Deep learning-based charging pile power dynamic allocation method and system

Abstract

The present application relates to the technical field of charging pile, especially to a charging pile power dynamic allocation method and system based on deep learning. Real-time operation data, historical load data and power grid capacity constraints of charging piles in the charging station are obtained, representation learning is performed through a deep residual learning model to output global state features and individual power demand prediction, the charging piles are decoupled into multiple power management partitions accordingly, a hierarchical power decision framework is established, the lower layer decision maker defines the partition power allocation boundary and generates a power allocation candidate scheme, the upper layer coordinator tracks the cross-partition power conflict and constructs a cross-partition power transfer strategy reconstruction candidate scheme, and finally the power adjustment instruction is distributed and the load data is fed back. Efficient dynamic allocation of charging pile power is realized, and the operation stability of the charging station and the adaptability of the power grid are improved.

Description

Technical Field

[0001] This invention relates to the field of charging pile technology, and in particular to a method and system for dynamic power allocation of charging piles based on deep learning. Background Technology

[0002] In existing charging pile power allocation technologies, conventional practices mainly rely on preset rules or centralized optimization algorithms. Preset rules include fixed power limits, round-robin scheduling, or time-sharing allocation based on charging priorities. For example, charging piles are allocated fixed power sequentially according to their numbers, or priority levels are set based on the vehicle's remaining battery power. Centralized optimization algorithms, on the other hand, establish mathematical models, such as linear programming or dynamic programming, to solve for the power allocation scheme for each charging pile with the goal of minimizing the total charging time or maximizing power utilization. Some schemes also introduce traditional machine learning models, such as support vector machines or random forests, to predict short-term loads and adjust the allocation strategy. These methods typically perform unified scheduling for the entire charging station, ignoring the local coupling characteristics between charging piles and the temporal continuity of charging behavior.

[0003] Existing conventional practices have significant drawbacks. In electric vehicle charging stations where charging demand is highly dynamic, user arrival times, dwell times, and charging power demands exhibit strong randomness and nonlinearity, making it difficult for preset rules to respond in real-time to sudden load changes. For example, when multiple vehicles with high power demands connect simultaneously, simple polling or priority rules can easily lead to some charging piles remaining idle for extended periods, while other charging piles are unable to meet the demand due to insufficient power, resulting in resource waste and low charging efficiency. While centralized optimization algorithms can achieve global optima, their computational complexity increases exponentially with the number of charging piles. In large charging stations (such as those with dozens or even hundreds of charging piles), solving the optimization problem in real-time requires substantial computing power and time, failing to meet the real-time requirements of millisecond-level power adjustment. Furthermore, these algorithms rely on accurate prediction models and real-time data collection across the entire network. However, in practice, data communication delays, noise interference, and prediction errors can cause allocation schemes to deviate from reality, triggering frequent adjustments, exacerbating charging station load fluctuations, and increasing equipment wear. These shortcomings limit the practical application effectiveness of existing conventional practices in large-scale, highly dynamic charging scenarios. Summary of the Invention

[0004] This invention provides a method and system for dynamic power allocation of charging piles based on deep learning, which can solve the problems in the prior art.

[0005] A first aspect of this invention provides a deep learning-based method for dynamic power allocation in charging stations, comprising:

[0006] Obtain real-time operating data, historical load data, and grid capacity constraint information for all charging piles within the charging station;

[0007] A deep residual learning model is constructed to perform representation learning on the real-time operating data and the historical load data. Cross-layer feature paths are established in the deep network to prevent gradient degradation, and global state features and individual power demand predictions for each charging pile are output.

[0008] Based on the global state characteristics and the grid capacity constraint information, the charging pile group is decoupled into multiple power management zones, and a hierarchical power decision framework is established. The lower-level decision-maker defines the power allocation boundary of the corresponding zone based on the global state characteristics, and generates power allocation candidate schemes by coupling the individual power demand prediction of each charging pile in the corresponding zone.

[0009] The upper-level coordinator tracks the cross-partition power conflict pattern between the power allocation candidate schemes of each partition based on the global state characteristics. When a power conflict is triggered or the total load exceeds the capacity constraint, the coordinator uses the global state characteristics to construct a cross-partition power transfer strategy and reconstruct the power allocation candidate schemes of each partition, and outputs the final power allocation scheme.

[0010] The power adjustment command corresponding to the final power allocation scheme is distributed to each charging pile, and the load feedback data after execution is fed back to the deep residual learning model and the upper-level coordinator to drive the strategy evolution.

[0011] A deep residual learning model is constructed to perform representation learning on the real-time operating data and the historical load data. Cross-layer feature pathways are established in the deep network to prevent gradient degradation. The output includes global state features and individual power demand predictions for each charging pile:

[0012] The real-time running data and historical load data are segmented into time-series windows, and after time-series alignment and feature dimension standardization, the deep residual learning model is driven to start the encoding process of the feature extraction layer.

[0013] In the feature extraction layer, feature transformation is performed layer by layer through the residual unit sequence. Each residual unit allows the features of the preceding layer to directly penetrate to the subsequent layer through the cross-layer feature path and performs semantic co-coding with the transformed features of the subsequent layer to construct a fusion feature representation that carries multi-level semantic information.

[0014] The fused feature representation is injected into the state decoupling module. By hierarchically identifying the contribution of different levels of features in the fused feature representation, the global state features representing the overall operation mode of the charging station and the individual state features representing the load fluctuation mode of each charging pile are parsed out.

[0015] Based on the global state features, a time-series evolution trajectory is constructed. The decision boundary of the power prediction module is defined according to the time-series evolution trajectory. The power prediction module integrates the individual state features and the time-series evolution trajectory and generates individual power demand predictions for each charging pile through a nonlinear mapping relationship.

[0016] Based on the global state features, a time-series evolution trajectory is constructed. The decision boundary of the power prediction module is defined according to the time-series evolution trajectory. The power prediction module then fuses the individual state features and the time-series evolution trajectory and generates individual power demand predictions for each charging pile through a nonlinear mapping relationship, including:

[0017] The global state characteristics are decomposed into a time-series dimension to extract a time-series base pattern that characterizes the load evolution law of the charging station. Based on the phase change and amplitude fluctuation characteristics of the time-series base pattern, the state evolution path of future periods is deduced, and a time-series evolution trajectory containing the evolution direction and evolution rate is constructed.

[0018] The decision boundary of the power prediction module is determined based on the evolution direction and evolution rate in the time-series evolution trajectory. The decision boundary is formed by converting the evolution rate into a power change rate constraint and combining it with the evolution direction to define the reachable region of the prediction space, thereby forming a boundary constraint domain that dynamically shrinks or expands with the time-series evolution trajectory.

[0019] Individual state features are mapped to the boundary constraint domain. The relative position of the individual state features within the boundary constraint domain is converted into a power prediction value through a nonlinear mapping relationship. The nonlinear mapping relationship adjusts the mapping curvature according to the degree of deviation between the individual state features and the time-series evolution trajectory. When the degree of deviation increases, the mapping curvature is enhanced to make the prediction value converge to the time-series evolution trajectory, thereby generating individual power demand predictions for each charging pile that are limited by the boundary constraint domain and coordinated with the time-series evolution trajectory.

[0020] The lower-level decision-maker defines the power allocation boundary of the corresponding partition based on the global state features, and generates power allocation candidate schemes by coupling the individual power demand predictions of each charging pile in the corresponding partition, including:

[0021] The state subspace of the corresponding partition is extracted from the global state features. Based on the coupling relationship between load saturation and power fluctuation margin in the state subspace, the morphological features of the power allocation boundary are determined. The morphological features form a dynamic antagonistic balance through the regulation of the boundary contraction degree by the load saturation and the release of the boundary expansion potential by the power fluctuation margin. An adaptive boundary constraint domain containing the morphological evolution trajectory is constructed.

[0022] The individual power demand predictions of each charging pile in the corresponding partition are mapped to the topology of the adaptive boundary constraint domain. By analyzing the spatial distribution characteristics of the demand of each charging pile in the topology, demand clustering regions and demand sparse regions are identified. The boundary pressure distribution is quantified according to the edge proximity of the demand clustering region to the adaptive boundary constraint domain. The priority response sequence of power allocation is determined according to the boundary pressure distribution.

[0023] Power quotas are iteratively allocated according to the priority response sequence. After each allocation, the remaining available space of the adaptive boundary constraint domain is reconstructed based on the change of the boundary pressure distribution. The geometric features of the remaining available space are fed back to the next allocation process to form a constraint transmission chain. When the distance of the geometric center of the remaining available space from the initial position exceeds the offset threshold determined by the morphological evolution trajectory, quota compression is triggered to generate a power allocation candidate scheme that takes into account both the priority of demand response and the integrity of boundary constraints.

[0024] The upper-level coordinator tracks cross-partition power conflict patterns among the power allocation candidate schemes of each partition based on the global state characteristics. When a power conflict is triggered or the total load exceeds the capacity constraint, the coordinator uses the global state characteristics to construct a cross-partition power transfer strategy and reconstruct the power allocation candidate schemes of each partition, outputting the final power allocation scheme, including:

[0025] High-order coupling correlations of power demand in each partition are extracted from global state features. A conflict topology hypergraph reflecting the power competition game relationship between partitions is constructed based on the high-order coupling correlations. The source of cross-partition power conflict mode is located by identifying abnormal regions of the hyperedge weight gradient in the conflict topology hypergraph. The propagation path of the cross-partition power conflict mode from the source to other partitions is tracked based on the distribution characteristics of the hyperedge weight gradient in the conflict topology hypergraph.

[0026] When the cross-zone power conflict mode is triggered or the total load exceeds the capacity constraint, the urgency phase field of power demand is analyzed based on the phase difference of the temporal evolution of each zone in the global state characteristics. The power transfer potential energy gradient is generated by coupling the urgency phase field with the topological manifold structure of the conflict topology hypergraph, and a power transfer strategy is constructed to drive the power to flow across zones along the downward direction of the power transfer potential energy gradient.

[0027] According to the power transfer strategy, the power allocation candidate schemes of each partition are iteratively reconstructed, and the final power allocation scheme that takes into account the topological manifold structure invariance of the conflicting topological hypergraph and the total load capacity constraint is output.

[0028] Based on the coupling of the urgency phase field with the topological manifold structure of the conflict topology hypergraph, a power transfer potential energy gradient is generated. A power transfer strategy is constructed to drive power flow across partitions along the descending direction of this power transfer potential energy gradient, including:

[0029] The urgency phase field is embedded into the fiber bundle structure of the conflict topological hypergraph. By maintaining the symplectic structure invariance of phase evolution, a gauge covariance relationship between phase dynamics and topological geometry is established in the fiber coordinate system of the topological manifold structure. Based on the gauge covariance relationship, the potential energy functional of each partition node is constructed. According to the potential energy functional, the evolution trajectory of the phase dynamics in the fiber coordinate system is coupled with the curvature characteristics of the topological geometry.

[0030] The covariant derivative of the potential functional is performed along the fiber bundle cross section of the topological manifold. The gauge invariant variation of the potential functional is calculated in the tangent bundle space of the fiber bundle cross section by a parallel translation operator. The gauge invariant variation is mapped to the horizontal distribution corresponding to each hyperedge through the horizontal lifting of the connection to obtain the power transfer potential gradient. A power transfer strategy is constructed based on the power transfer potential gradient.

[0031] Distributing power adjustment instructions corresponding to the final power allocation scheme to each charging pile, and feeding back the executed load feedback data to the deep residual learning model and the upper-layer coordinator to drive policy evolution includes:

[0032] According to the final power allocation scheme, a power adjustment instruction carrying a power adjustment trajectory is generated for each charging pile. The power adjustment instruction is distributed to the corresponding charging pile and load feedback data is collected during the execution process. The load feedback data includes the deviation trajectory between the actual power output value and the power adjustment trajectory.

[0033] Based on the deviation trajectory, a policy response manifold characterizing the execution dynamics is constructed, and a residual manifold embedding reflecting the prediction error propagation mode is extracted by analyzing the local geometric properties of the policy response manifold.

[0034] The residual manifold is embedded and fed back to the deep residual learning model to correct the manifold structure of the residual mapping. The policy response manifold is fed back to the upper coordinator to reconstruct the decision space geometry of the cross-regional power transfer policy. The policy evolution is driven based on the corrected manifold structure, the reconstructed decision space geometry, and the curvature information of the shared policy response manifold.

[0035] A second aspect of this invention provides a deep learning-based dynamic power allocation system for charging piles, comprising:

[0036] The data acquisition unit is used to acquire real-time operating data, historical load data, and grid capacity constraint information of all charging piles in the charging station;

[0037] The predictive modeling unit is used to construct a deep residual learning model to perform representation learning on the real-time operating data and the historical load data. It prevents gradient degradation by establishing cross-layer feature paths in the deep network and outputs global state features and individual power demand predictions for each charging pile.

[0038] The partition decision unit is used to decouple the charging pile group into multiple power management partitions based on the global state characteristics and the grid capacity constraint information, establish a hierarchical power decision framework, and the lower-level decision unit defines the power allocation boundary of the corresponding partition based on the global state characteristics, and couples the individual power demand prediction of each charging pile in the corresponding partition to generate power allocation candidate schemes.

[0039] The power coordination unit is used by the upper-layer coordinator to track the cross-partition power conflict mode between the power allocation candidate schemes of each partition based on the global state characteristics. When a power conflict is triggered or the total load exceeds the capacity constraint, the unit uses the global state characteristics to construct a cross-partition power transfer strategy and reconstruct the power allocation candidate schemes of each partition, and outputs the final power allocation scheme.

[0040] The instruction feedback unit is used to distribute the power adjustment instruction corresponding to the final power allocation scheme to each charging pile, and to feed back the executed load feedback data to the deep residual learning model and the upper-level coordinator to drive the strategy evolution.

[0041] A third aspect of the present invention provides an electronic device, comprising:

[0042] processor;

[0043] Memory used to store processor-executable instructions;

[0044] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0045] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0046] The deep residual learning model effectively prevents gradient degradation through cross-layer feature pathways, enabling deep representation learning of real-time operation data and historical load data of charging piles. This significantly improves the accuracy of global state feature extraction and the precision of individual power demand prediction. The model maintains gradient stability in the deep network, avoiding the information loss problem of traditional deep networks, thus providing a more reliable prediction basis for subsequent power allocation and adapting to the complex nonlinear changes in charging station load.

[0047] The hierarchical power decision-making framework decouples the charging pile cluster into multiple power management zones, significantly reducing the decision-making complexity of global coordination. The lower-level decision-maker accurately defines the power allocation boundaries of each zone based on global state characteristics, coupling individual demand predictions to generate candidate solutions. This makes the power allocation of each zone more closely aligned with actual operational constraints, reducing computational overhead and improving response speed. The local optimization after partitioning avoids direct global optimization calculations for all charging piles, improving real-time allocation efficiency.

[0048] The upper-level coordinator uses global state characteristics to track cross-regional power conflict patterns in real time. Once a power conflict is triggered or the total load exceeds the grid capacity, it immediately constructs a cross-regional power transfer strategy and reconstructs candidate schemes to ensure consistent power allocation across regions and avoid local overload or idleness. The load feedback data after execution continuously drives the evolution of the model and coordinator, enabling the deep residual learning model and hierarchical decision framework to automatically adapt to load changes and grid constraint fluctuations, achieving dynamic self-optimization and long-term stable operation of power allocation. Attached Figure Description

[0049] Figure 1 This is a flowchart illustrating the deep learning-based dynamic power allocation method for charging piles according to an embodiment of the present invention.

[0050] Figure 2 This is a flowchart illustrating the dynamic power allocation process for charging piles based on deep learning, as described in an embodiment of the present invention. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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.

[0052] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0053] Figure 1 This is a flowchart illustrating the deep learning-based dynamic power allocation method for charging piles according to an embodiment of the present invention.

[0054] Deep learning-based methods for dynamic power allocation in charging stations include:

[0055] Obtain real-time operating data, historical load data, and grid capacity constraint information for all charging piles within the charging station;

[0056] A deep residual learning model is constructed to perform representation learning on the real-time operating data and the historical load data. Cross-layer feature paths are established in the deep network to prevent gradient degradation, and global state features and individual power demand predictions for each charging pile are output.

[0057] Based on the global state characteristics and the grid capacity constraint information, the charging pile group is decoupled into multiple power management zones, and a hierarchical power decision framework is established. The lower-level decision-maker defines the power allocation boundary of the corresponding zone based on the global state characteristics, and generates power allocation candidate schemes by coupling the individual power demand prediction of each charging pile in the corresponding zone.

[0058] The upper-level coordinator tracks the cross-partition power conflict pattern between the power allocation candidate schemes of each partition based on the global state characteristics. When a power conflict is triggered or the total load exceeds the capacity constraint, the coordinator uses the global state characteristics to construct a cross-partition power transfer strategy and reconstruct the power allocation candidate schemes of each partition, and outputs the final power allocation scheme.

[0059] The power adjustment command corresponding to the final power allocation scheme is distributed to each charging pile, and the load feedback data after execution is fed back to the deep residual learning model and the upper-level coordinator to drive the strategy evolution.

[0060] In one optional implementation, a deep residual learning model is constructed to perform representation learning on the real-time operating data and the historical load data. Cross-layer feature paths are established in the deep network to prevent gradient degradation, and the output of global state features and individual power demand predictions for each charging pile includes:

[0061] The real-time running data and historical load data are segmented into time-series windows, and after time-series alignment and feature dimension standardization, the deep residual learning model is driven to start the encoding process of the feature extraction layer.

[0062] In the feature extraction layer, feature transformation is performed layer by layer through the residual unit sequence. Each residual unit allows the features of the preceding layer to directly penetrate to the subsequent layer through the cross-layer feature path and performs semantic co-coding with the transformed features of the subsequent layer to construct a fusion feature representation that carries multi-level semantic information.

[0063] The fused feature representation is injected into the state decoupling module. By hierarchically identifying the contribution of different levels of features in the fused feature representation, the global state features representing the overall operation mode of the charging station and the individual state features representing the load fluctuation mode of each charging pile are parsed out.

[0064] Based on the global state features, a time-series evolution trajectory is constructed. The decision boundary of the power prediction module is defined according to the time-series evolution trajectory. The power prediction module integrates the individual state features and the time-series evolution trajectory and generates individual power demand predictions for each charging pile through a nonlinear mapping relationship.

[0065] When segmenting real-time operational data and historical load data into time-series windows, a fixed-length sliding window is used to sample the original time series in segments. The window length is determined based on the typical fluctuation cycle of the charging station load. A certain step size overlap is maintained between adjacent windows to ensure the continuity of the time-series context. Time-series alignment processing interpolates and completes the sensor data from different sampling frequencies, ensuring that the current, voltage, power, and other multi-dimensional features of each charging pile at the same timestamp strictly correspond on the time axis. Feature dimension standardization employs a normalization method based on historical statistics, mapping each feature dimension to a unified numerical range to eliminate the interference of dimensional differences on subsequent network weight learning. After completing the above preprocessing, the standardized time-series window samples are organized into tensor form, driving the deep residual learning model to initiate the encoding process of the feature extraction layer.

[0066] The feature extraction layer consists of multiple residual units connected in series, with each residual unit containing two parallel information flow paths. The main path sequentially performs convolutional transformations, batch normalization, and nonlinear activation operations on the input features, gradually extracting local temporal patterns and cross-dimensional correlations. The cross-layer feature path, acting as a bypass, directly transmits the input features of the current residual unit to its output, fusing them with the transformation results of the main path through element-wise addition. When the input and output dimensions of the main path are inconsistent, linear projection is used to match the dimensions of the features on the cross-layer feature path, ensuring dimensionality compatibility of the addition operation. This structure allows gradients to flow directly to shallower layers along the cross-layer feature path during backpropagation, effectively mitigating the degradation problem caused by gradient decay layer by layer in deep networks. As the residual unit sequence is stacked layer by layer, the network gradually transitions from low-level local statistical features to high-level abstract semantic features. The feature representation of each level is transmitted to subsequent levels through the cross-layer feature path, ensuring that deep features are always superimposed with fine-grained information from shallow layers. Finally, the feature outputs of each level are concatenated or weighted and aggregated to form a fusion feature representation that carries multi-level semantic information. This representation includes both local features reflecting short-term power fluctuations and global features reflecting long-term load trends.

[0067] After the fused feature representation is injected into the state decoupling module, it is decomposed into two semantically clear state features by hierarchically identifying the contribution of features at different levels. The hierarchical identification mechanism is based on attention weighting: attention scores are calculated for sub-features from different depth levels in the fused feature representation. The attention score reflects the contribution of the feature at that level to the description of the global operation mode. Deep abstract features with higher scores are mainly used to construct global state features, while shallow fine-grained features with higher scores are mainly used to construct individual state features. The global state features aggregate macroscopic operation information such as the overall load level, peak-valley distribution pattern, and multi-pile collaborative charging mode of the charging station. The dimensions are relatively compact, and it can represent the overall operation status of the charging station in a low-redundancy manner. Individual state features retain microscopic information such as the load fluctuation pattern, charging stage transition characteristics, and user behavior differences of each charging pile. The individual state features corresponding to different charging piles have a distinguishable distribution structure in the feature space, providing differentiated input representations for subsequent independent power demand prediction of each pile.

[0068] When constructing a time-series evolution trajectory based on global state features, the global state feature sequence of multiple consecutive time steps is input into a cyclic structure or time-series modeling unit to capture the evolution pattern of the overall operating state of the charging station in the time dimension, and output a set of time-series evolution trajectory vectors describing the future power demand trend. The time-series evolution trajectory defines the decision boundary of the power prediction module. That is, under the current global operating situation, the reasonable range and direction of change of the power demand prediction of each charging pile are constrained by the time-series evolution trajectory, avoiding individual prediction results from deviating from the overall load trend. The power prediction module receives individual state features and time-series evolution trajectories as joint inputs. Through a nonlinear mapping relationship composed of multiple fully connected networks, the individual state features and time-series evolution trajectories are cross-fused to explore the interaction dependency between them. Specifically, after concatenating the individual state features and time-series evolution trajectories in the feature dimension, they are gradually compressed to the power prediction output dimension through multiple nonlinear transformations. An activation function is applied after each transformation to introduce nonlinear expressive power. Finally, the output layer converts the feature vector into a sequence of power demand prediction values ​​for each charging pile in the prediction time domain through linear mapping. The predicted values ​​are inversely normalized to restore them to the actual power dimensions, forming individual power demand prediction results for each charging pile, which are then used by the subsequent hierarchical power decision-making framework.

[0069] During model training, real power records from historical load data are used as supervision labels. End-to-end optimization of all parameters of the deep residual learning model is performed by minimizing the regression loss between predicted and actual power. The loss function considers both the mean square and absolute value terms of the prediction error, balancing the penalty strength for large-biased samples and robustness to noisy samples. Attention weight parameters in the state decoupling module and nonlinear mapping parameters in the power prediction module both participate in gradient backpropagation, ensuring continuous improvement in the decoupling quality between global and individual state features during end-to-end optimization. As load feedback data from charging stations accumulates, incremental fine-tuning of the model is performed using new samples, enabling the model to adapt to the slow drift of charging station operation modes over time and maintain the long-term stability of individual power demand prediction accuracy.

[0070] In one optional implementation, a time-series evolution trajectory is constructed based on the global state features, the decision boundary of the power prediction module is defined according to the time-series evolution trajectory, and the power prediction module fuses the individual state features and the time-series evolution trajectory and generates individual power demand predictions for each charging pile through a nonlinear mapping relationship, including:

[0071] The global state characteristics are decomposed into a time-series dimension to extract a time-series base pattern that characterizes the load evolution law of the charging station. Based on the phase change and amplitude fluctuation characteristics of the time-series base pattern, the state evolution path of future periods is deduced, and a time-series evolution trajectory containing the evolution direction and evolution rate is constructed.

[0072] The decision boundary of the power prediction module is determined based on the evolution direction and evolution rate in the time-series evolution trajectory. The decision boundary is formed by converting the evolution rate into a power change rate constraint and combining it with the evolution direction to define the reachable region of the prediction space, thereby forming a boundary constraint domain that dynamically shrinks or expands with the time-series evolution trajectory.

[0073] Individual state features are mapped to the boundary constraint domain. The relative position of the individual state features within the boundary constraint domain is converted into a power prediction value through a nonlinear mapping relationship. The nonlinear mapping relationship adjusts the mapping curvature according to the degree of deviation between the individual state features and the time-series evolution trajectory. When the degree of deviation increases, the mapping curvature is enhanced to make the prediction value converge to the time-series evolution trajectory, thereby generating individual power demand predictions for each charging pile that are limited by the boundary constraint domain and coordinated with the time-series evolution trajectory.

[0074] When performing time-series decomposition on the global state characteristics, the global state characteristics are expanded along the time axis. An orthogonal basis decomposition method is used to decompose the time-series signal of the charging station load into a linear combination of several time-series base patterns. Each time-series base pattern corresponds to a typical load evolution law, such as the morning and evening peak-valley cycle, the switching pattern between weekdays and non-weekdays, and the pulse pattern caused by sudden high-power access. The phase change characteristics and amplitude fluctuation characteristics of each time-series base pattern are extracted: the phase change reflects the degree of lead or lag of the pattern on the time axis, while the amplitude fluctuation characterizes the strength of the pattern's contribution to the overall load. The phase change sequence and amplitude fluctuation sequence of each time-series base pattern are jointly input into the state inference module. Based on the current phase and amplitude state and combined with historical evolution trends, the state evolution path of the charging station load within several future time steps is inferred. The evolution path is presented in the form of a vector field in the time-series space. Each time step corresponds to an evolution direction vector and a scalar evolution rate, which together constitute the time-series evolution trajectory. The evolution direction vector describes the trend of load state change in the power space, while the evolution rate quantifies the speed of this trend. A larger evolution rate means that the load will change significantly in a short period of time, while a smaller evolution rate means that the load is in a relatively stable maintenance phase.

[0075] When determining the decision boundary of the power prediction module based on the evolution direction and rate in the time-series evolution trajectory, the evolution rate is converted into a power change rate constraint. Specifically, if the evolution rate at a certain time step is... Then the maximum allowable power change within the corresponding time step. Evolution rate and preset conversion coefficient Joint decision, expressed as ,in The rate-to-power conversion coefficient is determined by a combination of the charging pile hardware response characteristics and power grid safety regulations. Under the constraint of the evolution direction, the prediction space is divided into reachable and inaccessible regions: the forward half-space along the evolution direction is the reachable region, while the region in the reverse evolution direction is limited by the power change rate constraint, allowing only reverse adjustment when the evolution rate is sufficiently large. The intersection of the reachable region and the power change rate constraint is defined as the boundary constraint domain. This constraint domain contracts or expands dynamically with the evolution trajectory: when the evolution rate increases, the constraint domain expands along the evolution direction, allowing for a wider range of power prediction values; when the evolution rate decreases or the evolution direction tends to stabilize, the constraint domain contracts, limiting the prediction value to a narrower power range, thereby suppressing excessive fluctuations in the prediction results. This dynamic contraction and expansion mechanism allows the decision boundary to adaptively follow the real-time evolution of the charging station load, avoiding prediction distortion problems caused by fixed boundaries.

[0076] When mapping individual state features to the boundary constraint domain, a coordinate transformation is performed on the individual state feature vector of each charging pile, projecting it from the original feature space to a local coordinate system with the current boundary constraint domain as the reference frame. The relative position of the individual state feature in the local coordinate system reflects the relationship between the current state of the charging pile and the global temporal evolution trajectory: if the projected coordinates of the individual state feature are near the center of the boundary constraint domain, it indicates that the state of the charging pile is highly consistent with the global evolution trend; if the projected coordinates are biased towards the edge of the boundary constraint domain, it indicates that the charging pile has a tendency to deviate from the global evolution trajectory. The relative position of the individual state feature in the boundary constraint domain is converted into a power prediction value through a nonlinear mapping relationship, which is implemented using a mapping function with adjustable curvature. The mapping curvature is determined by the degree of deviation between the individual state feature and the temporal evolution trajectory. Dynamic regulation: when When the value is small, the mapping function is close to linear, and the predicted value smoothly follows the changes in individual state characteristics; when... As the curvature increases, the mapping curvature strengthens accordingly, and the mapping function exerts a stronger pull on the predicted values ​​that are far from the time-series evolution trajectory, causing the predicted values ​​to converge towards the time-series evolution trajectory. This curvature adaptive control mechanism corresponds, in a physical sense, to the coordination constraint of the overall load management of the charging station: the power demand of a single charging pile should not deviate significantly from the overall load evolution trend of the station for a long period of time, otherwise it will disrupt the power balance of the entire station.

[0077] Degree of deviation The calculation is based on the distance metric between the individual state feature vector and the temporal evolution trajectory in the feature space, and can be expressed as: ,in Let be the individual state feature vector of a certain charging pile. This is the reference vector in the feature space corresponding to the temporal evolution trajectory at the current time step. This represents the Euclidean norm. When... Exceeding the preset threshold At that time, the mapped curvature is derived from the fundamental curvature. Increased proportionally to ,in This is the curvature enhancement coefficient, which controls the strength of the pull-back force. When... Not exceeding At that time, the mapped curvature remains the basic curvature. This ensures that the prediction results respond sensitively to changes in individual state characteristics.

[0078] The final generated individual power demand predictions for each charging pile simultaneously satisfy two constraints: First, the predicted values ​​fall within the boundary constraint domain defined by the evolution rate and direction, ensuring the prediction results are physically achievable. Second, the predicted values ​​are coordinated with the temporal evolution trajectory through curvature adaptive nonlinear mapping, preventing the predicted values ​​of individual charging piles from deviating significantly from the global trend due to transient anomalies in their individual states. The combined effect of these two constraints enables the power prediction module to maintain overall consistency in power prediction across the entire station while respecting the individual differences of each charging pile, providing high-quality individual power demand prediction input for the subsequent hierarchical power decision framework.

[0079] In one optional implementation, the lower-level decision-maker defines the power allocation boundary of the corresponding partition based on the global state features, and generates power allocation candidate schemes by coupling the individual power demand predictions of each charging pile within the corresponding partition, including:

[0080] The state subspace of the corresponding partition is extracted from the global state features. Based on the coupling relationship between load saturation and power fluctuation margin in the state subspace, the morphological features of the power allocation boundary are determined. The morphological features form a dynamic antagonistic balance through the regulation of the boundary contraction degree by the load saturation and the release of the boundary expansion potential by the power fluctuation margin. An adaptive boundary constraint domain containing the morphological evolution trajectory is constructed.

[0081] The individual power demand predictions of each charging pile in the corresponding partition are mapped to the topology of the adaptive boundary constraint domain. By analyzing the spatial distribution characteristics of the demand of each charging pile in the topology, demand clustering regions and demand sparse regions are identified. The boundary pressure distribution is quantified according to the edge proximity of the demand clustering region to the adaptive boundary constraint domain. The priority response sequence of power allocation is determined according to the boundary pressure distribution.

[0082] Power quotas are iteratively allocated according to the priority response sequence. After each allocation, the remaining available space of the adaptive boundary constraint domain is reconstructed based on the change of the boundary pressure distribution. The geometric features of the remaining available space are fed back to the next allocation process to form a constraint transmission chain. When the distance of the geometric center of the remaining available space from the initial position exceeds the offset threshold determined by the morphological evolution trajectory, quota compression is triggered to generate a power allocation candidate scheme that takes into account both the priority of demand response and the integrity of boundary constraints.

[0083] After receiving the global state features, the lower-level decision-maker first extracts the state subspace of the corresponding partition. The global state features include multi-dimensional information such as the overall load distribution, voltage level, and current margin of the charging station, while the state subspace of each partition is a slice of features directly related to that partition. Within this subspace, load saturation and power fluctuation margin are two core coupled variables. Load saturation reflects the proportion of the current partition's used power to its rated capacity. When the saturation is high, the available power space tends to narrow, and the boundary shows a contraction trend. Power fluctuation margin describes the flexible margin for power adjustment within the partition. When the margin is sufficient, the boundary has the potential to expand outward. A dynamic antagonistic balance is formed between the two: the increase in saturation continuously compresses the available range of the boundary, while the release of margin provides the driving force for the elastic expansion of the boundary.

[0084] Let the load saturation be The power fluctuation margin is The morphological characteristics of the power distribution boundary can then be characterized by the coupling function between the two. Specifically, the degree of boundary contraction... and Positive correlation, potential for boundary expansion and They are positively correlated, and together they determine the current form of the adaptive boundary constraint domain. In actual calculations, they will... and The ratio is introduced into the boundary shape equation, such that when When the boundary is larger, it tends to tighten. Smaller boundary values ​​allow for more flexible accommodation. Based on this, the sequence of boundary morphological changes throughout historical time steps is recorded as a morphological evolution trajectory. This trajectory includes the history of boundary contraction and expansion under different operating states, providing a basis for the dynamic determination of subsequent offset thresholds. Finally, an adaptive boundary constraint domain incorporating the morphological evolution trajectory is constructed. This constraint domain is not a static rectangle or fixed polygon, but a dynamic topology that adjusts in real time according to the operating state.

[0085] After constructing the adaptive boundary constraint domain, the individual power demand predictions for each charging pile within the corresponding partition are mapped to the topology of this constraint domain. The individual power demand prediction for each charging pile is output by a deep residual learning model, representing the expected power required by the charging pile in the current and future time steps. These demand values ​​are projected onto the topological coordinate system of the constraint domain according to their magnitude and physical location within the partition, forming a demand distribution map. In this distribution map, demand points for multiple charging piles are densely clustered in some areas, forming demand clustered regions; while demand points are sparse or even absent in other areas, forming demand sparse regions.

[0086] The approximation degree between the demand cluster region and the edge of the adaptive boundary constraint domain is quantified to obtain the boundary pressure distribution. Let the distance from the center point of a certain demand cluster region to the edge of the constraint domain be... subscript Indicates the first In a clustered region, the boundary pressure and The pressure is inversely proportional to the edge; that is, the closer to the edge, the greater the pressure. Normalizing the boundary pressure of all clustered regions within the partition yields a normalized boundary pressure distribution vector. The components of this vector reflect the pressure exerted by different charging pile demands on the boundary of the constraint domain. Charging piles with higher boundary pressure mean that their demand is closer to the limit of the currently allocable power, and failure to respond preferentially will lead to the breach of the constraint domain. Accordingly, charging piles are arranged in descending order of boundary pressure to determine the priority response sequence for power allocation. Charging piles with higher priority response sequences receive power quotas first during the iterative allocation process.

[0087] Power quotas are iteratively allocated according to a priority response sequence. In each iteration, the charging pile with the highest priority is selected from the priority response sequence, and its power quota is allocated based on its individual power demand prediction and the available space in the current constraint domain. After one allocation, the remaining available space in the constraint domain decreases, requiring reconstruction of the geometric features of the remaining available space. The geometric features of the remaining available space include its volume (i.e., the total remaining allocable power), geometric center location, and shape profile. These geometric features are fed back as constraint information to the next allocation process, enabling subsequent charging pile allocation decisions to be aware of the impact of previous allocations on the constraint domain, thus forming a constraint propagation chain. This constraint propagation chain ensures the consistency of the entire iterative allocation process and avoids the problem of exceeding the total limit when each charging pile is allocated independently.

[0088] During the iterative allocation process, the offset of the geometric center of the remaining available space is continuously monitored. Let the initial position of the geometric center of the constraint domain be... The geometric center position of the current iteration step is The Euclidean distance between the two is The offset threshold determined by the morphological evolution trajectory is: This threshold is adaptively determined based on the statistical regularity of geometric center offset in the historical morphological evolution trajectory, reflecting the maximum allowable drift range of the geometric center of the constraint domain under the current operating state. Exceed When this occurs, it indicates that the shape of the remaining available space has changed significantly. Continuing to allocate according to the original quota will result in subsequent charging piles not receiving a reasonable response, thus triggering the quota compression mechanism.

[0089] The core of the quota compression mechanism is to proportionally or differentially compress allocated but unexecuted quotas to bring the geometric center of the remaining available space back to a reasonable range. Differential compression prioritizes charging piles with later priority response sequences and relatively ample allocated quotas, while maintaining a smaller compression range for charging piles with earlier priority response sequences and higher boundary pressures. This ensures the response quality of high-priority demands while satisfying constraint integrity. After quota compression adjustment, the total allocation is re-verified to meet the capacity limit of the constraint domain. If it does, subsequent iterations continue; if it still exceeds the limit, compression is triggered again until the constraint conditions are fully met.

[0090] The iteration process terminates when all charging piles have completed power quota allocation or the remaining available space is insufficient to support further allocations. The final power quotas obtained by each charging pile are aggregated to form a power allocation candidate scheme for that partition. This candidate scheme balances demand response priority (ensuring the needs of high-pressure charging piles through a priority response sequence) and boundary constraint integrity (ensuring the total amount does not exceed the limit through constraint propagation chains and quota compression mechanisms), providing a reliable input basis for cross-partition coordination by the upper-level coordinator. The candidate scheme also includes updated information on the morphological evolution trajectory of the current constraint domain, used for boundary morphology calculations in subsequent time steps, enabling cross-time-step state memory and continuous policy optimization.

[0091] In one optional implementation, the upper-layer coordinator tracks cross-partition power conflict patterns among the power allocation candidate schemes of each partition based on the global state characteristics. When a power conflict is triggered or the total load exceeds the capacity constraint, the coordinator uses the global state characteristics to construct a cross-partition power transfer strategy and reconstruct the power allocation candidate schemes of each partition, outputting the final power allocation scheme including:

[0092] High-order coupling correlations of power demand in each partition are extracted from global state features. A conflict topology hypergraph reflecting the power competition game relationship between partitions is constructed based on the high-order coupling correlations. The source of cross-partition power conflict mode is located by identifying abnormal regions of the hyperedge weight gradient in the conflict topology hypergraph. The propagation path of the cross-partition power conflict mode from the source to other partitions is tracked based on the distribution characteristics of the hyperedge weight gradient in the conflict topology hypergraph.

[0093] When the cross-zone power conflict mode is triggered or the total load exceeds the capacity constraint, the urgency phase field of power demand is analyzed based on the phase difference of the temporal evolution of each zone in the global state characteristics. The power transfer potential energy gradient is generated by coupling the urgency phase field with the topological manifold structure of the conflict topology hypergraph, and a power transfer strategy is constructed to drive the power to flow across zones along the downward direction of the power transfer potential energy gradient.

[0094] According to the power transfer strategy, the power allocation candidate schemes of each partition are iteratively reconstructed, and the final power allocation scheme that takes into account the topological manifold structure invariance of the conflicting topological hypergraph and the total load capacity constraint is output.

[0095] After obtaining candidate power allocation schemes for each partition, the upper-level coordinator extracts high-order coupling relationships of power demand for each partition from the global state features. High-order coupling relationships refer to power competition dependencies that go beyond simple pairings between partitions and involve three or more partitions simultaneously. Specifically, the global state feature vector is partitioned along the partition dimension, and the joint mutual information of any partition subset in the power demand dimension is calculated to quantify the intensity of synchronous competition among multiple partitions. Unlike traditional graph structures where each edge connects only two nodes, hypergraphs allow a single hyperedge to connect any number of nodes, thus naturally expressing the collective coupling relationships of power competition among multiple partitions. A conflict topology hypergraph is constructed using each partition as a node and high-order coupling relationships as hyperedges. The weight of the hyperedge is determined by the joint power competition intensity of the corresponding partition subset; a higher weight indicates a greater conflict potential in power allocation for that partition combination.

[0096] After the conflict topology hypergraph is established, the excitation source of power conflict is located by calculating the gradient field of the weights of each hyperedge in the spatial distribution. Let the hyperedge in the hypergraph be the... The weight of each superedge is The weight difference between adjacent hyperedges constitutes the weight gradient. When a certain local area Significantly deviates from the global mean and exceeds the anomaly detection threshold. At this point, the region is marked as an anomalous region of hyperedge weight gradient, and the partition combination associated with the corresponding hyperedge is the excitation source of the cross-partition power conflict mode. Precise location of the excitation source allows subsequent power transfer strategies to target the root cause of the conflict, rather than indiscriminately adjusting the global allocation scheme. After locating the excitation source, the propagation path of the conflict mode is traced along the topological connectivity of the hypergraph: starting from the excitation source hyperedge, adjacent hyperedges sharing the same nodes are identified sequentially, extending gradually according to the direction of the weight gradient, forming a conflict propagation link radiating outward from the excitation source. The propagation path tracing results are stored as an ordered sequence of hyperedges for use in constructing the subsequent power transfer potential gradient.

[0097] When a cross-regional power conflict mode is triggered, or the sum of the total loads of candidate schemes across regions exceeds the grid capacity constraint, the power transfer strategy construction phase begins. Phase information on the time-varying power demand of each region is extracted from the global state characteristics. Since the power demand of each region does not change synchronously in time, and the peak demand times of different regions differ, this temporal difference is quantified as a phase difference. ,in , This represents the index of any two partitions. A larger phase difference indicates a greater temporal misalignment in the urgency of power demands between the two partitions, resulting in lower costs for transferring power from the high-urgency partition to the low-urgency partition. The phase difference information of all partitions is integrated to construct an urgency phase field. This field describes the urgency of power demand in each partition at the current moment and their interrelationships within the power allocation space. The urgency phase field is essentially a scalar field defined on the set of partitions. Partitions with high field values ​​indicate that their power demand is the most urgent at the current moment and should be given priority in power supply; partitions with low field values ​​have the capacity to transfer power to other partitions.

[0098] The generation of the power transfer potential gradient is achieved by coupling the urgency phase field with the topological manifold structure of the conflicting topological hypergraph. The topological manifold structure of the conflicting topological hypergraph describes the geometry of the interval power competition relationship in high-dimensional space, including the manifold's curvature distribution, connectivity, and local neighborhood structure. The urgency phase field is then coupled with the topological manifold structure. Projecting this onto the topological manifold, aligning the gradient direction of the phase field with the local tangent space of the manifold, yields the power transfer potential gradient field. . At each point on the manifold, the optimal direction for power transfer is indicated, i.e., the direction of flow from high-potential-energy regions (regions with high urgency and high conflict pressure) to low-potential-energy regions (regions with low urgency and power-acceptance capacity). The power transfer strategy dictates that power flows along... The downward flow crosses the partitions. In each iteration, the amount of power transferred out or received by each partition is determined by its local gradient magnitude in the potential energy gradient field. The larger the gradient magnitude, the more power is transferred in a single iteration.

[0099] Driven by the power transfer strategy, the power allocation candidate schemes for each partition are iteratively reconstructed, with each iteration following the... The guidelines adjust the power allocation of each zone, and after adjustment, re-examine two constraints: first, the topological manifold invariance, meaning that the coupling topology between zones does not change qualitatively after power adjustment, the connectivity structure and hyperedge relationships of the hypergraph remain stable, and the power transfer does not destroy the original cooperative dependencies between zones; second, the total load capacity constraint, meaning that the sum of the adjusted power of all zones does not exceed the maximum allowable capacity limit of the power grid. If both constraints are satisfied, the iteration terminates, and the current power allocation scheme for each partition becomes the final power allocation scheme. If either constraint is not satisfied, the power transfer amount in this round is reduced, and the next iteration continues until convergence.

[0100] The final power allocation scheme needs to undergo consistency verification before output to confirm that the gradient weights of the abnormal hyperedges originally marked as excitation sources in the conflicting topology hypergraph have fallen back to the normal range. No longer exceeding the threshold This indicates that the cross-regional power conflict mode has been effectively resolved. Simultaneously, it verifies that the power allocation amounts for each region fall within the power allocation boundaries determined by the lower-level decision-maker, ensuring the legitimacy of the final scheme within the hierarchical power decision-making framework. Through the complete process of constructing the conflict topology hypergraph, locating the excitation source, tracing the propagation path, analyzing the urgency phase field, generating the power transfer potential energy gradient, and iteratively reconstructing, the upper-level coordinator can systematically resolve cross-regional power conflicts and satisfy grid capacity constraints while maintaining the stability of the inter-regional topology relationships, outputting the globally optimal final power allocation scheme.

[0101] In one optional implementation, a power transfer potential gradient is generated based on the coupling of the urgency phase field with the topological manifold structure of the conflict topological hypergraph. A power transfer strategy that drives power to flow across partitions along the descending direction of the power transfer potential gradient includes:

[0102] The urgency phase field is embedded into the fiber bundle structure of the conflict topological hypergraph. By maintaining the symplectic structure invariance of phase evolution, a gauge covariance relationship between phase dynamics and topological geometry is established in the fiber coordinate system of the topological manifold structure. Based on the gauge covariance relationship, the potential energy functional of each partition node is constructed. According to the potential energy functional, the evolution trajectory of the phase dynamics in the fiber coordinate system is coupled with the curvature characteristics of the topological geometry.

[0103] The covariant derivative of the potential functional is performed along the fiber bundle cross section of the topological manifold. The gauge invariant variation of the potential functional is calculated in the tangent bundle space of the fiber bundle cross section by a parallel translation operator. The gauge invariant variation is mapped to the horizontal distribution corresponding to each hyperedge through the horizontal lifting of the connection to obtain the power transfer potential gradient. A power transfer strategy is constructed based on the power transfer potential gradient.

[0104] like Figure 2 As shown, the method includes:

[0105] In obtaining the urgent phase field After the conflicting topological hypergraph, the two need to be deeply coupled at the mathematical structure level in order to generate a geometrically meaningful power transfer potential gradient field. Therefore, the urgency phase field This process involves embedding the hypergraph into the topological manifold structure corresponding to the conflicting topological hypergraph, specifically through fiber bundle embedding. The topological manifold structure serves as the base space, with each partition node forming a base point set on the manifold, while the urgency phase field... The phase evolution information carried exists as fiber components attached to each base point. The core of fiber bundle embedding lies in maintaining the symplectic structure invariance of the phase evolution process, that is, the symplectic form in the phase space remains conserved during coordinate transformation or topological deformation, thereby ensuring that the physical interpretability of phase dynamics remains consistent before and after geometric operations.

[0106] In fiber coordinates, the relationship between phase dynamics and topological geometry is described by gauge covariance. Consider a partition node on a topological manifold. The fiber coordinates at the location are This coordinate describes the phase state of the node along the fiber direction. (Introduction of a connection) As a gauge field on the fiber bundle, it is used to characterize the parallel translation rules of the fiber coordinate system between different points in the base space. The gauge covariance relationship indicates that the evolution of phase dynamics in the fiber coordinate system is driven not only by the time derivative of the phase itself, but also by the connection... Modulation of the curvature of the base space. Specifically, for nodes... Its potential functional Defined as the phase state at the node Connection curvature tensor and the weight of the hyperedge in which the node participates in the hypergraph. The scalar functional obtained by coupled integration. Potential energy functional. The construction of this model allows for an explicit coupling between the evolution trajectory of phase dynamics in the fiber coordinate system and the curvature characteristics of the bottom space topology: when the topological curvature of a region containing a node is large, it means that the hyperedge connectivity around the node is more complex, the topological pressure of power conflict is higher, and the corresponding potential energy functional... The value is also relatively large, thus generating a stronger power transfer driving force in subsequent gradient calculations.

[0107] For potential energy functional Performing covariant derivative calculations along the fiber bundle cross-section of the topological manifold is used to obtain the power transfer potential gradient field. The key step. Covariant derivative calculations differ from ordinary partial derivatives; they are performed through a connection during the calculation process. The variation of fiber components is compensated for to ensure the covariance of the derivative under gauge transformation, meaning the physical meaning of the result does not change with the choice of coordinate system. A parallel translation operator is introduced in the tangent bundle space of the fiber bundle section. This operator describes how to transfer fiber coordinates from nodes. Move parallel to the node along the bottom space path The process. Parallel translation operator. By contact The line integral along the path is given by exponentiation, and its value reflects the curvature accumulation effect of the region traversed by the path. This is achieved using the parallel translation operator. For potential energy functional Perform gauge-invariant variational calculations on the fiber bundle cross section: for nodes The potential functional at a given point is used to calculate the variational response of the functional along each direction of the tangent bundle space in its neighborhood, while simultaneously using the parallel shift operator. Adjacent nodes The fiber coordinates at the location are pulled back to the node. The fiber coordinate system at that location is then compared again, thereby eliminating spurious variational components introduced purely by coordinate system differences and preserving the true geometrically driven variation. This gauge-invariant variation is denoted as... Its direction and amplitude remain unchanged under gauge transformation, exhibiting good geometric stability.

[0108] The norm-invariant variation After contact The horizontal lifting operation is mapped to the horizontal distribution corresponding to each hyperedge, thereby obtaining the power transfer potential gradient field defined at the hyperedge level of the hypergraph. Horizontal lifting is a standard operation in fiber bundle theory. It lifts a tangent vector in the base space to a horizontal tangent vector in the entire space, making the lifted vector orthogonal to the fiber direction. This removes redundant degrees of freedom in the fiber direction, retaining only the driving components determined by the geometry of the base space. For the _th ... A hyperedge, the horizontal distribution of which is formed by the union of the horizontal tangent spaces of each partition node connected by the hyperedge. Horizontal elevation to super-edge After the horizontal distribution, the potential energy gradient components on the hyperedge are obtained. Its direction points to the partition node where the potential energy functional decreases fastest along the hyperedge direction, and its magnitude reflects the rate of change of potential energy in that direction. The potential energy gradient components on all hyperedges are then combined. This forms a complete power transfer potential energy gradient field. .

[0109] Based on power transfer potential energy gradient field When constructing the power transfer strategy, the basic principle of power flowing across partitions along the direction of the potential energy gradient descent is followed. For partition nodes with higher potential energy functional values, it means that the partition is currently experiencing greater power conflict pressure or a more urgent phase field. The phase lead at this node is relatively high, requiring a partial power transfer to adjacent partition nodes with lower potential energy to achieve overall potential energy balance. The amount of power transferred is determined based on the potential energy gradient components on each hyperedge. The magnitudes of the gradients are proportionally distributed: hyperedges with larger gradient magnitudes correspond to larger power transfers, and hyperedges with smaller gradient magnitudes correspond to smaller power transfers, thus matching the intensity of the power flow with the conflict pressure distribution revealed by the topological geometry. When performing cross-regional power transfer, the hyperedge weights must also be considered. Weighted adjustment of transfer amounts, super-edge weights A larger hyperedge indicates a stronger power coupling relationship between the corresponding partition pairs, allowing for a larger power transfer flux. This is based on the aforementioned potential energy gradient field... The power transfer strategy is constructed, the power allocation candidate schemes of each zone are reconstructed, the cross-zone power conflict mode is effectively resolved, and the total load distribution tends to achieve a topological equilibrium state within the grid capacity constraint. Finally, it provides a reliable mathematical basis for the upper-level coordinator to output the final power allocation scheme that satisfies the global constraints.

[0110] In one optional implementation, distributing power adjustment instructions corresponding to the final power allocation scheme to each charging pile, and feeding back the executed load feedback data to the deep residual learning model and the upper-level coordinator to drive policy evolution includes:

[0111] According to the final power allocation scheme, a power adjustment instruction carrying a power adjustment trajectory is generated for each charging pile. The power adjustment instruction is distributed to the corresponding charging pile and load feedback data is collected during the execution process. The load feedback data includes the deviation trajectory between the actual power output value and the power adjustment trajectory.

[0112] Based on the deviation trajectory, a policy response manifold characterizing the execution dynamics is constructed, and a residual manifold embedding reflecting the prediction error propagation mode is extracted by analyzing the local geometric properties of the policy response manifold.

[0113] The residual manifold is embedded and fed back to the deep residual learning model to correct the manifold structure of the residual mapping. The policy response manifold is fed back to the upper coordinator to reconstruct the decision space geometry of the cross-regional power transfer policy. The policy evolution is driven based on the corrected manifold structure, the reconstructed decision space geometry, and the curvature information of the shared policy response manifold.

[0114] After obtaining the final power allocation scheme, it needs to be converted into power adjustment commands that can be directly executed by each charging pile. Unlike simple power setpoint commands, the power adjustment commands generated here carry a power adjustment trajectory, that is, the power change path that the charging pile should follow within the effective time window of the command. The power adjustment trajectory provides a sequence of target power values ​​at each moment in terms of time resolution, enabling the charging pile controller to smoothly adjust the output power according to a predetermined rhythm, rather than abruptly changing after receiving the command, thereby reducing the impact on the power grid. The power adjustment command is distributed to the local controller of the corresponding charging pile through the charging station communication bus. During execution, the local controller continuously collects the actual power output value at a fixed sampling period and compares it point by point with the target power value at the corresponding moment in the power adjustment trajectory to form a deviation sequence. This deviation sequence, when expanded in the time dimension, constitutes the deviation trajectory, which records the dynamic difference between the actual response and the expected trajectory during execution, and is the original data source for subsequent feedback modeling.

[0115] Deviation trajectories not only reflect the execution errors of individual charging piles but also implicitly contain the combined influence of various execution dynamics factors, such as charging pile hardware response characteristics, battery state of charge changes, and local temperature compensation. To extract structurally meaningful information from these complex deviation signals, the deviation trajectories of all charging piles are embedded in a high-dimensional feature space to construct a policy response manifold. Each point in the policy response manifold corresponds to a feature representation of a deviation trajectory. The local geometric properties of the manifold—including tangent space direction, curvature distribution, and geodesic structure—combined to characterize the similarities and differences in execution dynamics among different charging piles. Specifically, statistical moment features, frequency domain energy distribution features, and temporal autocorrelation features are extracted from the deviation trajectories of each charging pile. These are then concatenated to form a high-dimensional feature vector. Manifold learning methods are used to project these high-dimensional feature vectors onto a low-dimensional manifold coordinate system, ensuring that charging piles with similar execution dynamics are close to each other on the manifold, while charging piles with significantly different execution dynamics are far apart on the manifold.

[0116] After establishing the policy response manifold, residual manifold embeddings are extracted by analyzing its local geometric properties. The analysis of local geometric properties focuses on the principal curvature directions within the neighborhood of each point on the manifold. If the principal curvature of a certain region is large, it indicates drastic changes in the execution dynamics within that region, corresponding to a systematic deviation between the actual response of the charging pile and the expected trajectory. This deviation often stems from structural errors in the deep residual learning model's prediction of the power demand of this type of charging pile. By extracting the local tangent space basis vectors of these high-curvature regions, a residual manifold embedding can be constructed. This embedding describes the propagation pattern of the prediction error in the feature space in a compact, low-dimensional representation—that is, from which feature dimensions the model prediction error is input, along which directions it propagates, and on which types of charging piles it accumulates and amplifies. Let the local principal curvature of a point on the manifold be... The corresponding neighborhood radius is Then the residual embedding weight at that point satisfy Points with larger weights are given higher priority in the residual manifold embedding, ensuring that the execution dynamics that have the most significant impact on the propagation of prediction errors are fully characterized.

[0117] The residual manifold embedding is fed back to the deep residual learning model to correct the manifold structure of the residual mapping. In the deep residual learning model, the curvature distribution of the feature manifolds of the cross-layer feature pathways implicitly determines the model's prediction bias pattern for the power demand of different types of charging piles when learning the residual mapping. Using the residual manifold embedding as an additional supervision signal guides the model to adjust the distribution of the output features of each residual block on the manifold, aligning the curvature structure of the model's output feature manifold with the error propagation pattern revealed by the actual execution dynamics. Specifically, a manifold alignment regularization term is added to the model's loss function. This regularization term measures the mean geodesic distance between the model's current output feature manifold and the target residual manifold embedding, denoted as . By minimizing The residual mapping of the model gradually converges to a manifold structure consistent with the execution dynamics. This correction process does not change the network topology of the model, but only adjusts the parameters of each layer through gradient backpropagation. Therefore, it can be carried out continuously in an online incremental learning manner during the normal operation of the charging station without interrupting the service.

[0118] The policy response manifold is simultaneously fed back to the upper-level coordinator to reconstruct the decision space geometry of the cross-sector power transfer policy. When tracking cross-sector power conflict patterns, the upper-level coordinator's decision space is spanned by the power state vectors of each sector. Before the introduction of the policy response manifold, the decision space was implicitly assumed to be Euclidean space, and the distance and direction relationships between the power states of each sector were measured linearly. The curvature information of the policy response manifold reveals the nonlinear coupling relationship between the execution dynamics of different sectors: when the execution dynamics of two sectors are adjacent on the manifold, power transfer between them is easier to achieve at the actual execution level, with lower transfer costs; conversely, when power transfer occurs between sectors with significantly different execution dynamics, the actual execution deviation is greater, requiring more power margin to be reserved in the decision-making stage. The upper-level coordinator redefines the power transfer cost metric between sectors based on the curvature information of the policy response manifold, replacing the original cost matrix based on Euclidean distance with a cost matrix based on the geodesic distance of the manifold, and denotes the cost of each sector node. and The distance between the manifold geodesics is Power transfer cost and Positive correlation makes the construction of cross-regional power transfer strategies closer to the actual execution capability boundary.

[0119] After manifold structure correction and decision space geometry reconstruction are completed, the deep residual learning model and the upper-level coordinator share the curvature information of the policy response manifold, thus enabling collaborative policy evolution between the two modules. The driving mechanism of policy evolution is reflected in the following: the individual power demand prediction output by the deep residual learning model after correction more accurately reflects the actual response capability of each charging pile; the cross-regional power transfer strategy generated by the upper-level coordinator based on the reconstructed decision space geometry is more in line with the actual execution dynamics constraints. The combined effect of these two factors results in a smaller deviation trajectory for the next round of power adjustment commands at the execution level, and the curvature of the policy response manifold gradually flattens, indicating that the execution dynamics of the overall system tend to stabilize. This closed-loop feedback mechanism enables the power dynamic allocation capability of the charging station to continuously improve with the accumulation of operating time, ultimately achieving a self-consistent steady-state equilibrium among model prediction accuracy, decision space geometry, and actual execution dynamics.

[0120] A second aspect of this invention provides a deep learning-based dynamic power allocation system for charging piles, comprising:

[0121] The data acquisition unit is used to acquire real-time operating data, historical load data, and grid capacity constraint information of all charging piles in the charging station;

[0122] The predictive modeling unit is used to construct a deep residual learning model to perform representation learning on the real-time operating data and the historical load data. It prevents gradient degradation by establishing cross-layer feature paths in the deep network and outputs global state features and individual power demand predictions for each charging pile.

[0123] The partition decision unit is used to decouple the charging pile group into multiple power management partitions based on the global state characteristics and the grid capacity constraint information, establish a hierarchical power decision framework, and the lower-level decision unit defines the power allocation boundary of the corresponding partition based on the global state characteristics, and couples the individual power demand prediction of each charging pile in the corresponding partition to generate power allocation candidate schemes.

[0124] The power coordination unit is used by the upper-layer coordinator to track the cross-partition power conflict mode between the power allocation candidate schemes of each partition based on the global state characteristics. When a power conflict is triggered or the total load exceeds the capacity constraint, the unit uses the global state characteristics to construct a cross-partition power transfer strategy and reconstruct the power allocation candidate schemes of each partition, and outputs the final power allocation scheme.

[0125] The instruction feedback unit is used to distribute the power adjustment instruction corresponding to the final power allocation scheme to each charging pile, and to feed back the executed load feedback data to the deep residual learning model and the upper-level coordinator to drive the strategy evolution.

[0126] A third aspect of the present invention provides an electronic device, comprising:

[0127] processor;

[0128] Memory used to store processor-executable instructions;

[0129] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0130] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0131] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0132] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A deep learning-based method for dynamic power allocation in charging piles, characterized in that, include: Obtain real-time operating data, historical load data, and grid capacity constraint information for all charging piles within the charging station; A deep residual learning model is constructed to perform representation learning on the real-time operating data and the historical load data. Cross-layer feature paths are established in the deep network to prevent gradient degradation, and global state features and individual power demand predictions for each charging pile are output. Based on the global state characteristics and the grid capacity constraint information, the charging pile group is decoupled into multiple power management zones, and a hierarchical power decision framework is established. The lower-level decision-maker defines the power allocation boundary of the corresponding zone based on the global state characteristics, and generates power allocation candidate schemes by coupling the individual power demand prediction of each charging pile in the corresponding zone. The upper-level coordinator tracks the cross-partition power conflict pattern between the power allocation candidate schemes of each partition based on the global state characteristics. When a power conflict is triggered or the total load exceeds the capacity constraint, the coordinator uses the global state characteristics to construct a cross-partition power transfer strategy and reconstruct the power allocation candidate schemes of each partition, and outputs the final power allocation scheme. The power adjustment command corresponding to the final power allocation scheme is distributed to each charging pile, and the load feedback data after execution is fed back to the deep residual learning model and the upper-level coordinator to drive the strategy evolution.

2. The method according to claim 1, characterized in that, A deep residual learning model is constructed to perform representation learning on the real-time operating data and the historical load data. Cross-layer feature pathways are established in the deep network to prevent gradient degradation. The output includes global state features and individual power demand predictions for each charging pile: The real-time running data and historical load data are segmented into time-series windows, and after time-series alignment and feature dimension standardization, the deep residual learning model is driven to start the encoding process of the feature extraction layer. In the feature extraction layer, feature transformation is performed layer by layer through the residual unit sequence. Each residual unit allows the features of the preceding layer to directly penetrate to the subsequent layer through the cross-layer feature path and performs semantic co-coding with the transformed features of the subsequent layer to construct a fusion feature representation that carries multi-level semantic information. The fused feature representation is injected into the state decoupling module. By hierarchically identifying the contribution of different levels of features in the fused feature representation, the global state features representing the overall operation mode of the charging station and the individual state features representing the load fluctuation mode of each charging pile are parsed out. Based on the global state features, a time-series evolution trajectory is constructed. The decision boundary of the power prediction module is defined according to the time-series evolution trajectory. The power prediction module integrates the individual state features and the time-series evolution trajectory and generates individual power demand predictions for each charging pile through a nonlinear mapping relationship.

3. The method according to claim 2, characterized in that, Based on the global state features, a time-series evolution trajectory is constructed. The decision boundary of the power prediction module is defined according to the time-series evolution trajectory. The power prediction module then fuses the individual state features and the time-series evolution trajectory and generates individual power demand predictions for each charging pile through a nonlinear mapping relationship, including: The global state characteristics are decomposed into a time-series dimension to extract a time-series base pattern that characterizes the load evolution law of the charging station. Based on the phase change and amplitude fluctuation characteristics of the time-series base pattern, the state evolution path of future periods is deduced, and a time-series evolution trajectory containing the evolution direction and evolution rate is constructed. The decision boundary of the power prediction module is determined based on the evolution direction and evolution rate in the time-series evolution trajectory. The decision boundary is formed by converting the evolution rate into a power change rate constraint and combining it with the evolution direction to define the reachable region of the prediction space, thereby forming a boundary constraint domain that dynamically shrinks or expands with the time-series evolution trajectory. Individual state features are mapped to the boundary constraint domain. The relative position of the individual state features within the boundary constraint domain is converted into a power prediction value through a nonlinear mapping relationship. The nonlinear mapping relationship adjusts the mapping curvature according to the degree of deviation between the individual state features and the time-series evolution trajectory. When the degree of deviation increases, the mapping curvature is enhanced to make the prediction value converge to the time-series evolution trajectory, thereby generating individual power demand predictions for each charging pile that are limited by the boundary constraint domain and coordinated with the time-series evolution trajectory.

4. The method according to claim 1, characterized in that, The lower-level decision-maker defines the power allocation boundary of the corresponding partition based on the global state features, and generates power allocation candidate schemes by coupling the individual power demand predictions of each charging pile in the corresponding partition, including: The state subspace of the corresponding partition is extracted from the global state features. Based on the coupling relationship between load saturation and power fluctuation margin in the state subspace, the morphological features of the power allocation boundary are determined. The morphological features form a dynamic antagonistic balance through the regulation of the boundary contraction degree by the load saturation and the release of the boundary expansion potential by the power fluctuation margin. An adaptive boundary constraint domain containing the morphological evolution trajectory is constructed. The individual power demand predictions of each charging pile in the corresponding partition are mapped to the topology of the adaptive boundary constraint domain. By analyzing the spatial distribution characteristics of the demand of each charging pile in the topology, demand clustering regions and demand sparse regions are identified. The boundary pressure distribution is quantified according to the edge proximity of the demand clustering region to the adaptive boundary constraint domain. The priority response sequence of power allocation is determined according to the boundary pressure distribution. Power quotas are iteratively allocated according to the priority response sequence. After each allocation, the remaining available space of the adaptive boundary constraint domain is reconstructed based on the change of the boundary pressure distribution. The geometric features of the remaining available space are fed back to the next allocation process to form a constraint transmission chain. When the distance of the geometric center of the remaining available space from the initial position exceeds the offset threshold determined by the morphological evolution trajectory, quota compression is triggered to generate a power allocation candidate scheme that takes into account both the priority of demand response and the integrity of boundary constraints.

5. The method according to claim 1, characterized in that, The upper-level coordinator tracks cross-partition power conflict patterns among the power allocation candidate schemes of each partition based on the global state characteristics. When a power conflict is triggered or the total load exceeds the capacity constraint, the coordinator uses the global state characteristics to construct a cross-partition power transfer strategy and reconstruct the power allocation candidate schemes of each partition, outputting the final power allocation scheme, including: High-order coupling correlations of power demand in each partition are extracted from global state features. A conflict topology hypergraph reflecting the power competition game relationship between partitions is constructed based on the high-order coupling correlations. The source of cross-partition power conflict mode is located by identifying abnormal regions of the hyperedge weight gradient in the conflict topology hypergraph. The propagation path of the cross-partition power conflict mode from the source to other partitions is tracked based on the distribution characteristics of the hyperedge weight gradient in the conflict topology hypergraph. When the cross-zone power conflict mode is triggered or the total load exceeds the capacity constraint, the urgency phase field of power demand is analyzed based on the phase difference of the temporal evolution of each zone in the global state characteristics. The power transfer potential energy gradient is generated by coupling the urgency phase field with the topological manifold structure of the conflict topology hypergraph, and a power transfer strategy is constructed to drive the power to flow across zones along the downward direction of the power transfer potential energy gradient. According to the power transfer strategy, the power allocation candidate schemes of each partition are iteratively reconstructed, and the final power allocation scheme that takes into account the topological manifold structure invariance of the conflicting topological hypergraph and the total load capacity constraint is output.

6. The method according to claim 5, characterized in that, Based on the coupling of the urgency phase field with the topological manifold structure of the conflict topology hypergraph, a power transfer potential energy gradient is generated. A power transfer strategy is constructed to drive power flow across partitions along the descending direction of this power transfer potential energy gradient, including: The urgency phase field is embedded into the fiber bundle structure of the conflict topological hypergraph. By maintaining the symplectic structure invariance of phase evolution, a gauge covariance relationship between phase dynamics and topological geometry is established in the fiber coordinate system of the topological manifold structure. Based on the gauge covariance relationship, the potential energy functional of each partition node is constructed. According to the potential energy functional, the evolution trajectory of the phase dynamics in the fiber coordinate system is coupled with the curvature characteristics of the topological geometry. The covariant derivative of the potential functional is performed along the fiber bundle cross section of the topological manifold. The gauge invariant variation of the potential functional is calculated in the tangent bundle space of the fiber bundle cross section by a parallel translation operator. The gauge invariant variation is mapped to the horizontal distribution corresponding to each hyperedge through the horizontal lifting of the connection to obtain the power transfer potential gradient. A power transfer strategy is constructed based on the power transfer potential gradient.

7. The method according to claim 1, characterized in that, Distributing power adjustment instructions corresponding to the final power allocation scheme to each charging pile, and feeding back the executed load feedback data to the deep residual learning model and the upper-layer coordinator to drive policy evolution includes: According to the final power allocation scheme, a power adjustment instruction carrying a power adjustment trajectory is generated for each charging pile. The power adjustment instruction is distributed to the corresponding charging pile and load feedback data is collected during the execution process. The load feedback data includes the deviation trajectory between the actual power output value and the power adjustment trajectory. Based on the deviation trajectory, a policy response manifold characterizing the execution dynamics is constructed, and a residual manifold embedding reflecting the prediction error propagation mode is extracted by analyzing the local geometric properties of the policy response manifold. The residual manifold is embedded and fed back to the deep residual learning model to correct the manifold structure of the residual mapping. The policy response manifold is fed back to the upper coordinator to reconstruct the decision space geometry of the cross-regional power transfer policy. The policy evolution is driven based on the corrected manifold structure, the reconstructed decision space geometry, and the curvature information of the shared policy response manifold.

8. A deep learning-based dynamic power allocation system for charging piles, used to implement the method as described in any one of claims 1-7, characterized in that, include: The data acquisition unit is used to acquire real-time operating data, historical load data, and grid capacity constraint information of all charging piles in the charging station; The predictive modeling unit is used to construct a deep residual learning model to perform representation learning on the real-time operating data and the historical load data. It prevents gradient degradation by establishing cross-layer feature paths in the deep network and outputs global state features and individual power demand predictions for each charging pile. The partition decision unit is used to decouple the charging pile group into multiple power management partitions based on the global state characteristics and the grid capacity constraint information, establish a hierarchical power decision framework, and the lower-level decision unit defines the power allocation boundary of the corresponding partition based on the global state characteristics, and couples the individual power demand prediction of each charging pile in the corresponding partition to generate power allocation candidate schemes. The power coordination unit is used by the upper-layer coordinator to track the cross-partition power conflict mode between the power allocation candidate schemes of each partition based on the global state characteristics. When a power conflict is triggered or the total load exceeds the capacity constraint, the unit uses the global state characteristics to construct a cross-partition power transfer strategy and reconstruct the power allocation candidate schemes of each partition, and outputs the final power allocation scheme. The instruction feedback unit is used to distribute the power adjustment instruction corresponding to the final power allocation scheme to each charging pile, and to feed back the executed load feedback data to the deep residual learning model and the upper-level coordinator to drive the strategy evolution.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images