A new energy vehicle charging intelligent scheduling method based on deep learning

Abstract

The application discloses a new energy vehicle charging intelligent scheduling method based on deep learning, which comprises the following steps: obtaining multi-source charging scheduling characteristics; inputting the multi-source charging scheduling characteristics into a deep learning prediction model to construct a vehicle charging demand margin; searching for predicted available resources based on a charging scheduling prediction result to obtain a continuous charging chain set; calculating a bearable margin of a chained transmission atom and updating a resource margin to obtain an enhanced resource margin; constraining a vehicle charging demand and a matchable range of the chained transmission atom according to a chained matching state to establish an enhanced Greenkhorn matching matrix; calculating a power occupation amount of the chained transmission atom, performing self-inhibition correction on a projection correction transmission kernel, and performing Greenkhorn greedy iteration to obtain an iteration matrix; and generating a new energy vehicle charging intelligent scheduling scheme. The application improves the matching precision between new energy vehicle charging demands and resources.

Description

Technical Field

[0001] This invention relates to the field of new energy vehicle charging scheduling technology, and in particular to a new energy vehicle charging intelligent scheduling method based on deep learning. Background Technology

[0002] With the continuous increase in the number of new energy vehicles, the usage frequency of public charging stations, park charging stations, commercial parking lot charging areas, and centralized charging facilities in residential communities is constantly increasing. Existing new energy vehicle charging dispatch systems typically recommend charging stations or reserve charging time slots for users based on information such as the vehicle's current location, distance to charging stations, number of available charging piles, number of vehicles in the queue, and time-of-use electricity pricing.

[0003] In existing technologies, deep learning-based charging prediction methods can predict vehicle arrival demand, charging station load, and queuing conditions to a certain extent. However, most solutions still remain at the prediction level, and subsequent scheduling usually adopts rule-based sorting, comprehensive scoring, linear programming, or ordinary heuristic optimization methods. These methods generally treat charging resources as independent sites, time slots, or charging piles, and do not fully consider the overall requirements of a vehicle's single charging task for continuous time periods, interface occupancy, power occupancy, and grid load margin. This can easily lead to scheduling results that are mathematically allocable but discontinuous or unusable in actual execution.

[0004] Furthermore, existing charging scheduling solutions typically employ queuing, re-recommendation, or simple postponement when dealing with resource shortages or load constraints, lacking a structured transfer mechanism for charging demands from unserved vehicles. When available resources at a charging station or during a specific scheduling period are insufficient, the system struggles to establish stable resource boundaries between actual charging resources, buffer resources, and unserved states. This leads to vehicle charging demands being forcibly allocated to unservable resources or directly entering an inefficient queuing state, reducing charging resource utilization and user experience.

[0005] Therefore, how to provide a deep learning-based intelligent scheduling method for new energy vehicle charging is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a deep learning-based intelligent scheduling method for new energy vehicle charging. This invention reconstructs predicted available resources into a continuous charging chain set by chaining transmission atoms, enhances the marginal capacity of resources to accommodate the charging needs of unloaded vehicles, and improves the Greenkhorn matching process by projection correction transmission kernel and load potential self-inhibition correction, so that the matching between vehicle charging needs and charging resources takes into account continuous executability, resource carrying capacity, and grid load security.

[0007] A deep learning-based intelligent scheduling method for charging new energy vehicles according to an embodiment of the present invention includes the following steps: Acquire new energy vehicle charging scheduling data, perform data preprocessing, and obtain multi-source charging scheduling characteristics; The multi-source charging scheduling features are input into a deep learning prediction model to obtain the charging scheduling prediction results and construct the vehicle charging demand margin. Based on the charging scheduling prediction results, the available resources are retrieved, the retrieval results are merged into chained transmission atoms, and the chained matching status is recorded to obtain a set of continuous charging chains. The carrying capacity of chained transmission atoms is calculated based on the continuous charging chain set. When the carrying capacity is insufficient, the uncarried vehicle charging demand is transferred to the associated buffer chain, and the resource margin is updated to obtain the enhanced resource margin. Using the vehicle charging demand margin and the enhanced resource margin as the matching margin, the matching range between the vehicle charging demand and the chained transmission atoms is constrained according to the chained matching state, and the corresponding scheduling cost is calculated. The transmission cost matrix and the projection correction transmission kernel are constructed to establish the enhanced Greenkhorn matching matrix. The power occupancy of the chained transport atom is calculated based on the enhanced Greenkhorn matching matrix and the chained matching state. Load potential energy is generated and the projection correction transport kernel is self-inhibited. Greenkhorn greedy iteration is performed using the corrected transport kernel to obtain the iteration matrix. When the iterative matrix satisfies the preset convergence condition, the matching relationship between vehicle charging demand and enhanced resource margin is determined, and an intelligent scheduling scheme for new energy vehicle charging is generated.

[0008] Optionally, obtaining the multi-source charging scheduling feature specifically includes: Collect multi-source charging operation data according to the preset scheduling cycle, and add the same scheduling cycle identifier and the same spatial area identifier to the collected data to obtain the original data of new energy vehicle charging scheduling. Using the original data of new energy vehicle charging scheduling as input, data preprocessing is performed according to the scheduling cycle identifier and spatial area identifier to obtain multi-source charging scheduling characteristics.

[0009] Optionally, the specific methods for obtaining the charging scheduling prediction results and the marginal vehicle charging demand include: The multi-source charging scheduling features are invoked, and the features corresponding to the same vehicle, the same charging station, and the same power distribution area are arranged in time sequence according to the order of the scheduling cycle to generate a time sequence feature sequence. The time-series feature sequence is fed into the deep learning prediction model, and the time change features between adjacent scheduling cycles, the supply and demand relationship features between vehicles and charging stations, and the load relationship features between charging stations and power distribution areas are extracted in sequence to obtain the charging operation status representation. Based on the charging operation status representation, a deep learning prediction model is used to generate charging scheduling prediction results. Based on the charging dispatch prediction results, the charging demand of the vehicles participating in the dispatch is homogenized and then weighted and fused according to preset weights to obtain the vehicle demand intensity of each vehicle participating in the dispatch. Based on the vehicle demand intensity, the vehicle demand intensity of each participating vehicle is normalized, and the vehicle charging demand margin is generated according to the normalization result.

[0010] Optionally, obtaining the continuous charging chain set specifically includes: Using the predicted available resources in the charging scheduling prediction results as the processing object, a resource grid is established according to the scheduling time slot and the charging interface. The resource grid that remains available in adjacent scheduling time slots without exceeding the grid load margin is marked as a chainable resource grid. For vehicles participating in the scheduling, a sliding window search is performed on the chainable resource grid between the vehicle's expected arrival time and the target departure time to select the resource window that can complete one charging task for the vehicle within a continuous scheduling time slot, thus obtaining the candidate chain window. Perform intra-chain locking processing on the candidate chain window, binding the charging interface occupation, scheduling time slot occupation, and power occupation within the same candidate chain window into an indivisible chained transmission atom; Perform inter-chain mutual exclusion marking on chained transmission atoms. When different chained transmission atoms occupy the same charging interface or the same scheduling time slot resource, record the resource conflict relationship between different chained transmission atoms and restrict chained transmission atoms with resource conflict relationship from being selected in the same matching result at the same time. Based on the vehicle chaining adaptation relationship between the vehicle and the chained transmission atom, a chaining matching state is generated; The chained transmission atoms that have undergone intra-chain locking, inter-chain mutual exclusion marking, and chained matching state constraints are aggregated into a continuous charging chain set.

[0011] Optionally, obtaining the enhanced resource margin specifically includes: Retrieve the continuous charging chain set and chain matching state, process the chain transmission atoms in the continuous charging chain set one by one, identify the time slots and power portions of the current chain transmission atom that cannot be used to carry the vehicle charging demand within the scheduling cycle from the chain matching state, and deduct the time slots and power portions that cannot be used to carry the vehicle charging demand from the predicted available resources of the current chain transmission atom to obtain the carrying capacity margin of the current chain transmission atom. Write the vehicle charging demand that matches the current chained transmission atom in the vehicle charging demand margin into the carrying capacity. The carrying capacity is occupied in the order of writing the vehicle charging demand until the carrying capacity is full or the corresponding vehicle charging demand is fully carried, thus obtaining the carried vehicle charging demand and the uncarried vehicle charging demand. When there is an unattended vehicle charging demand, continue searching for available predictable resources along the chained transmission atom that generated the current unattended vehicle charging demand. Connect the searched available predictable resources into a buffer chain and write the unattended vehicle charging demand into the buffer chain. After writing unloaded vehicle charging requests into the buffer chain, the buffer chain carrying compatibility relationship between the buffer chain and the corresponding vehicle is re-verified. Vehicle charging requests that pass the verification are retained in the buffer chain, while vehicle charging requests that fail the verification are retained as unloaded. Based on the vehicle charging needs already carried by the chained transmission atoms, the vehicle charging needs already carried by the buffer chain, and the vehicle charging needs that are retained in the uncarried state, the resource margin is updated to obtain the enhanced resource margin.

[0012] Optionally, obtaining the Greenkhorn matching matrix specifically includes: Retrieve the vehicle charging demand margin, enhanced resource margin, and chained matching status; register the vehicle charging demand margin as the row margin of the matrix; register the enhanced resource margin as the column margin of the matrix; and establish the row and column correspondence between vehicle charging demand and resource-side matching units. The row and column correspondence is verified item by item according to the chain matching state. The row and column positions that meet the chain matching state are reserved as transmittable positions, and the row and column positions that do not meet the chain matching state are set as transmittable positions. A chain transmission mask is generated based on the transmittable and transmittable positions. Within the transmittable locations defined by the chained transmission mask, the scheduling cost is calculated based on the scheduling consumption generated by the vehicle charging demand entering the resource-side matching unit. The scheduling cost is then written into the corresponding transmittable location, and the shielding cost is written into the prohibited transmission location to obtain the transmission cost matrix. The initial transmission weights between vehicle charging demand and resource-side matching units are calculated based on the transmission cost matrix to obtain the initial transmission core. Bregman-Dykstra cyclic projection is performed on the initial transmission core. The transmission weights located in the forbidden transmission positions in the initial transmission core are reset to zero, and the transmission weights located in the transmittable positions are retained as weights to be corrected. The weights to be corrected are sequentially subjected to chained bearing compression, buffer chain writing, unbearable state writing, and projection correction amount back-addition. In chained bearing compression, the transmission weights exceeding the chained transmission atom's bearable margin are separated. In buffer chain writing, the separated transmission weights are written to the buffer chain along the corresponding marginal discharge path. In unbearable state writing, the transmission weights that the buffer chain cannot bear are written to the unbearable state. In projection correction amount back-addition, the difference in transmission weights before and after the current round of processing is back-added to the weights to be corrected in the next round. This process is repeated until the transmission weight changes of two adjacent rounds meet the preset projection error threshold, thus obtaining the projection-corrected transmission core. Using the vehicle charging demand margin and the enhanced resource margin as the matching margin, and the projection-corrected transmission kernel as the kernel function, the initial matching weights between the vehicle charging demand and the resource-side matching units are generated according to the demand-side scaling, the transmission weights in the projection-corrected transmission kernel, and the resource-side scaling. Row margin calibration and column margin calibration are performed on the initial matching weights to obtain the enhanced Greenkhorn matching matrix.

[0013] Optionally, obtaining the iteration matrix specifically includes: According to the chained matching state, the matching weights in the enhanced Greenkhorn matching matrix are assigned to the scheduling slots of the chained transmission atoms. The power share occupied by vehicle charging demand in each chained transmission atom in each scheduling slot is calculated to obtain the chain slot power occupancy spectrum. The power occupancy spectrum of the chain time slot is compared with the grid load margin in the charging scheduling prediction results on a time slot basis. The power occupancy below the load safety boundary retains the carrying direction, the power occupancy reaching the load safety boundary is marked with the back pressure direction, and the power occupancy exceeding the load safety boundary is marked with the discharge direction. The load potential energy is generated according to the carrying direction, back pressure direction and discharge direction. Based on the load potential energy, the projection correction transmission core is used to maintain the transmission capacity of the transmission weight corresponding to the bearing direction, reduce the transmission capacity of the transmission weight corresponding to the back pressure direction, and separate the transmission weight corresponding to the discharge direction from the corresponding position of the chained transmission atom. The separated transmission weight is written along the buffer chain in the edge of the enhanced resource and the unbearing state to obtain the self-inhibiting transmission core. Based on the self-inhibiting transmission kernel, the row direction marginal deviation, column direction marginal deviation, and load direction marginal deviation of the enhanced Greenkhorn matching matrix are recalculated. The deviation with the largest value among the three types of marginal deviations is determined as the greedy update object. The update method of this iteration is determined according to the greedy update object. The update method includes demand compensation update, resource absorption update, and load release update. When the current iteration performs demand compensation update, the matching weight of the corresponding row is increased according to the vehicle charging demand margin. When the current iteration performs resource absorption update, the matching weight of the corresponding column is increased according to the enhanced resource margin. When the current iteration performs load release update, the matching weight of the corresponding chained transmission atom is reduced according to the load potential energy, and the reduced matching weight is transferred to the buffer chain and unloaded state via the self-inhibiting transmission kernel. For the matching weights that have completed this round of iteration, perform chained transmission mask verification, marginal bleed verification, and load capacity verification. Write the matching weights that pass the verification into the current round of iteration matrix, and re-correct the matching weights that fail the verification using the self-inhibition transmission kernel until the current round of iteration matrix meets the preset iteration error threshold, thus obtaining the iteration matrix.

[0014] Optionally, the intelligent scheduling scheme for charging new energy vehicles specifically includes: Read the iteration matrix, vehicle charging demand margin, and enhanced resource margin; calculate the marginal deviation between the row direction of the iteration matrix and the vehicle charging demand margin; calculate the marginal deviation between the column direction of the iteration matrix and the enhanced resource margin; and calculate the weight change between adjacent iteration matrices. The marginal deviation and weight change are compared with the preset convergence thresholds respectively. When the marginal deviation in the row direction, the marginal deviation in the column direction, and the weight change all meet the preset convergence thresholds, the iteration matrix is ​​determined as the target matching matrix. Based on the target matching matrix and chained matching state, the matching relationship between vehicle charging demand and enhanced resource margin is determined. The matching relationship is then converted into the corresponding continuous charging arrangement, buffer chain arrangement and unloaded state of the vehicle, and an intelligent scheduling scheme for new energy vehicle charging is generated.

[0015] The beneficial effects of this invention are: This invention generates charging scheduling prediction results through a deep learning prediction model, and constructs vehicle charging demand margins based on these results. This allows charging scheduling to no longer rely solely on the current number of available charging stations or their distance, but to reflect changes in vehicle charging demand, charging resource availability, and grid load in advance, thereby improving the foresight and demand identification accuracy of new energy vehicle charging scheduling.

[0016] This invention reconstructs predicted available resources into a continuous charging chain set by chaining the transmission of atoms and chaining the matching state. This allows vehicle charging demand to be matched with resource units that can continuously perform a charging task, avoiding the splitting of vehicle charging demand into discontinuous time slots, conflicting interfaces, or unbearable resources, thereby improving the continuous executability of the scheduling scheme and the reliability of charging resource allocation.

[0017] When the available capacity is insufficient, this invention updates the resource margin through a buffer chain and forms an enhanced resource margin, enabling unloaded vehicle charging demands to be orderly received between the real chained transmission atoms, the buffer chain, and the unloaded state. This avoids vehicle charging demands being forcibly allocated to resources with limited load or insufficient capacity, thereby improving the stability of charging scheduling in resource-scarce scenarios.

[0018] This invention establishes an enhanced Greenkhorn matching matrix by projecting and correcting the transmission kernel, and combines load potential energy to perform self-inhibition correction on the transmission kernel. This enables the Greenkhorn greedy iteration process to actively reduce the matching tendency of high-power occupied chained transmission atoms, reduce local charging station congestion and distribution network load peaks, and improve the matching accuracy between new energy vehicle charging demand and charging resources, the utilization rate of charging resources, and the grid load balancing capability. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a deep learning-based intelligent scheduling method for new energy vehicle charging proposed in this invention. Figure 2 This is a schematic diagram of the enhanced Greenkhorn matching method for a new energy vehicle charging intelligent scheduling method based on deep learning proposed in this invention. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0021] refer to Figures 1-2 A deep learning-based intelligent scheduling method for charging new energy vehicles includes the following steps: Acquire new energy vehicle charging scheduling data, perform data preprocessing, and obtain multi-source charging scheduling characteristics; The multi-source charging scheduling features are input into a deep learning prediction model to obtain the charging scheduling prediction results and construct the vehicle charging demand margin. Based on the charging scheduling prediction results, the available resources are retrieved, the retrieval results are merged into chained transmission atoms, and the chained matching status is recorded to obtain a set of continuous charging chains. The carrying capacity of chained transmission atoms is calculated based on the continuous charging chain set. When the carrying capacity is insufficient, the uncarried vehicle charging demand is transferred to the associated buffer chain, and the resource margin is updated to obtain the enhanced resource margin. Using the vehicle charging demand margin and the enhanced resource margin as the matching margin, the matching range between the vehicle charging demand and the chained transmission atoms is constrained according to the chained matching state, and the corresponding scheduling cost is calculated. The transmission cost matrix and the projection correction transmission kernel are constructed to establish the enhanced Greenkhorn matching matrix. The power occupancy of the chained transport atom is calculated based on the enhanced Greenkhorn matching matrix and the chained matching state. Load potential energy is generated and the projection correction transport kernel is self-inhibited. Greenkhorn greedy iteration is performed using the corrected transport kernel to obtain the iteration matrix. When the iterative matrix satisfies the preset convergence condition, the matching relationship between vehicle charging demand and enhanced resource margin is determined, and an intelligent scheduling scheme for new energy vehicle charging is generated.

[0022] In this embodiment, obtaining the multi-source charging scheduling feature specifically includes: Multi-source charging operation data is collected according to a preset scheduling cycle, and the collected data is marked with the same scheduling cycle identifier and the same spatial area identifier to obtain the original data of new energy vehicle charging scheduling. The multi-source charging operation data includes vehicle operation data, charging station operation data, reservation time period data, road traffic data, electricity price data and distribution network operation data. Using the raw data of new energy vehicle charging scheduling as input, data preprocessing is performed according to the scheduling cycle identifier and spatial area identifier. The data preprocessing includes time alignment of data from different sources, spatial association of vehicles, charging stations and distribution network nodes, missing value imputation, outlier removal, duplicate record merging, data unit unification, numerical data normalization, categorical data encoding, time series data sorting and multi-source feature fusion of the time-aligned and spatially associated data to obtain multi-source charging scheduling features.

[0023] In this embodiment, the specific methods for obtaining the charging scheduling prediction results and the marginal vehicle charging demand include: The multi-source charging scheduling features are invoked, and the features corresponding to the same vehicle, the same charging station, and the same power distribution area are arranged in time sequence according to the order of the scheduling cycle to generate a time sequence feature sequence. The time-series feature sequence is fed into the deep learning prediction model, and the time change features between adjacent scheduling cycles, the supply and demand correlation features between vehicles and charging stations, and the load correlation features between charging stations and distribution areas are extracted in sequence to obtain the charging operation status representation. The deep learning prediction model is a multi-source spatiotemporal charging scheduling prediction model, including an input embedding module, a time feature encoding module, a spatial correlation encoding module, a multi-source feature fusion module, a charging demand prediction branch, a charging resource prediction branch, and a power grid load prediction branch. The input embedding module is used to map multi-source charging scheduling features to a unified feature space. The time feature encoding module is used to extract the charging demand change features within continuous scheduling cycles. The spatial correlation encoding module is used to extract the spatial correlation features between vehicles, charging stations, and distribution areas. The multi-source feature fusion module is used to generate the charging operation status representation. The charging demand prediction branch outputs vehicle charging demand prediction information based on the charging operation status representation. The charging resource prediction branch outputs predicted available resource information based on the charging operation status representation. The power grid load prediction branch outputs power grid load prediction information based on the charging operation status representation. Based on the charging operation status representation, a deep learning prediction model is used to generate charging scheduling prediction results. The charging scheduling prediction results represent the changes in vehicle charging demand, charging resource availability, and grid load during the scheduling period. Based on the charging dispatch prediction results, the charging demand of the vehicles participating in the dispatch is processed to be in the same direction, and then weighted and fused according to the preset weights to obtain the vehicle demand intensity of each vehicle participating in the dispatch. The vehicle charging demand includes the target charging capacity gap, the urgency of leaving the station, the degree of waiting constraint, and the reliability of the arrival demand. Based on the vehicle demand intensity, the vehicle demand intensity of each participating vehicle is normalized, and the vehicle charging demand margin is generated according to the normalization result.

[0024] In this embodiment, obtaining the continuous charging chain set specifically includes: Using the predicted available resources in the charging scheduling prediction results as the processing object, a resource grid is established according to the scheduling time slot and the charging interface. The resource grid that remains available in adjacent scheduling time slots without exceeding the grid load margin is marked as a chainable resource grid. For vehicles participating in the scheduling, a sliding window search is performed on the chainable resource grid between the vehicle's expected arrival time and the target departure time to select the resource window that can complete one charging task for the vehicle within a continuous scheduling time slot, thus obtaining the candidate chain window. Perform intra-chain locking processing on the candidate chain window, binding the charging interface occupation, scheduling time slot occupation, and power occupation within the same candidate chain window into an indivisible chained transmission atom; Perform inter-chain mutual exclusion marking on chained transmission atoms. When different chained transmission atoms occupy the same charging interface or the same scheduling time slot resource, record the resource conflict relationship between different chained transmission atoms and restrict chained transmission atoms with resource conflict relationship from being selected in the same matching result at the same time. Based on the vehicle chaining adaptation relationship between the vehicle and the chaining transmission atom, a chaining matching state is generated. The vehicle chaining adaptation relationship includes the adaptation relationship between the vehicle arrival time and the chaining transmission atom start time, the adaptation relationship between the vehicle target departure time and the chaining transmission atom end time, the adaptation relationship between the vehicle charging demand and the power supply available by the chaining transmission atom, and the grid load margin adaptation relationship after the vehicle matches the chaining transmission atom. The chained transmission atoms that have undergone intra-chain locking, inter-chain mutual exclusion marking, and chained matching state constraints are aggregated into a continuous charging chain set.

[0025] In this embodiment, enhancing resource margins specifically includes: Retrieve the continuous charging chain set and chain matching state, process the chain transmission atoms in the continuous charging chain set one by one, identify the time slots and power portions of the current chain transmission atom that cannot be used to carry the vehicle charging demand within the scheduling cycle from the chain matching state, and deduct the time slots and power portions that cannot be used to carry the vehicle charging demand from the predicted available resources of the current chain transmission atom to obtain the carrying capacity margin of the current chain transmission atom. Write the vehicle charging demand that matches the current chained transmission atom in the vehicle charging demand margin into the carrying capacity. The carrying capacity is occupied in the order of writing the vehicle charging demand until the carrying capacity is full or the corresponding vehicle charging demand is fully carried, thus obtaining the carried vehicle charging demand and the uncarried vehicle charging demand. When there is an unattended vehicle charging demand, continue searching for available predictable resources along the chained transmission atom that generated the current unattended vehicle charging demand. Connect the searched available predictable resources into a buffer chain and write the unattended vehicle charging demand into the buffer chain. After the unloaded vehicle charging demand is written into the buffer chain, the buffer chain load adaptation relationship between the buffer chain and the corresponding vehicle is re-verified. The vehicle charging demand that passes the verification is retained in the buffer chain, and the vehicle charging demand that fails the verification is retained as unloaded. The buffer chain load adaptation relationship includes the adaptation relationship between the service time of the buffer chain and the service time of the vehicle, the adaptation relationship between the location of the buffer chain and the reachable range of the vehicle, the adaptation relationship between the power supply of the buffer chain and the vehicle charging demand, and the adaptation relationship between the power occupied by the buffer chain and the grid load margin. Based on the vehicle charging needs already carried by the chained transmission atoms, the vehicle charging needs already carried by the buffer chain, and the vehicle charging needs that are retained in the uncarried state, the resource margin is updated to obtain the enhanced resource margin.

[0026] This invention identifies the actual carrying capacity of chained transmission atoms within the scheduling cycle by using the carrying capacity margin, and introduces uncarried vehicle charging demands into the buffer chain and uncarried state. This enables the enhanced resource margin to simultaneously express real charging resources, successor buffer resources, and uncarried demands, avoiding the forced allocation of vehicle charging demands to time-slot-limited, power-limited, or load-limited resources. This improves the elasticity of resource margin construction, scheduling stability in supply-demand imbalance scenarios, and the executability of the enhanced Greenkhorn matching matrix.

[0027] In this embodiment, obtaining the Greenkhorn matching matrix specifically includes: Retrieve the vehicle charging demand margin, enhanced resource margin, and chained matching status; register the vehicle charging demand margin as the row margin of the matrix; register the enhanced resource margin as the column margin of the matrix; and establish the row and column correspondence between vehicle charging demand and resource-side matching units. The row and column correspondence is verified item by item according to the chain matching state. The row and column positions that meet the chain matching state are reserved as transmittable positions, and the row and column positions that do not meet the chain matching state are set as transmittable positions. A chain transmission mask is generated based on the transmittable and transmittable positions. Within the transmittable locations defined by the chained transmission mask, the scheduling cost is calculated based on the scheduling consumption generated by the vehicle charging demand entering the resource-side matching unit. The scheduling cost is then written into the corresponding transmittable location, and the shielding cost is written into the prohibited transmission location to obtain the transmission cost matrix. The initial transmission weights between vehicle charging demand and resource-side matching units are calculated based on the transmission cost matrix to obtain the initial transmission core. Bregman-Dykstra cyclic projection is performed on the initial transmission core. The transmission weights located in the forbidden transmission positions in the initial transmission core are reset to zero, and the transmission weights located in the transmittable positions are retained as weights to be corrected. The weights to be corrected are sequentially subjected to chained bearing compression, buffer chain writing, unbearable state writing, and projection correction amount back-addition. In chained bearing compression, the transmission weights exceeding the chained transmission atom's bearable margin are separated. In buffer chain writing, the separated transmission weights are written to the buffer chain along the corresponding marginal discharge path. In unbearable state writing, the transmission weights that the buffer chain cannot bear are written to the unbearable state. In projection correction amount back-addition, the difference in transmission weights before and after the current round of processing is back-added to the weights to be corrected in the next round. This process is repeated until the transmission weight changes of two adjacent rounds meet the preset projection error threshold, thus obtaining the projection-corrected transmission core. Using the vehicle charging demand margin and the enhanced resource margin as matching margins, and the projection-corrected transmission kernel as the kernel function, initial matching weights between vehicle charging demand and resource-side matching units are generated according to the demand-side scaling, the transmission weights in the projection-corrected transmission kernel, and the resource-side scaling. Row margin calibration and column margin calibration are performed on the initial matching weights to ensure that the row-direction matching weights satisfy the vehicle charging demand margin and the column-direction matching weights satisfy the enhanced resource margin, thus obtaining the enhanced Greenkhorn matching matrix.

[0028] This invention limits the transmittable range between vehicle charging demand and resource-side matching units by using a chained transmission mask. It also utilizes Bregman-Dykstra cyclic projection to perform chained carrying compression, buffer chain writing, uncarried state writing, and projection correction addition on the initial transmission core. This allows the projection-corrected transmission core to simultaneously adapt to the vehicle charging demand margin, the enhanced resource margin, and the chained matching state, reducing the number of unexecutable matches entering the enhanced Greenkhorn matching matrix and improving the matching accuracy, marginal consistency, and scheduling executability between vehicle charging demand and continuously bearable resources.

[0029] In this embodiment, obtaining the iteration matrix specifically includes: According to the chained matching state, the matching weights in the enhanced Greenkhorn matching matrix are assigned to the scheduling slots of the chained transmission atoms. The power share occupied by vehicle charging demand in each chained transmission atom in each scheduling slot is calculated to obtain the chain slot power occupancy spectrum. The power occupancy spectrum of the chain time slot is compared with the grid load margin in the charging scheduling prediction results on a time slot basis. The power occupancy below the load safety boundary retains the carrying direction, the power occupancy reaching the load safety boundary is marked with the back pressure direction, and the power occupancy exceeding the load safety boundary is marked with the discharge direction. The load potential energy is generated according to the carrying direction, back pressure direction and discharge direction. Based on the load potential energy, the projection correction transmission core is used to maintain the transmission capacity of the transmission weight corresponding to the bearing direction, reduce the transmission capacity of the transmission weight corresponding to the back pressure direction, and separate the transmission weight corresponding to the discharge direction from the corresponding position of the chained transmission atom. The separated transmission weight is written along the buffer chain in the edge of the enhanced resource and the unbearing state to obtain the self-inhibiting transmission core. Based on the self-inhibiting transmission kernel, the row direction marginal deviation, column direction marginal deviation, and load direction marginal deviation of the enhanced Greenkhorn matching matrix are recalculated. The deviation with the largest value among the three types of marginal deviations is determined as the greedy update object. The update method for this iteration is determined according to the greedy update object. The update methods include demand compensation update, resource absorption update, and load release update. When the current iteration performs demand compensation update, the matching weight of the corresponding row is increased according to the vehicle charging demand margin. When the current iteration performs resource absorption update, the matching weight of the corresponding column is increased according to the enhanced resource margin. When the current iteration performs load release update, the matching weight of the corresponding chained transmission atom is reduced according to the load potential energy, and the reduced matching weight is transferred to the buffer chain and unloaded state via the self-inhibiting transmission kernel. For the matching weights that have completed this round of iteration, perform chained transmission mask verification, marginal bleed verification, and load capacity verification. Write the matching weights that pass the verification into the current round of iteration matrix, and re-correct the matching weights that fail the verification using the self-inhibition transmission kernel until the current round of iteration matrix meets the preset iteration error threshold, thus obtaining the iteration matrix.

[0030] This invention couples the matching weights in the enhanced Greenkhorn matching matrix with the actual power occupancy of chained transmission atoms by using the chain time slot power occupancy spectrum, and uses load potential energy to perform self-inhibition correction on the projection correction transmission kernel. This allows the matching weights of high-load chained transmission atoms to be orderly transferred to buffer chains and unloaded states, avoiding the continuous concentration of vehicle charging demand on charging resources with loads close to the upper limit. This improves the response capability of the Greenkhorn greedy iteration process to changes in grid load and enhances the load balance, continuous executability, and resource matching stability of the scheduling scheme.

[0031] In this embodiment, the intelligent scheduling scheme for charging new energy vehicles specifically includes: Read the iteration matrix, vehicle charging demand margin, and enhanced resource margin; calculate the marginal deviation between the row direction of the iteration matrix and the vehicle charging demand margin; calculate the marginal deviation between the column direction of the iteration matrix and the enhanced resource margin; and calculate the weight change between adjacent iteration matrices. The marginal deviation and weight change are compared with the preset convergence thresholds respectively. When the marginal deviation in the row direction, the marginal deviation in the column direction, and the weight change all meet the preset convergence thresholds, the iteration matrix is ​​determined as the target matching matrix. Based on the target matching matrix and chained matching state, the matching relationship between vehicle charging demand and enhanced resource margin is determined. The matching relationship is then converted into the corresponding continuous charging arrangement, buffer chain arrangement and unloaded state of the vehicle, and an intelligent scheduling scheme for new energy vehicle charging is generated.

[0032] Example 1: To verify the feasibility of the present invention in practice, it was applied to a centralized charging operation area for new energy vehicles. This area includes public fast charging stations, commercial parking lot charging areas, park charging areas, and community centralized charging areas. During peak scheduling periods, there is a significant problem of uneven distribution of charging resources in this area. Some charging stations closer to the main roads are in a long-term queue, while some adjacent charging stations still have idle charging piles. At the same time, some distribution branches approach the load safety boundary during the centralized fast charging period, which requires the operation platform to temporarily reduce the charging power, affecting both user waiting time and charging completion time.

[0033] In this application scenario, traditional scheduling methods primarily generate recommendations based on the vehicle's current location, distance to charging stations, number of available charging piles, and number of vehicles in the queue. When a charging station displays available charging piles at a given time, the system may guide a large number of vehicles to that station simultaneously. However, due to time differences in user arrival, the station may already be in a queue when the vehicles actually arrive. Furthermore, traditional methods treat reservation time slots as independent resources, failing to consider the continuous time slots, charging interface occupancy, and power consumption required for a single charging task for the same vehicle as a whole. This can easily lead to scheduling results being allocable within the system but not being able to be executed continuously during actual charging. Especially when the distribution network load is close to the safety boundary, ordinary scheduling methods typically perform power limiting or manual adjustments after generating the scheduling results, resulting in repeated revisions of the scheduling scheme, reducing user acceptance and station operating efficiency.

[0034] In this embodiment, the operation platform first collects vehicle operation data, charging station operation data, reservation time slot data, road traffic data, electricity price data, and distribution network operation data. After data preprocessing, multi-source charging scheduling features are obtained. The deep learning prediction model outputs charging scheduling prediction results based on the multi-source charging scheduling features and constructs vehicle charging demand margins according to changes in vehicle charging demand. Subsequently, the system retrieves predicted available resources based on the charging scheduling prediction results, merges resource windows that can continuously complete a vehicle's one-time charging task into chained transmission atoms, and records the chained matching status to obtain a set of continuous charging chains. For chained transmission atoms with insufficient capacity, the system continues to retrieve predictable available resources along the corresponding chained transmission atom, writes unloaded vehicle charging demands into a buffer chain, and retains demands that cannot pass the buffer chain verification as unloaded, thereby forming an enhanced resource margin.

[0035] When establishing the matching relationship, this embodiment does not directly use the ordinary Greenkhorn algorithm to match vehicle demand with charging resources. Instead, it uses the vehicle charging demand margin and the enhanced resource margin as the matching margin. Based on the chained matching state constraint, it constructs a transmission cost matrix and a projection-corrected transmission kernel to determine the matchable range of vehicle charging demand and chained transmission atoms. The projection-corrected transmission kernel corrects the initial transmission kernel through Bregman-Dykstra cyclic projection, setting the transmission weights at forbidden positions to zero, separating transmission weights exceeding the carrying capacity of chained transmission atoms into buffer chains, and allowing transmission weights that the buffer chains cannot carry to enter an uncarried state, thereby obtaining an enhanced Greenkhorn matching matrix that better conforms to actual charging constraints. Furthermore, the system calculates the power occupancy of chained transmission atoms based on the enhanced Greenkhorn matching matrix and the chained matching state, generates load potential energy, and performs self-inhibition correction on the projection-corrected transmission kernel. This reduces the attractiveness of chained transmission atoms with loads close to the safety boundary in the Greenkhorn greedy iteration, preventing more vehicles from continuously flowing to high-load resources.

[0036] To verify the effectiveness, the method of this invention was compared with the traditional distance-first recommendation method, the current available charging station priority recommendation method, and the ordinary Greenkhorn matching method. The traditional distance-first recommendation method generates charging station recommendations based on the road distance between the vehicle's current location and the charging station, from closest to farthest. The current available charging station priority method generates charging station recommendations based on the number of currently available charging stations, from most to least. The ordinary Greenkhorn matching method performs Greenkhorn iterative matching based solely on the vehicle's charging demand margin, charging resource margin, and a fixed transmission cost matrix. Specific comparative experiments are shown in Table 1. Table 1. Performance Comparison of Different Charging Scheduling Methods in Centralized Charging Scenarios for New Energy Vehicles

[0037] As shown in Table 1, under the same condition of 4216 vehicles participating in the daily scheduling, the average waiting time of the method of this invention is 8.7 minutes, which is significantly lower than the 18.6 minutes of the distance-first recommendation method, the 15.9 minutes of the current available charging pile priority recommendation method, and the 12.8 minutes of the ordinary Greenkhorn matching method. The average waiting time during peak scheduling periods is reduced to 13.9 minutes, and the maximum number of vehicles queuing at popular stations is reduced to 18, indicating that this invention can alleviate peak charging congestion. At the same time, the on-time completion rate of charging tasks of this invention is increased to 95.8%, the average utilization rate of charging piles is increased to 82.4%, and the proportion of unexecutable scheduling is reduced to 1.7%, indicating that chained transmission atoms, enhanced resource margins, projection correction transmission kernels, and load potential self-inhibition correction can improve the matching stability between vehicle charging demand and executable charging resources. In addition, the maximum load ratio of distribution branches is reduced from 96.2% of the distance-first recommendation method to 84.7%, indicating that this invention can reduce the load impact of centralized charging on the local distribution network.

[0038] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A deep learning-based intelligent scheduling method for charging new energy vehicles, characterized in that, Includes the following steps: Acquire new energy vehicle charging scheduling data, perform data preprocessing, and obtain multi-source charging scheduling characteristics; The multi-source charging scheduling features are input into a deep learning prediction model to obtain the charging scheduling prediction results and construct the vehicle charging demand margin. Based on the charging scheduling prediction results, the available resources are retrieved, the retrieval results are merged into chained transmission atoms, and the chained matching status is recorded to obtain a set of continuous charging chains. The carrying capacity of chained transmission atoms is calculated based on the continuous charging chain set. When the carrying capacity is insufficient, the uncarried vehicle charging demand is transferred to the associated buffer chain, and the resource margin is updated to obtain the enhanced resource margin. Using the vehicle charging demand margin and the enhanced resource margin as the matching margin, the matching range between the vehicle charging demand and the chained transmission atoms is constrained according to the chained matching state, and the corresponding scheduling cost is calculated. The transmission cost matrix and the projection correction transmission kernel are constructed to establish the enhanced Greenkhorn matching matrix. The power occupancy of the chained transport atom is calculated based on the enhanced Greenkhorn matching matrix and the chained matching state. Load potential energy is generated and the projection correction transport kernel is self-inhibited. Greenkhorn greedy iteration is performed using the corrected transport kernel to obtain the iteration matrix. When the iterative matrix satisfies the preset convergence condition, the matching relationship between vehicle charging demand and enhanced resource margin is determined, and an intelligent scheduling scheme for new energy vehicle charging is generated.

2. The intelligent scheduling method for charging new energy vehicles based on deep learning according to claim 1, characterized in that, The acquisition of the multi-source charging scheduling feature specifically includes: Collect multi-source charging operation data according to the preset scheduling cycle, and add the same scheduling cycle identifier and the same spatial area identifier to the collected data to obtain the original data of new energy vehicle charging scheduling. Using the original data of new energy vehicle charging scheduling as input, data preprocessing is performed according to the scheduling cycle identifier and spatial area identifier to obtain multi-source charging scheduling characteristics.

3. The intelligent scheduling method for charging new energy vehicles based on deep learning according to claim 1, characterized in that, The specific methods for obtaining the charging scheduling prediction results and the vehicle charging demand margin include: The multi-source charging scheduling features are invoked, and the features corresponding to the same vehicle, the same charging station, and the same power distribution area are arranged in time sequence according to the order of the scheduling cycle to generate a time sequence feature sequence. The time-series feature sequence is fed into the deep learning prediction model, and the time change features between adjacent scheduling cycles, the supply and demand relationship features between vehicles and charging stations, and the load relationship features between charging stations and power distribution areas are extracted in sequence to obtain the charging operation status representation. Based on the charging operation status representation, a deep learning prediction model is used to generate charging scheduling prediction results. Based on the charging dispatch prediction results, the charging demand of the vehicles participating in the dispatch is homogenized and then weighted and fused according to preset weights to obtain the vehicle demand intensity of each vehicle participating in the dispatch. Based on the vehicle demand intensity, the vehicle demand intensity of each participating vehicle is normalized, and the vehicle charging demand margin is generated according to the normalization result.

4. The intelligent scheduling method for charging new energy vehicles based on deep learning according to claim 1, characterized in that, The specific steps to obtain the continuous charging chain set include: Using the predicted available resources in the charging scheduling prediction results as the processing object, a resource grid is established according to the scheduling time slot and the charging interface. The resource grid that remains available in adjacent scheduling time slots without exceeding the grid load margin is marked as a chainable resource grid. For vehicles participating in the scheduling, a sliding window search is performed on the chainable resource grid between the vehicle's expected arrival time and the target departure time to select the resource window that can complete one charging task for the vehicle within a continuous scheduling time slot, thus obtaining the candidate chain window. Perform intra-chain locking processing on the candidate chain window, binding the charging interface occupation, scheduling time slot occupation, and power occupation within the same candidate chain window into an indivisible chained transmission atom; Perform inter-chain mutual exclusion marking on chained transmission atoms. When different chained transmission atoms occupy the same charging interface or the same scheduling time slot resource, record the resource conflict relationship between different chained transmission atoms and restrict chained transmission atoms with resource conflict relationship from being selected in the same matching result at the same time. Based on the vehicle chaining adaptation relationship between the vehicle and the chained transmission atom, a chaining matching state is generated; The chained transmission atoms that have undergone intra-chain locking, inter-chain mutual exclusion marking, and chained matching state constraints are aggregated into a continuous charging chain set.

5. The intelligent scheduling method for charging new energy vehicles based on deep learning according to claim 1, characterized in that, The enhancement of resource margins specifically includes: Retrieve the continuous charging chain set and chain matching state, process the chain transmission atoms in the continuous charging chain set one by one, identify the time slots and power portions of the current chain transmission atom that cannot be used to carry the vehicle charging demand within the scheduling cycle from the chain matching state, and deduct the time slots and power portions that cannot be used to carry the vehicle charging demand from the predicted available resources of the current chain transmission atom to obtain the carrying capacity margin of the current chain transmission atom. Write the vehicle charging demand that matches the current chained transmission atom in the vehicle charging demand margin into the carrying capacity. The carrying capacity is occupied in the order of writing the vehicle charging demand until the carrying capacity is full or the corresponding vehicle charging demand is fully carried, thus obtaining the carried vehicle charging demand and the uncarried vehicle charging demand. When there is an unattended vehicle charging demand, continue searching for available predictable resources along the chained transmission atom that generated the current unattended vehicle charging demand. Connect the searched available predictable resources into a buffer chain and write the unattended vehicle charging demand into the buffer chain. After writing unloaded vehicle charging requests into the buffer chain, the buffer chain carrying compatibility relationship between the buffer chain and the corresponding vehicle is re-verified. Vehicle charging requests that pass the verification are retained in the buffer chain, while vehicle charging requests that fail the verification are retained as unloaded. Based on the vehicle charging needs already carried by the chained transmission atoms, the vehicle charging needs already carried by the buffer chain, and the vehicle charging needs that are retained in the uncarried state, the resource margin is updated to obtain the enhanced resource margin.

6. The intelligent scheduling method for charging new energy vehicles based on deep learning according to claim 1, characterized in that, The Greenkhorn matching matrix is ​​obtained specifically through the following steps: Retrieve the vehicle charging demand margin, enhanced resource margin, and chained matching status; register the vehicle charging demand margin as the row margin of the matrix; register the enhanced resource margin as the column margin of the matrix; and establish the row and column correspondence between vehicle charging demand and resource-side matching units. The row and column correspondence is verified item by item according to the chain matching state. The row and column positions that meet the chain matching state are reserved as transmittable positions, and the row and column positions that do not meet the chain matching state are set as transmittable positions. A chain transmission mask is generated based on the transmittable and transmittable positions. Within the transmittable locations defined by the chained transmission mask, the scheduling cost is calculated based on the scheduling consumption generated by the vehicle charging demand entering the resource-side matching unit. The scheduling cost is then written into the corresponding transmittable location, and the shielding cost is written into the prohibited transmission location to obtain the transmission cost matrix. The initial transmission weights between vehicle charging demand and resource-side matching units are calculated based on the transmission cost matrix to obtain the initial transmission core. Bregman-Dykstra cyclic projection is performed on the initial transmission core. The transmission weights located in the forbidden transmission positions in the initial transmission core are reset to zero, and the transmission weights located in the transmittable positions are retained as weights to be corrected. The weights to be corrected are sequentially subjected to chained bearing compression, buffer chain writing, unbearable state writing, and projection correction amount back-addition. In chained bearing compression, the transmission weights exceeding the chained transmission atom's bearable margin are separated. In buffer chain writing, the separated transmission weights are written to the buffer chain along the corresponding marginal discharge path. In unbearable state writing, the transmission weights that the buffer chain cannot bear are written to the unbearable state. In projection correction amount back-addition, the difference in transmission weights before and after the current round of processing is back-added to the weights to be corrected in the next round. This process is repeated until the transmission weight changes of two adjacent rounds meet the preset projection error threshold, thus obtaining the projection-corrected transmission core. Using the vehicle charging demand margin and the enhanced resource margin as the matching margin, and the projection-corrected transmission kernel as the kernel function, the initial matching weights between the vehicle charging demand and the resource-side matching units are generated according to the demand-side scaling, the transmission weights in the projection-corrected transmission kernel, and the resource-side scaling. Row margin calibration and column margin calibration are performed on the initial matching weights to obtain the enhanced Greenkhorn matching matrix.

7. The intelligent scheduling method for charging new energy vehicles based on deep learning according to claim 1, characterized in that, The specific steps to obtain the iteration matrix include: According to the chained matching state, the matching weights in the enhanced Greenkhorn matching matrix are assigned to the scheduling slots of the chained transmission atoms. The power share occupied by vehicle charging demand in each chained transmission atom in each scheduling slot is calculated to obtain the chain slot power occupancy spectrum. The power occupancy spectrum of the chain time slot is compared with the grid load margin in the charging scheduling prediction results on a time slot basis. The power occupancy below the load safety boundary retains the carrying direction, the power occupancy reaching the load safety boundary is marked with the back pressure direction, and the power occupancy exceeding the load safety boundary is marked with the discharge direction. The load potential energy is generated according to the carrying direction, back pressure direction and discharge direction. Based on the load potential energy, the projection correction transmission core is used to maintain the transmission capacity of the transmission weight corresponding to the bearing direction, reduce the transmission capacity of the transmission weight corresponding to the back pressure direction, and separate the transmission weight corresponding to the discharge direction from the corresponding position of the chained transmission atom. The separated transmission weight is written along the buffer chain in the edge of the enhanced resource and the unbearing state to obtain the self-inhibiting transmission core. Based on the self-inhibiting transmission kernel, the row direction marginal deviation, column direction marginal deviation, and load direction marginal deviation of the enhanced Greenkhorn matching matrix are recalculated. The deviation with the largest value among the three types of marginal deviations is determined as the greedy update object. The update method of this iteration is determined according to the greedy update object. The update method includes demand compensation update, resource absorption update, and load release update. When the current iteration performs demand compensation update, the matching weight of the corresponding row is increased according to the vehicle charging demand margin. When the current iteration performs resource absorption update, the matching weight of the corresponding column is increased according to the enhanced resource margin. When the current iteration performs load release update, the matching weight of the corresponding chained transmission atom is reduced according to the load potential energy, and the reduced matching weight is transferred to the buffer chain and unloaded state via the self-inhibiting transmission kernel. For the matching weights that have completed this round of iteration, perform chained transmission mask verification, marginal bleed verification, and load capacity verification. Write the matching weights that pass the verification into the current round of iteration matrix, and re-correct the matching weights that fail the verification using the self-inhibition transmission kernel until the current round of iteration matrix meets the preset iteration error threshold, thus obtaining the iteration matrix.

8. The intelligent scheduling method for charging new energy vehicles based on deep learning according to claim 1, characterized in that, The intelligent scheduling scheme for charging new energy vehicles specifically includes: Read the iteration matrix, vehicle charging demand margin, and enhanced resource margin; calculate the marginal deviation between the row direction of the iteration matrix and the vehicle charging demand margin; calculate the marginal deviation between the column direction of the iteration matrix and the enhanced resource margin; and calculate the weight change between adjacent iteration matrices. The marginal deviation and weight change are compared with the preset convergence thresholds respectively. When the marginal deviation in the row direction, the marginal deviation in the column direction, and the weight change all meet the preset convergence thresholds, the iteration matrix is ​​determined as the target matching matrix. Based on the target matching matrix and chained matching state, the matching relationship between vehicle charging demand and enhanced resource margin is determined. The matching relationship is then converted into the corresponding continuous charging arrangement, buffer chain arrangement and unloaded state of the vehicle, and an intelligent scheduling scheme for new energy vehicle charging is generated.

Images