduling decisions strictly follow physical laws and ensure the feasibility of the solution. Summary of the Invention
[0005] The purpose of this invention is to provide an intelligent scheduling system for traffic charging stations that integrates machine learning, so as to solve the problems mentioned in the background art.
[0006] To achieve the above objectives, this invention provides an intelligent scheduling system for transportation charging stations that integrates machine learning. The system includes: a data acquisition and preprocessing module, used to acquire raw monitoring records containing multiple charging station operating parameters, hierarchically decompose the parameters in the raw monitoring records into equipment-level parameters, station-level aggregated parameters, and regional-level statistical parameters, and asynchronously align the decomposed equipment-level parameters, station-level aggregated parameters, and regional-level statistical parameters based on timestamps to form a time-synchronized hierarchical parameter sequence; a conflict pattern identification module, used to identify potential conflict patterns representing the imbalance between charging demand and supply from the time-synchronized hierarchical parameter sequence; a scheduling relationship graph construction module, used to construct a scheduling relationship graph reflecting the charging station network topology and power flow paths; a task description generation module, used to associate and map potential conflict patterns with the scheduling relationship graph to generate scheduling task descriptions with topological constraints; and an intelligent scheduling decision module, used to use a machine learning model to parse the scheduling task descriptions with topological constraints and output a collaborative scheduling scheme containing power redistribution instructions and equipment start / stop commands.
[0007] Preferably, the hierarchical decomposition of parameters in the original monitoring records specifically includes: extracting voltage, current, power, occupancy status, and temperature parameters belonging to a single charging pile and classifying them as equipment-level parameters; summing, averaging, and calculating the variance of the equipment-level parameters of all charging piles belonging to the same charging station to obtain station-level aggregated parameters describing the overall load, average utilization rate, and operational stability of the charging station; and further summarizing and statistically analyzing the station-level aggregated parameters of all charging stations in the target area to generate regional-level statistical parameters reflecting the total regional demand, load distribution, and supply and demand trends.
[0008] Preferably, the asynchronous alignment processing of the decomposed equipment-level parameters, station-level aggregated parameters, and regional-level statistical parameters based on timestamps specifically includes: establishing independent time axes for equipment-level parameters, station-level aggregated parameters, and regional-level statistical parameters respectively; using the time axis with the highest sampling frequency as a reference, using an interpolation method to fill in the time points of the parameter sequences with lower sampling frequencies; and unifying the parameter sequences of all levels to the same and equally spaced time point sequences to form a time-synchronized hierarchical parameter sequence.
[0009] Preferably, identifying potential conflict patterns characterizing the imbalance between charging demand and supply from a time-synchronized hierarchical parameter sequence specifically includes: detecting charging pile clusters that continuously exceed thresholds in the device-level parameter sequence within the time-synchronized hierarchical parameter sequence; marking charging stations whose load rate is consistently higher than a preset capacity and which are predicted to have queuing vehicles in adjacent time periods in the station-level aggregated parameter sequence; locating the time periods and geographical areas where the total demand power curve rises sharply while the regional power supply curve shows a stable or declining trend in the regional statistical parameter sequence; and comprehensively defining the results of the above detection, marking, and location as potential conflict patterns.
[0010] Preferably, the construction of a scheduling relationship graph reflecting the network topology of charging stations and the path of power flow specifically includes: using each charging station as a graph node, with node attributes including the location, rated total power, current available power, and a list of connected piles; using the connection relationships of power distribution network lines, the geographical adjacency relationships between charging stations, and the connection relationships of shareable backup energy storage units as graph edges, with edge attributes including line capacity, physical distance, and connection type; and assigning real-time state attributes to each node and edge in the graph to form a dynamically weighted scheduling relationship graph.
[0011] Preferably, the step of associating and mapping potential conflict patterns with the scheduling relationship graph to generate a scheduling task description with topological constraints specifically includes: mapping the specific charging stations, time periods, and power deficit information involved in the potential conflict patterns to the corresponding nodes and time slices in the scheduling relationship graph; searching for a set of adjacent or connected nodes that satisfy the power replenishment path constraints based on the edge attributes and real-time status on the scheduling relationship graph, with the conflict node as the center; and integrating the conflict node, power deficit, replenishment path constraints, and the set of callable adjacent nodes to generate a scheduling task description with topological constraints.
[0012] Preferably, the step of using a machine learning model to parse the scheduling task description with topological constraints specifically includes: converting the scheduling task description with topological constraints into a structured model input containing node feature vectors, edge feature matrices, and global constraint vectors; feeding the structured model input into a pre-trained graph neural network and sequence decision combination model; the graph neural network and sequence decision combination model traversing feasible paths in the scheduling relationship graph and evaluating the impact of different power allocation sequences on network topology stability; and outputting the power adjustment amount, adjustment timing, and start / stop commands for specific charging piles specified for each relevant charging station, thus constituting a collaborative scheduling scheme.
[0013] Preferably, the specific content of the coordinated scheduling scheme further includes: specifying a source of power supplementation for charging stations with power shortages, including adjacent light-load charging stations, regional energy storage units, or specific lines of the power grid; specifying specific power output values and time windows for charging stations or units that serve as power supplementation sources; and allocating detailed power limits or start / stop time points for each charging pile within the charging station affected by the scheduling to ensure load balance within the station.
[0014] Preferably, the step of assigning real-time state attributes to each node and edge in the graph to form a dynamically weighted scheduling relationship graph specifically includes: assigning real-time state attributes to each node in the graph, including current available power, current number of vehicles waiting in the queue, and average service time within the station; assigning real-time state attributes to each edge in the graph, including current available line transmission capacity and current actual transmission power value; and dynamically calculating and updating the weight values of each edge in the graph based on the real-time state attributes of the nodes and edges, wherein the weight values are negatively correlated with the available line capacity and positively correlated with the actual transmission power, thereby forming a dynamically weighted scheduling relationship graph.
[0015] Preferably, the step of feeding the structured model input into the pre-trained graph neural network and sequence decision combination model specifically includes: the graph neural network and sequence decision combination model being pre-trained using historical scheduling data and corresponding network topology state data; the graph neural network receiving the structured model input containing node feature vectors and edge feature matrices, embedding and aggregating the features of nodes and edges in the scheduling relationship graph, and outputting a latent vector representation reflecting the state of each node in the current topology state; the sequence decision model receiving the latent vector representation output by the graph neural network and the global constraint vector, and gradually outputting a sequence of power redistribution instructions that satisfy the topology constraints through sequence generation, which is the cooperative scheduling scheme.
[0016] Compared with existing technologies, the beneficial effects of this invention are as follows: Raw parameters collected from terminals such as charging piles and transformers are structurally decomposed according to three levels: individual equipment, site aggregation, and regional statistics. Asynchronous data is calibrated and aligned based on timestamps to form a time-synchronized hierarchical parameter sequence. This eliminates the temporal disorder caused by differences in collection frequency and time among multi-source data, constructing a collaborative representation of equipment status, site load, and regional supply and demand under a unified time benchmark. It achieves a precise and consistent description of the charging network's operational status from micro to macro levels, providing a reliable multi-dimensional data view for identifying real supply and demand imbalances.
[0017] A scheduling relationship graph reflecting the electrical connections between charging stations and feasible power paths is constructed. Identified demand conflict points are mapped to this graph, generating scheduling task descriptions that clearly define source points, sink points, and feasible transmission paths. Power grid physical topology constraints and line capacity limitations are directly embedded into the problem definition. This enables the subsequent intelligent decision-making model to automatically avoid electrically infeasible scheduling schemes within the solution space, ensuring that the generated power redistribution commands and equipment start / stop commands strictly conform to the actual operating conditions of the power grid, achieving safe and feasible scheduling under complex physical constraints. Attached Figure Description
[0018] Figure 1 This is a timing diagram of the intelligent scheduling system for traffic charging stations that integrates machine learning, as described in this invention.
[0019] Figure 2 This is a flowchart for asynchronous alignment processing based on timestamps.
[0020] Figure 3 A flowchart for constructing a scheduling relationship graph.
[0021] Figure 4 This is a graph showing the relationship between the edge weights of the scheduling path and the actual transmission power.
[0022] Figure 5 A bar chart comparing the power utilization rate of charging stations before and after scheduling. Detailed Implementation
[0023] 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.
[0024] Please see Figure 1This invention provides an intelligent scheduling system for traffic charging stations that integrates machine learning. The system includes: a data acquisition and preprocessing module that acquires raw monitoring records containing multiple charging station operating parameters, performs hierarchical decomposition of the parameters in the raw monitoring records into equipment-level parameters, station-level aggregated parameters, and regional-level statistical parameters, and performs asynchronous alignment processing on the decomposed equipment-level parameters, station-level aggregated parameters, and regional-level statistical parameters based on timestamps to form a time-synchronized hierarchical parameter sequence; a conflict pattern identification module that identifies potential conflict patterns representing the imbalance between charging demand and supply from the time-synchronized hierarchical parameter sequence; a scheduling relationship graph construction module that constructs a scheduling relationship graph reflecting the charging station network topology and power flow path; a task description generation module that associates and maps potential conflict patterns with the scheduling relationship graph to generate a scheduling task description with topological constraints; and an intelligent scheduling decision module that uses a machine learning model to parse the scheduling task description with topological constraints and outputs a collaborative scheduling scheme containing power redistribution instructions and equipment start / stop commands.
[0025] In one embodiment of the present invention, see [reference] Figure 2 In the data acquisition and preprocessing module, the parameters of the original monitoring records are hierarchically decomposed to extract voltage, current, power, occupancy status, and temperature parameters belonging to a single charging pile, which are then classified as equipment-level parameters. The equipment-level parameters of all charging piles belonging to the same charging station are summed, averaged, and variance calculated to obtain station-level aggregated parameters describing the overall load, average utilization rate, and operational stability of the charging station. The station-level aggregated parameters of all charging stations within the target area are then summarized and statistically analyzed again to generate regional-level statistical parameters reflecting the total regional demand, load distribution, and supply-demand trends. Based on timestamps, the decomposed equipment-level parameters, station-level aggregated parameters, and regional-level statistical parameters are asynchronously aligned. Independent time axes are established for each of these parameters. Using the time axis with the highest sampling frequency as a reference, interpolation methods are used to fill in the time points for parameter sequences with lower sampling frequencies, unifying all hierarchical parameter sequences to the same and equally spaced time points, forming a time-synchronized hierarchical parameter sequence.
[0026] In practice, the raw monitoring records acquired by the data acquisition and preprocessing module contain operating parameters from multiple charging stations. These parameters may be asynchronous in time and differ in granularity. The implementation process first decomposes the parameters hierarchically, extracting voltage, current, power, occupancy status Boolean value, and temperature values belonging to a single charging pile. These values are categorized as equipment-level parameters. Specific calculations are performed on the equipment-level parameters of all charging piles belonging to the same charging station. The instantaneous power values of all charging piles are summed to obtain the station-level total power describing the overall load of the charging station. The power utilization rate of all charging piles is averaged to obtain the station-level average load rate describing the average utilization rate of the charging station. The variance of the voltage values of all charging piles is calculated to obtain the station-level voltage fluctuation value describing the operational stability of the charging station. These calculation results together constitute the station-level aggregated parameters. The station-level aggregated parameters of all charging stations in the target area are summarized and statistically analyzed again. The total station-level power of all charging stations in the area is accumulated to generate the total regional demand power reflecting the total regional demand. The distribution of the number of charging stations in different load rate intervals in the area is statistically analyzed to generate a regional load distribution histogram reflecting the load distribution. The rate of change of the total regional demand power over time is calculated to generate a regional demand trend slope reflecting the supply and demand trend. These statistical results together constitute the regional statistical parameters.
[0027] In some embodiments, the decomposed equipment-level parameters, station-level aggregated parameters, and regional-level statistical parameters are asynchronously aligned based on timestamps. A dense time axis with second intervals is established for the equipment-level parameters, a medium time axis with minute intervals is established for the station-level aggregated parameters, and a sparse time axis with ten-minute intervals is established for the regional-level statistical parameters. Using the second-level time axis of the equipment-level parameters as the highest sampling frequency benchmark, interpolation methods are used to fill in the missing time points on the medium time axis of the station-level aggregated parameters and the sparse time axis of the regional-level statistical parameters. For missing second-level time points in the station-level aggregated parameter sequence, linear interpolation is performed using the nearest preceding and following minute-level data points to fill in the missing values. The linear interpolation formula can be understood as follows:
[0028]
[0029] in: Representative at The interpolated parameter values at time 10:00. Represents the previous valid time point The parameter values, Represents the next valid time point The parameter values are then used to unify the parameter sequences at all levels into a sequence of identical and equally spaced time points, with intervals of seconds, thus forming a time-synchronized hierarchical parameter sequence.
[0030] Optionally, the sampling frequency of equipment-level parameters is the highest, followed by station-level aggregated parameters, and then regional-level statistical parameters. During asynchronous alignment, the time axis of the equipment-level parameters is always used as the reference time axis. This means that when the sampling time of station-level aggregated parameters or regional-level statistical parameters does not completely coincide with the time point of the equipment-level parameters, the interpolation method can generate estimated values for the corresponding times, ensuring that parameters at all levels have corresponding numerical records at every identical timestamp.
[0031] In one embodiment of the present invention, in the conflict pattern identification module, potential conflict patterns characterizing the imbalance between charging demand and supply are identified from the time-synchronized hierarchical parameter sequence. In the time-synchronized hierarchical parameter sequence, charging pile clusters that continuously exceed the threshold are detected in the device-level parameter sequence. In the station-level aggregated parameter sequence, charging stations whose load rate is continuously higher than the preset capacity and which are predicted to have queuing vehicles in adjacent time periods are marked. In the regional statistical parameter sequence, the time periods and geographical ranges in which the total demand power curve rises sharply while the regional power supply curve shows a stable or declining trend are located. The results of the above detection, marking and location are comprehensively defined as potential conflict patterns.
[0032] In practical implementation, the conflict pattern recognition module receives a time-synchronized hierarchical parameter sequence generated by the data acquisition and preprocessing module, and identifies potential conflict patterns representing the imbalance between charging demand and supply from this sequence. Within the time-synchronized hierarchical parameter sequence, it detects charging pile clusters that continuously exceed a threshold in the device-level parameter sequence. The implementation process sets the power threshold to 95% of the charging pile's rated power. When the power value of a single charging pile exceeds the power threshold for six consecutive sampling periods, it is determined that the charging pile is in a continuous over-threshold operation state. Multiple charging piles that are physically adjacent and simultaneously in a continuous over-threshold operation state are marked as a charging pile cluster. In the station-level aggregated parameter sequence, charging stations with a load rate that is consistently higher than the preset capacity and with predicted queuing vehicles in adjacent time periods are marked. The implementation process sets the station-level load rate threshold to 85% of the rated total capacity of the charging station. When the station-level average load rate of a charging station is higher than the station-level load rate threshold for three consecutive aggregated time windows, it is determined that the load rate is consistently higher than the preset capacity. At the same time, based on the historical vehicle arrival data and current time characteristics of the charging station, the number of vehicles arriving at the charging station in the next time period is predicted. When the predicted number of queuing vehicles is greater than the number of available charging piles, it is determined that there is a predicted queuing vehicle in adjacent time periods. Charging stations that meet the above two conditions are marked.
[0033] In some embodiments, within a regional statistical parameter sequence, the time period and geographical range during which the total demand power curve rises sharply while the regional power supply curve shows a stable or declining trend are identified. The implementation process calculates the rate of change of the regional total demand power within a continuous time window.
[0034]
[0035] in: This represents the total regional power demand value at the current moment. This represents the total regional power demand value at the previous moment. Represents a fixed time interval between two moments, when When the value exceeds a preset threshold for a sharp increase, the regional total demand power curve is determined to be in a state of sharp increase. This can be understood as the regional power supply curve being obtained from grid input data; when the rate of change of regional power supply within the same time window... The absolute value is less than a very small stability threshold, or When the value is negative, it indicates that the regional power supply curve is showing a stable or declining trend. The continuous time window that simultaneously satisfies the conditions of a sharp increase in regional total power demand and a stable or declining regional power supply, along with the geographical range of all charging stations covered by it, is identified as a potential supply-demand trend conflict.
[0036] Optionally, the results of the above detection, marking, and location are collectively defined as potential conflict patterns. A complete potential conflict pattern record includes the following elements: a list of detected charging pile cluster identifiers and their over-threshold durations; marked high-load charging station identifiers and their predicted queuing start times; and located supply-demand trend conflict periods and their geographical polygon coordinates. It can be understood that detecting device-level parameter sequences focuses on overload risks at the device level, marking station-level aggregated parameter sequences focuses on service capacity saturation at the station level, and locating regional-level statistical parameter sequences focuses on macro-level supply-demand imbalances. The outputs of these three operations characterize the pressure points in system operation from different dimensions. The comprehensive definition process correlates and integrates these discrete pressure point information to form a complete description of potential conflicts.
[0037] In one embodiment of the present invention, see [reference] Figure 3In the scheduling relationship graph construction module, a scheduling relationship graph reflecting the charging station network topology and power flow path is constructed. Each charging station is used as a graph node, and node attributes include the charging station's location, rated total power, current available power, and a list of connected piles. The power distribution network line connections, the geographical adjacency between charging stations, and the shared backup energy storage unit connections are used as graph edges, with edge attributes including line capacity, physical distance, and connection type. Real-time state attributes are assigned to each node and edge in the graph, forming a dynamically weighted scheduling relationship graph. Each node is assigned real-time state attributes including current available power, current number of vehicles waiting in the queue, and average service time within the station. Each edge is assigned real-time state attributes including current available transmission capacity and current actual transmission power. Based on the real-time state attributes of nodes and edges, the weight values of each edge in the graph are dynamically calculated and updated. The weight values are negatively correlated with available line capacity and positively correlated with actual transmission power, forming a dynamically weighted scheduling relationship graph.
[0038] In practical implementation, the scheduling relationship graph construction module constructs a scheduling relationship graph reflecting the network topology of charging stations and the power flow path. Each charging station is used as a graph node, and the node attributes include the geographical coordinates of the charging station, the rated total power value of the charging station, the current available power value of the charging station, and a list of connected piles of the charging station. The list of connected piles records the unique identifiers of all charging piles belonging to that charging station. The graph edges are the power distribution line connection relationship, the geographical adjacency relationship between charging stations, and the shared backup energy storage unit connection relationship. The power distribution line connection relationship is determined based on the power grid topology data, the geographical adjacency relationship is determined based on the distance between charging stations being less than a set threshold, and the backup energy storage unit connection relationship is defined by a pre-configured energy storage allocation strategy. The edge attributes include the maximum carrying capacity value of the line, the physical distance value between charging station nodes, and the enumerated value of the connection type. The connection type enumerated value is used to distinguish between the three types of relationships: "power grid line", "geographical adjacency", and "energy storage link".
[0039] In some embodiments, each node and edge in the graph is assigned real-time state attributes to form a dynamically weighted scheduling relationship graph. Each node in the graph is assigned real-time state attributes including the current available power value, the current number of vehicles waiting in the queue (integer), and the average service time within the station. The current available power value is obtained in real-time from station-level aggregation parameters. The current number of vehicles waiting in the queue is obtained in real-time through vehicle detection equipment at the charging station entrance or data from the reservation system. The average service time within the station is calculated based on vehicle records that have completed charging services in the most recent hour. Each edge in the graph is assigned real-time state attributes including the current available transmission capacity value and the current actual transmission power value. The current available transmission capacity value is the remaining capacity obtained by subtracting the current actual transmission power value from the maximum carrying capacity value of the line. The current actual transmission power value is read in real-time by sensors deployed at the line measurement points.
[0040] In one embodiment of the present invention, in the task description generation module, potential conflict patterns are associated and mapped with the scheduling relationship graph to generate a scheduling task description with topological constraints. The specific charging stations, time periods, and power deficit information involved in the potential conflict patterns are mapped to the corresponding nodes and time slices in the scheduling relationship graph. On the scheduling relationship graph, with the conflict node as the center, the set of adjacent or connected nodes that satisfy the power replenishment path constraints are searched according to the edge attributes and real-time status. The conflict node, power deficit, replenishment path constraints, and the set of callable adjacent nodes are integrated and described to generate a scheduling task description with topological constraints.
[0041] In practical implementation, the task description generation module receives potential conflict patterns from the conflict pattern identification module and a dynamically weighted scheduling relationship graph from the scheduling relationship graph construction module. It then associates and maps the potential conflict patterns with the scheduling relationship graph to generate a scheduling task description with topological constraints. The specific charging station identifiers involved in the potential conflict patterns, the time period of the conflict, and the calculated power deficit information are mapped to the corresponding nodes and time slices in the scheduling relationship graph. During implementation, a unique node is located in the scheduling relationship graph based on the charging station identifier. The time period of the conflict is divided into a series of consecutive time slices and aligned with the temporal snapshot of the graph. Power deficit information is assigned as a dynamic attribute to the conflicting node.
[0042] In some embodiments, on the scheduling graph, centered on the conflict node, a set of adjacent or connected nodes satisfying power supplementation path constraints is searched based on edge attributes and real-time status. Power supplementation path constraints include the maximum allowed path hop count, the minimum available transmission capacity requirement for each edge on the path, and the upper limit of the total physical distance of the path. The search process starts from the conflict node, traverses all edges directly connected to it or indirectly connected through intermediate nodes, checks whether each potential path satisfies the above constraints, and includes nodes satisfying the constraints at all path endpoints into the set of callable adjacent nodes. It is understood that the search process excludes nodes with currently available power values of zero or negative, while prioritizing nodes connected via "grid lines" or "energy storage links," as these connections typically imply more stable and controllable power transmission capabilities. See Table 1.
[0043] Table 1: Candidate Node Information Table
[0044] Node identifier Node type Current available power (kW) Optimal path hop count Minimum available capacity for the route (kW) Total path distance (km) Feasibility assessment Station_Alpha charging station 150 1 200 1.5 8.32 Station_Beta charging station 80 2 100 3.2 1.56 Storage_Unit_01 Energy storage unit 200 1 250 0.8 15.63
[0045] It is understandable that the description of a scheduling task with topological constraints is no longer a simple requirement description, but a composite problem definition that embeds network topology, real-time status and transmission constraints. It transforms the abstract "insufficient power" problem into a specific graph computation task of finding feasible scheduling paths on a specific network graph, providing accurate and structured input for the subsequent intelligent scheduling decision module.
[0046] See Figure 4 This is a graph showing the relationship between edge weights and actual transmission power of a scheduling path. It's a data analysis chart primarily used to illustrate the relationship between edge weights and actual transmission power for different paths in a smart charging station scheduling system. For the path "Energy Storage Unit - Conflict Node": both the edge weight (≈0.9) and actual transmission power (≈150kW) are the highest, indicating sufficient available capacity and high transmission efficiency. The edge weight and actual transmission power show a positive correlation, consistent with the design logic of "positive correlation between edge weight and actual transmission power." This type of chart is typically used for path selection evaluation in smart charging station scheduling systems, helping dispatchers quickly identify the optimal path with "high weight and high transmission capacity," thus assisting in the development of power allocation plans.
[0047] In one embodiment of the present invention, in the intelligent scheduling decision module, a machine learning model is used to parse the scheduling task description with topological constraints, converting the topologically constrained scheduling task description into a structured model input containing node feature vectors, edge feature matrices, and global constraint vectors. This structured model input is fed into a pre-trained graph neural network and sequence decision combination model. This graph neural network and sequence decision combination model traverses feasible paths in the scheduling relationship graph and evaluates the impact of different power allocation sequences on network topology stability. It outputs the power adjustment amount, adjustment timing, and start / stop commands for specific charging piles specified for each relevant charging station, constituting a collaborative scheduling scheme. The collaborative scheduling scheme also includes specifying power replenishment sources for charging stations with power shortages. These sources include adjacent lightly loaded charging stations, regional energy storage units, or specific lines in the power grid. Specific power output values and time windows are specified for the charging stations or units serving as power replenishment sources. Detailed power limits or start / stop times are allocated to each charging pile within the affected charging station to ensure load balance within the station. The structured model is fed into a pre-trained graph neural network and sequential decision-making combined model. This graph neural network and sequential decision-making combined model are pre-trained using historical scheduling data and corresponding network topology state data. The graph neural network receives the structured model input containing node feature vectors and edge feature matrices, performs feature embedding and aggregation of nodes and edges in the scheduling relationship graph, and outputs latent vector representations reflecting the state of each node in the current topology state. The sequential decision-making model receives the latent vector representations output by the graph neural network and global constraint vectors, and gradually outputs a sequence of power redistribution instructions that satisfy the topology constraints through sequence generation, which is the cooperative scheduling scheme.
[0048] In practical implementation, the intelligent scheduling decision module receives a topologically constrained scheduling task description from the task description generation module and uses a machine learning model to parse the topologically constrained scheduling task description, outputting a collaborative scheduling scheme that includes power redistribution instructions and equipment start / stop commands. The topologically constrained scheduling task description is converted into a structured model input containing node feature vectors, edge feature matrices, and global constraint vectors. Node feature vectors are extracted from the node attributes and real-time status attributes of the scheduling relationship graph, including dimensions such as the rated total power of the charging station, the current available power, the current number of vehicles waiting in the queue, and the average service time within the station. The edge feature matrix is extracted from the edge attributes and real-time status attributes of the scheduling relationship graph, including dimensions such as line capacity, physical distance, connection type enumeration, current available transmission capacity of the line, current actual transmission power, and dynamic weight. The global constraint vector is extracted from the power deficit, supplementary path constraints, and time slice information in the scheduling task description.
[0049] In some embodiments, a structured model is fed into a pre-trained graph neural network and sequence decision-making combined model. The graph neural network and sequence decision-making combined model are pre-trained using historical scheduling records and corresponding historical network topology data. The training objective is to ensure that the model's output scheduling scheme maximizes charging demand while minimizing network losses and device actions in the simulation environment. The graph neural network receives the structured model input, which includes node feature vectors and edge feature matrices. It performs feature embedding and aggregation on the scheduling relationship graph, updates the hidden state of each node in the graph through a multi-layer message passing mechanism, and finally outputs a latent vector representation reflecting the state of each node under the current topology. The sequence decision-making model receives the latent vector representation output by the graph neural network and the global constraint vector separated from the structured model input. Through sequence generation, it gradually outputs a sequence of power redistribution instructions that satisfy the topology constraints in an autoregressive manner. This power redistribution instruction sequence constitutes the cooperative scheduling scheme.
[0050] Optionally, a graph neural network and sequential decision-making combined model traverses feasible paths in the scheduling graph and evaluates the impact of different power allocation sequences on network topology stability. The evaluation process is implemented through an internal value network that scores the intermediate graph state after each decision step. The scoring function is as follows:
[0051]
[0052] in: The estimated comprehensive value representing the current decision sequence, Representing the The power utilization improvement benefits of each scheduled node Representing the The power transmission loss of the line being used This represents the total number of device state transitions triggered by the entire scheduling scheme. and The preset tradeoff coefficients, and These represent the number of nodes and edges involved, respectively. It can be understood that when the model outputs each instruction, it refers to the value network's score, tending to choose the option that will lead to the final result. The action maximizes the value, thus satisfying immediate needs while ensuring the long-term stability of the network. The outputs include a power adjustment value for each relevant charging station, an adjustment timing timestamp, and a Boolean sequence of start / stop commands for specific charging piles. These outputs collectively constitute the collaborative scheduling scheme.
[0053] The collaborative scheduling scheme also includes specifying power replenishment sources for charging stations with power shortages. These sources include adjacent lightly loaded charging stations, regional energy storage units, or specific grid lines. Source information is derived from the instruction sequence output by the sequence decision model. Specific power output values and time windows are specified for the charging stations or units acting as power replenishment sources. The specific power output value is the absolute value of the power adjustment, and the time window is the start and end time defined by the adjustment timing timestamp. Detailed power limits or start / stop times are allocated to each charging pile within the affected charging station to ensure load balancing within the station. The allocation process follows the charging station's internal load balancing strategy, decomposing the total power adjustment at the charging station level to each charging pile. For charging piles requiring load reduction, a lower power limit value is set in the instruction; for charging piles requiring temporary service suspension, a specific start / stop time point is set in the instruction.
[0054] See Figure 5 This is a bar chart comparing the power utilization of charging stations before and after scheduling, demonstrating the optimization effect of intelligent scheduling on the resource utilization efficiency of charging stations. High-load stations show reduced utilization; load-balanced stations maintain stable utilization. This chart reflects the core objective of intelligent scheduling for charging stations: balancing the load of each station by reducing the power utilization of originally high-load stations (A, B, C) to avoid equipment overload, while maintaining the stability of the load-balanced station (D), ultimately improving the operational stability and equipment lifespan of the entire charging network.
[0055] 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.
[0056] Although embodiments of the invention 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 invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A smart dispatching system for transportation charging stations integrating machine learning, characterized in that, The system includes: The data acquisition and preprocessing module is used to acquire raw monitoring records containing multiple charging station operating parameters, perform hierarchical decomposition of the parameters in the raw monitoring records into equipment-level parameters, station-level aggregated parameters, and regional-level statistical parameters, and perform asynchronous alignment processing on the decomposed equipment-level parameters, station-level aggregated parameters, and regional-level statistical parameters based on timestamps to form a time-synchronized hierarchical parameter sequence. The conflict pattern identification module is used to identify potential conflict patterns representing the imbalance between charging demand and supply from a time-synchronized hierarchical parameter sequence. The scheduling relationship graph construction module is used to construct a scheduling relationship graph that reflects the network topology of charging stations and the path of power flow. The task description generation module is used to associate and map potential conflict patterns with scheduling relationship graphs to generate scheduling task descriptions with topological constraints. The intelligent scheduling decision module is used to analyze the description of scheduling tasks with topological constraints using machine learning models and output a collaborative scheduling scheme that includes power redistribution instructions and equipment start-up and shutdown commands. The construction of the scheduling relationship graph reflecting the charging station network topology and power flow paths specifically includes: Each charging station is used as a graph node, and the node attributes include the location of the charging station, the rated total power, the current available power, and the list of connected piles. The graph edges are defined by the connection relationships of power distribution network lines, the geographical adjacency relationships between charging stations, and the connection relationships of shareable backup energy storage units. The attributes of the edges include line capacity, physical distance, and connection type. Assign real-time state attributes to each node and edge in the graph to form a dynamically weighted scheduling relationship graph; The step of associating and mapping potential conflict patterns with scheduling relationship graphs to generate scheduling task descriptions with topological constraints specifically includes: Map the specific charging stations, time periods, and power deficit information involved in potential conflict modes to the corresponding nodes and time slices in the scheduling relationship graph; On the scheduling relationship graph, with the conflict node as the center, based on the edge attributes and real-time status, search for the set of adjacent or connected nodes that satisfy the power supplement path constraint; The conflicting nodes, power deficits, supplementary path constraints, and the set of callable adjacent nodes are integrated and described to generate a scheduling task description with topological constraints.
2. The intelligent dispatching system for traffic charging stations integrating machine learning according to claim 1, characterized in that, The hierarchical decomposition of parameters in the original monitoring records specifically includes: Extract voltage, current, power, occupancy status, and temperature parameters belonging to a single charging pile and classify them as equipment-level parameters; The equipment-level parameters of all charging piles belonging to the same charging station are summed, averaged, and varianced to obtain station-level aggregated parameters describing the overall load, average utilization rate, and operational stability of the charging station. The station-level aggregated parameters of all charging stations within the target area are summarized and statistically analyzed again to generate regional-level statistical parameters that reflect the total demand, load distribution, and supply and demand trends in the region.
3. The intelligent dispatching system for traffic charging stations integrating machine learning according to claim 2, characterized in that, The asynchronous alignment process for the decomposed equipment-level parameters, station-level aggregated parameters, and regional-level statistical parameters based on timestamps specifically includes: Establish independent time axes for equipment-level parameters, station-level aggregated parameters, and regional-level statistical parameters; Using the time axis with the highest sampling frequency as a reference, the parameter sequences with lower sampling frequencies are interpolated to fill in the time points. All parameter sequences at all levels are unified to the same and equally spaced time sequence, forming a time-synchronized hierarchical parameter sequence.
4. The intelligent dispatching system for traffic charging stations integrating machine learning according to claim 3, characterized in that, The identification of potential conflict patterns characterizing the imbalance between charging demand and supply from a time-synchronized hierarchical parameter sequence specifically includes: In a time-synchronized hierarchical parameter sequence, a cluster of charging piles that continuously exceeds the threshold in the device-level parameter sequence is detected; In the station-level aggregated parameter sequence, charging stations with a load rate that is consistently higher than the preset capacity and with predicted queued vehicles in adjacent time periods are marked. In the regional statistical parameter series, locate the time period and geographical range where the total demand power curve rises sharply while the regional power supply curve shows a stable or declining trend; The results of the above detection, labeling and localization are collectively defined as potential conflict patterns.
5. The intelligent dispatching system for traffic charging stations integrating machine learning according to claim 4, characterized in that, The process of using a machine learning model to parse the scheduling task description with topological constraints specifically includes: The description of a scheduling task with topological constraints is converted into a structured model input containing node feature vectors, edge feature matrices, and global constraint vectors. The structured model is fed into a pre-trained graph neural network and a sequence decision combination model. The graph neural network and sequence decision combination model traverses feasible paths in the scheduling relationship graph and evaluates the impact of different power allocation sequences on network topology stability. The output includes the power adjustment amount, adjustment timing, and start / stop commands for specific charging piles specified for each relevant charging station, forming a collaborative scheduling scheme.
6. The intelligent dispatching system for traffic charging stations integrating machine learning according to claim 5, characterized in that, The specific details of the coordinated scheduling scheme also include: Designate sources of power replenishment for charging stations with power shortages, including adjacent light-load charging stations, regional energy storage units, or specific grid lines; Specify the exact value and time window for the power output of the charging station or unit that serves as a power replenishment source; Assign detailed power limits or start / stop times to each charging pile within the charging station affected by the scheduling to ensure balanced load within the station.
7. The intelligent dispatching system for traffic charging stations integrating machine learning according to claim 6, characterized in that, The process of assigning real-time state attributes to each node and edge in the graph to form a dynamically weighted scheduling relationship graph specifically includes: Assign each node in the graph real-time status attributes including current available power, current number of vehicles waiting in the queue, and average service time within the station; Assign each edge in the graph a real-time status attribute including the current available transmission capacity of the line and the current actual transmission power value; Based on the real-time state attributes of the nodes and edges, the weight values of each edge in the graph are dynamically calculated and updated. The weight values are negatively correlated with the available line capacity and positively correlated with the actual transmission power, forming a dynamically weighted scheduling relationship graph.
8. The intelligent dispatching system for traffic charging stations integrating machine learning according to claim 7, characterized in that, The step of feeding the structured model into a pre-trained graph neural network and sequence decision combination model specifically includes: The graph neural network and sequence decision combination model is pre-trained using historical scheduling data and corresponding network topology state data. The graph neural network receives a structured model input containing node feature vectors and edge feature matrices, performs feature embedding and aggregation of nodes and edges on the scheduling relationship graph, and outputs a latent vector representation reflecting the state of each node in the current topology. The sequential decision model receives the latent vector representation and global constraint vector output by the graph neural network. Through sequence generation, it gradually outputs a sequence of power redistribution instructions that satisfy the topological constraints, which is the cooperative scheduling scheme.