A Dynamic Power Allocation Method for Charging Station Clusters Based on Load Forecasting
By constructing a joint topology graph and distributed consensus iteration, the problem of power collaborative allocation in charging station clusters under complex topology and load prediction uncertainty is solved, realizing dynamic power allocation with autonomous collaboration and physical constraint security, and improving the stability and efficiency of the system.
Patent Information
- Application Number
- CN202611222294.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-12
- Publication Date
- 2026-11-10
AI Technical Summary
Under complex topology and load forecasting uncertainty, the stability and efficiency of power collaborative allocation in charging station clusters are insufficient, and existing solutions are difficult to achieve real-time autonomous collaboration and physical constraint satisfaction under dynamic topology.
A joint topology map integrating spatial proximity, electrical coupling, and communication accessibility is constructed. The load forecast residual is quantified as the power regulation elasticity coefficient. Decentralized power collaborative allocation is achieved through distributed consensus iteration. Physical constraints are met through feeder capacity over-limit verification and regulation pressure sharing.
It enables autonomous and coordinated allocation of charging station clusters under complex topologies and load forecast uncertainties, ensuring system consistency and physical constraint safety, and automatically iterating and reconverging when adapting to topology changes, thereby improving the stability and efficiency of power allocation.
Smart Images

Figure CN123073713A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power system dispatching and distributed energy management technology, and in particular to a dynamic power allocation method for charging station groups based on load forecasting. Background Technology
[0002] The large-scale integration of electric vehicle charging infrastructure has made power allocation for urban charging station clusters a crucial aspect of power distribution network operation and management. Currently, power allocation for charging station clusters is typically based on load forecasting results, using forecast data at multiple time scales, such as day-ahead and intraday, to generate dispatch instructions.
[0003] Existing solutions mainly fall into two categories: centralized optimization scheduling and preliminary distributed control. Centralized methods rely on a central node to aggregate data from all sites and then solve the problem uniformly, which results in heavy communication burden and high risk of single-point failure. Although distributed methods can alleviate the above problems, the coordination mechanism between sites in existing solutions is relatively simple, and cross-site power interaction is mostly limited to post-event adjustments after feeder safety limits are exceeded, making it difficult to achieve real-time autonomous coordination under dynamic topology.
[0004] Furthermore, existing solutions typically address load forecasting errors through global parameter correction or rolling plans, failing to establish a direct correlation with the actual adjustment capabilities of each station. This leads to the potential for localized overload or resource idleness when forecasting deviations accumulate. For common topology changes in large-scale systems, such as communication link fluctuations and station commissioning / discontinuation, existing methods suffer from slow convergence speed and insufficient adaptability, making it difficult to meet actual operational needs. Summary of the Invention
[0005] This application provides a dynamic power allocation method for charging station clusters based on load forecasting, aiming to solve the problem of insufficient stability and efficiency of power collaborative allocation in charging station clusters under the influence of complex topology and load forecasting uncertainty.
[0006] This application provides a dynamic power allocation method for charging station groups based on load forecasting, specifically including:
[0007] Obtain topological association data of multiple charging stations in the charging network, and construct a joint topological map based on the topological association data to characterize the spatial proximity, electrical coupling and communication reachability between charging stations.
[0008] Obtain load forecast residual sequences at multiple time scales, extract error persistence features and error mutation features from the load forecast residual sequences, and generate an error feature vector characterizing forecast uncertainty.
[0009] Based on the error feature vector, the influence of the prediction error on the power regulation capability of the multiple charging stations is quantified, and a power regulation elasticity coefficient is generated.
[0010] By utilizing the adjacency relationships in the joint topology graph and the power adjustment elasticity coefficient, and through distributed consensus iteration, the charging station exchanges power allocation suggestion values with neighboring stations and performs weighted updates to generate intermediate power command variables.
[0011] The intermediate variables of the power command are checked for over-limit feeder capacity. If there is a risk of over-limit, the pressure is distributed according to the topological adjacency relationship to generate a cooperative power command sequence that meets the feeder capacity constraints.
[0012] In response to changes in the charging station cluster topology, the joint topology graph is updated and the distributed consensus iterative reconvergence is triggered, outputting a dynamic power allocation instruction sequence.
[0013] The dynamic power allocation method for charging station groups based on load forecasting provided in this application has the following beneficial effects:
[0014] By constructing a joint topology map integrating spatial proximity, electrical coupling, and communication reachability, the load forecast residual is quantified into a power regulation elasticity coefficient to characterize the regulation capability of each station. Based on this map and elasticity coefficient, distributed consensus iteration is performed to achieve decentralized power collaborative allocation. Simultaneously, feeder capacity over-limit verification and regulation pressure sharing ensure the satisfaction of physical constraints. Iterative reconvergence is automatically triggered when the topology changes to maintain system consistency. This scheme solves the technical problems of insufficient stability and efficiency in collaborative allocation of charging station clusters under complex topologies and load forecast uncertainties, achieving autonomous collaborative power allocation, physical constraint safety, and dynamic topology adaptation for charging station clusters. Attached Figure Description
[0015] Figure 1 This is the main flowchart of a dynamic power allocation method for charging station clusters based on load forecasting.
[0016] Figure 2 This is a sub-flowchart of a dynamic power allocation method for charging station clusters based on load forecasting.
[0017] Figure 3 This is a flowchart of an algorithm for generating an amortization coefficient matrix based on the topological edge weight ratio, which is a dynamic power allocation method for charging station groups based on load forecasting. Detailed Implementation
[0018] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.
[0019] The following disclosure provides many different embodiments or examples for implementing various structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. These are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.
[0020] like Figure 1 As shown, this application provides a dynamic power allocation method for charging station groups based on load forecasting, specifically including:
[0021] S1: Obtain the geographical coordinates, access distribution network node number, transformer capacity and real-time communication link status data of each charging station in the urban charging network. Based on the electrical distance threshold, feeder affiliation and communication reachability conditions, construct a three-dimensional topology perception map that includes spatial proximity, electrical coupling and communication reachability, and generate a joint topology map composed of vertices and weighted bidirectional edges.
[0022] S2: Input the load forecast residual sequences of the day-ahead, intraday and real-time multi-timescales into the local lightweight error feature extraction module, calculate the number of consecutive over-limit periods as the error persistence index and the deviation change rate of adjacent periods as the error mutation intensity index, and generate an error feature vector characterizing the forecast uncertainty.
[0023] S3: Based on the error persistence index and error mutation intensity index in the error feature vector, perform nonlinear mapping operation to quantify the impact of prediction error on power regulation capability, and generate power regulation elasticity coefficients to characterize the adaptive regulation capability of each charging station in topology iteration.
[0024] S4: Utilize the adjacency relationships and power adjustment elasticity coefficients in the joint topology graph to initiate a distributed consensus iteration process, enabling each charging station to exchange current power allocation suggestions only with its topology neighbors, and generate local power command intermediate variables through a weighted average update rule.
[0025] S5: Determine whether the potential over-limit risk of the local power command intermediate variable and the total capacity of the feeder meets the soft constraint condition of power balance. If there is an over-limit risk, trigger the local scaling mechanism, distribute the pressure to neighboring stations according to the topology edge weight ratio, and correct the local power command intermediate variable to generate a coordinated power command sequence that meets the feeder capacity constraint.
[0026] S6: Monitor changes in the topology of the charging station group or the status of new station access, reinitialize the boundary conditions of the distributed consensus iteration process based on the updated joint topology graph, drive the cooperative power command sequence to automatically reconverge to the new global coordination solution, and output the final dynamic power allocation command sequence.
[0027] Step S1: Obtain the geographical coordinates, distribution network node numbers, transformer capacity, and real-time communication link status data of each charging station in the urban charging network. Based on electrical distance thresholds, feeder affiliation, and communication reachability conditions, construct a 3D topology-aware map that includes spatial proximity, electrical coupling, and communication reachability, generating a joint topology map composed of vertices and weighted bidirectional edges. Specifically, this includes:
[0028] S1.1: Synchronously acquire and process the geographic coordinate data, access distribution network node number data, transformer capacity data, and real-time communication link status data of each charging station in the urban charging network to generate a multi-dimensional basic dataset of the charging station group containing spatial location information, electrical connection identifiers, equipment capacity parameters, and communication connectivity status.
[0029] By using edge data acquisition gateways deployed on the urban charging network dispatch platform, the operational database of each charging station controller is periodically polled to extract the current latitude and longitude coordinates and the unique identifier ID of the transformer in the corresponding distribution network. Standardized interface protocols are used to read the rated transformer capacity parameters of each station, and the round-trip latency and packet loss rate of communication heartbeat packets between the station and the central server are monitored in real time. The topology association data includes at least the spatial location information, electrical connection identifiers, equipment capacity parameters, and communication link status data of each charging station.
[0030] The heterogeneous data was uniformly mapped to a standard timestamp sequence, and data cleaning was performed to remove null values or outliers caused by sensor malfunctions. The cleaned geographic coordinates were then transformed and projected onto a local Cartesian coordinate system to eliminate the influence of the Earth's curvature on short-distance calculations.
[0031] Transformer capacity data is normalized into scalar values, and communication link status is quantized into binary connectivity flags and average continuous delay values. Four types of data—spatial location, electrical identification, capacity scalars, and communication status—are integrated to construct a structured multidimensional basic data table with the charging station ID as the primary key.
[0032] This processing method transforms scattered, heterogeneous raw data from multiple sources into a unified, time-aligned, multi-dimensional basic dataset of charging station clusters, enabling the preparation of all-element data required for subsequent topology construction.
[0033] For example, a city area contains 50 charging stations, and the edge gateway collects data every 5 seconds. Station A has coordinates (116.40, 39.90), transformer ID T01, capacity 630kVA, communication latency 15ms, and packet loss rate 0.1%. Station B has coordinates (116.41, 39.91), transformer ID T02, capacity 800kVA, communication latency 20ms, and packet loss rate 0.2%. The system removes empty coordinate data reported by station C due to a fault and marks it as offline. The latitude and longitude of stations A and B are converted to local coordinates (100, 200) and (110, 210), respectively. The transformer capacity is normalized to 0.78 and 1.0 (baseline 800kVA). The communication status is quantified as a connectivity flag of 1 and average latency of 15ms and 20ms. The final result is a multidimensional basic dataset containing 50 valid nodes, each of which contains a 4-dimensional attribute vector, ensuring data integrity and consistency and providing accurate input for subsequent electrical distance calculations.
[0034] S1.2: Based on the multi-dimensional basic dataset of charging station clusters, perform electrical distance calculation and feeder affiliation matching operation. Use the preset electrical distance threshold to determine the electrical coupling strength between stations and combine the feeder homology analysis results to generate an electrical proximity correlation matrix that characterizes the physical connection tightness between stations.
[0035] The S1.2 sub-step aims to quantify the electrical coupling strength between charging stations and generate an electrical proximity correlation matrix based on a multidimensional basic dataset of charging station clusters.
[0036] The node numbers and transformer impedance parameters of the charging station cluster in the multidimensional basic dataset are analyzed to extract the electrical node identifiers and equivalent impedance values of each charging station in the distribution network topology.
[0037] For any two charging station nodes, retrieve all intermediate nodes and line segments on their physical connection path in the power distribution network, and construct a complete electrical path sequence from the source node to the target node.
[0038] Based on the physical parameter model of the distribution network line, the resistance and reactance values of each line segment in the path are calculated, and the total equivalent impedance modulus between the two stations is obtained by summing them up. The standardized electrical distance is obtained by standardizing the standard impedance value with reference to the system reference impedance value.
[0039] By comparing the feeder affiliation identifiers of the two stations, if the two stations are connected to the same 10kV feeder or the same low-voltage busbar of the distribution transformer, it is determined that they have feeder homology and are assigned a homology marking coefficient of 1; otherwise, it is 0.
[0040] The standardized electrical distance is compared with the preset electrical distance threshold. If the distance is less than the threshold and the feeders are of the same origin, then the two stations are considered to have a strong electrical coupling relationship.
[0041] Based on the determination results of strong electrical coupling, an N×N dimensional electrical proximity correlation matrix is constructed. The matrix element values reflect the tightness of the physical connection between stations, and non-coupled node pairs are assigned a value of zero.
[0042] Through the above-mentioned electrical distance calculation and source matching processing, the discrete distribution network topology information is transformed into a quantified electrical proximity correlation matrix, realizing an accurate characterization of the physical constraints of power interaction between charging stations.
[0043] For example, the system reference impedance Z_base is set to 10 ohms, and the preset electrical distance threshold is 0.5. For charging stations A and B connected to the same 10kV feeder, their equivalent impedance magnitude is measured to be 3 ohms, and the calculated normalized electrical distance is 0.3. Since 0.3 is less than 0.5 and the feeders of the two stations are from the same source (marked as 1), it is determined that A and B have strong coupling, and the corresponding position in the correlation matrix is assigned a value of 1. For charging station C with a different feeder, its equivalent impedance magnitude with A is 12 ohms, and the normalized distance is 1.2, which is greater than the threshold, so the corresponding position in the correlation matrix is assigned a value of 0. The finally generated correlation matrix clearly distinguishes between strongly coupled and weakly coupled nodes, providing an accurate electrical constraint basis for subsequent topology weight fusion.
[0044] S1.3: Perform communication reachability verification on the inter-station nodes in the electrical proximity association matrix, filter the node pairs with direct communication channels based on real-time communication link status data, and generate a candidate set of bidirectional adjacency relationships that integrates electrical coupling conditions and communication connectivity conditions.
[0045] Read the node pair identifiers in the electrical proximity correlation matrix and extract the real-time communication link status data of each charging station, including heartbeat response time, packet loss rate and bandwidth utilization indicators.
[0046] The extracted communication link status data is subjected to connectivity threshold determination. The maximum allowable delay threshold and the minimum available bandwidth threshold are set to filter out physical links that meet the requirements of low latency and high reliability transmission.
[0047] Perform a logical AND operation on the links that meet the communication requirements and the electrical proximity association matrix, retain only the node pairs that have both electrical coupling foundation and direct communication capability, and remove invalid edges that have electrical connection but communication interruption or normal communication but electrical isolation.
[0048] The symmetry of the filtered bidirectional adjacency relationships is checked to ensure that if node A can reach node B, then node B can also reach node A, thus constructing the preliminary adjacency list structure of the undirected graph.
[0049] A communication quality score is generated for each valid adjacent node pair. The confidence weight is calculated based on historical link stability data and used as input data for the communication reachability dimension of subsequent three-dimensional attribute weight fusion.
[0050] Through communication reachability verification, the electrical proximity matrix from the previous step is transformed into a bidirectional adjacency candidate set that integrates electrical coupling conditions and communication connectivity conditions, thereby achieving accurate mapping of the communication dimension in the topology graph and ensuring that the data interaction channels for subsequent distributed consistency iterations are genuine and effective.
[0051] For example, the maximum allowable latency threshold is set to 50ms, and the minimum available bandwidth threshold is set to 1Mbps. In a charging station cluster, station A and station B are electrically separated by 2km and share the same feeder. However, the packet loss rate of their 4G communication link reaches 15%, exceeding the 5% threshold. Therefore, communication is deemed unreachable, and this edge is removed from the candidate set. Station A and station C are electrically separated by 5km, share the same feeder but are directly connected via a dedicated fiber optic network. The latency is 2ms, and the packet loss rate is 0.1%. Communication is deemed reachable, and although the electrical coupling is weak, there is potential for collaboration (depending on the specific threshold setting in S1.2; here, we assume a relatively wide electrical threshold). This edge is retained. The final candidate set only includes node pairs such as AC and BD that satisfy both electrical proximity and high-quality communication links, ensuring the real-time performance and reliability of subsequent power command exchanges.
[0052] S1.4: Based on the candidate set of bidirectional adjacency relationships, perform three-dimensional attribute weight fusion calculation, and map spatial proximity index, electrical coupling index and communication reachability index into edge weight values through a weighted fusion algorithm to generate weighted bidirectional edge attribute data that reflects the power coordination potential between charging stations.
[0053] Spatial distance data, electrical impedance parameters, and communication link quality indicators of each pair of charging station nodes in the candidate set of bidirectional adjacency relationships are extracted as the original input variables for the fusion of three-dimensional attribute weights.
[0054] Normalization mapping is performed on spatial distance data, and spatial proximity weights are calculated using an inverse decay function. The closer the distance, the higher the weight, which represents the convenience of power support at the geographical level.
[0055] The electrical coupling weight is calculated based on electrical impedance parameters. The reciprocal of the sum of the transformer short-circuit impedance and the line impedance is used as a metric. The smaller the impedance, the tighter the electrical connection and the lower the power exchange loss.
[0056] Read real-time communication link status data, including packet loss rate and latency jitter value, and generate communication reachability weights through an exponential penalty function. High-reliability links are given higher weights to ensure low-latency transmission of coordinated commands.
[0057] A weighted fusion model is constructed, in which the spatial proximity weight, electrical coupling weight, and communication reachability weight are multiplied by preset influence coefficients. The influence coefficients are dynamically adjusted according to the scheduling scenario preferences, and the weight of the electrical coupling coefficient is increased when physical stability is emphasized.
[0058] The following formula is used to calculate the charging station With charging station The combined edge weight values between :
[0059]
[0060] in, For spatial proximity weights, For electrical coupling weights, For communication reachability weights, , , These are the corresponding normalized influence coefficients, and the sum of the three is 1.
[0061] The calculated comprehensive edge weight values are truncated by thresholding to remove weak connections that are below the minimum collaborative potential threshold, thus preventing invalid nodes from interfering with the convergence speed of subsequent distributed iterations.
[0062] The retained valid edge weight values are encapsulated into weighted bidirectional edge attribute data and bound to the corresponding node pairs to form a topology connection strength matrix that reflects the potential for power collaboration between charging stations.
[0063] By using a weighted fusion algorithm, the candidate set from the previous step is transformed into quantified edge weight numerical data, thereby realizing a digital representation of the multi-dimensional collaborative capabilities of the charging station group and providing accurate topological constraints for subsequent distributed consistency iterations.
[0064] For example, the spatial distance between charging station A and charging station B is set to 2km, with a normalized spatial weight of 0.8; the sum of electrical impedances is 0.5 ohms, with a normalized electrical weight of 0.9; the communication packet loss rate is 0.1%, with a communication weight of 0.95. Influence coefficients are set to α1=0.2, α2=0.5, and α3=0.3. Substituting these values into the formula, W is calculated. AB =0.20.8+0.50.9+0.3*0.95=0.16+0.45+0.285=0.895. This high weight value indicates that the two stations have a very strong power mutual assistance capability, and power balance adjustment will be prioritized in subsequent iterations.
[0065] S1.5: Utilize weighted bidirectional edge attribute data and vertex information of each charging station to perform graph structure assembly operations, construct a three-dimensional topology-aware graph that includes spatial proximity, electrical coupling, and communication reachability, and finally output a joint topology graph composed of vertices and weighted bidirectional edges.
[0066] Step S2: Input the load forecast residual sequences at multiple time scales (day-ahead, intraday, and real-time) into the local lightweight error feature extraction module. Calculate the number of consecutive periods exceeding limits as an error persistence index and the rate of change of deviation between adjacent periods as an error mutation intensity index, generating an error feature vector characterizing forecast uncertainty. Specifically, this includes:
[0067] S2.1: Obtain raw load forecast data and corresponding actual load monitoring data at multiple time scales, including day-ahead, intraday, and real-time data. Perform differential operation processing based on the timestamp alignment mechanism to generate a set of load forecast residual sequences containing deviation information in multiple time dimensions.
[0068] Substep S2.1 aims to extract residual sequences from multi-timescale forecast data, providing a foundation for subsequent feature quantization. This step receives the raw data streams of day-ahead, intraday, and real-time load forecasts, as well as the actual load monitoring data streams for the corresponding time periods, as the execution targets.
[0069] The received multi-source heterogeneous data is timestamped and standardized, and the UTC time format is uniformly adopted with accuracy down to the second to eliminate timing misalignment caused by data acquisition clock deviation.
[0070] Based on a unified timeline, a linear interpolation algorithm is used to fill in missing predicted or measured data points, ensuring that the data at the three time scales of day-ahead, intraday, and real-time are strictly aligned in the time dimension.
[0071] For the aligned dataset, the predicted load value and the actual load value at the same time are extracted, and point-by-point difference operation is performed to calculate the instantaneous prediction deviation at each time scale.
[0072] The load forecast residual at time t for the k-th time scale is calculated using the following formula:
[0073]
[0074] in, Indicates the first Each time scale The predicted residual at time, Indicates the first Each time scale Forecast load value at time of day express The actual load monitoring value at any given time.
[0075] The calculated residual data in the three dimensions of day-ahead, intraday, and real-time are arranged in chronological order to construct a three-dimensional residual matrix, forming a set of load forecast residual sequences containing multi-time dimension deviation information.
[0076] Through the above differential operations and sequence construction processing, the multi-source prediction and actual monitoring data from the previous step are transformed into a standardized set of load prediction residual sequences, realizing the structured expression of prediction uncertainty data and laying the data foundation for the subsequent feature extraction of error persistence and time-varying intensity.
[0077] For example, the day-ahead forecast time step is set to 15 minutes, the intraday time step to 5 minutes, and the real-time time step to 1 minute. At t=10:00, the day-ahead forecast load is 500kW, the intraday forecast load is 520kW, and the real-time forecast load is 530kW, while the actual monitored load is 510kW. After time alignment and interpolation, the calculated day-ahead residual is -10kW, the intraday residual is 10kW, and the real-time residual is 20kW. The system stores these three residual values into the corresponding residual sequence vectors, forming a multi-dimensional residual snapshot for that moment. If the measured data is missing at a certain moment, the system uses linear interpolation to generate imputation values based on the data from adjacent moments, ensuring the continuity of residual calculation. The final output residual sequence set accurately reflects the deviation distribution under different forecast periods, significantly improving the data completeness and timeliness of subsequent feature extraction.
[0078] S2.2: Perform threshold comparison logic judgment on each residual data point in the load forecast residual sequence set, count the number of time periods that continuously exceed the preset error tolerance threshold, and generate an error persistence index that characterizes the long-term cumulative effect of forecast deviation.
[0079] Receive the load forecast residual sequence set generated by S2.1, which contains deviation data points at three time scales: day-ahead, intraday, and real-time. Perform absolute value operation on each data point in the residual sequence to eliminate the directional difference between positive and negative deviations, generating a non-negative error amplitude sequence.
[0080] A dynamic error tolerance threshold is set, which is determined based on a preset proportional coefficient of the charging station's rated power, and is used to define the boundary between normal fluctuations and abnormal deviations. Each element in the error amplitude sequence is compared point-by-point with the dynamic error tolerance threshold to generate a Boolean-type over-limit flag sequence, where a true value indicates that the error exceeds the allowable range at the current moment. A sliding window is performed on the over-limit flag sequence to identify subsequence segments with consecutive true values.
[0081] The number of time steps contained in each continuous subsequence segment is counted as the number of consecutive out-of-limit periods corresponding to that segment. The length of the most recent continuous subsequence segment at the current time is selected and defined as the error persistence index for the current time. Among them, the error persistence feature is used to characterize the long-term cumulative effect of prediction bias, and the error mutation feature is used to characterize the short-term sharp fluctuation characteristics of prediction bias.
[0082] Through the threshold comparison and continuous counting process described above, the discrete residual data is transformed into a scalar index that characterizes the long-term cumulative effect of prediction bias. This enables the quantitative extraction of the time-domain characteristics of prediction uncertainty and provides horizontal axis component data for the subsequent construction of error feature vectors.
[0083] For example, assuming the rated power of the charging station is 100kW and the preset proportional coefficient is 5%, the dynamic error tolerance threshold is 5kW. At a certain moment, the predicted residual sequence is [6,7,8,2,9,10]kW, and its absolute value remains [6,7,8,2,9,10]. Point-by-point comparison generates an over-limit flag sequence [1,1,1,0,1,1]. Two continuous subsequences are identified: the first three are [1,1,1] with a length of 3; the last two are [1,1] with a length of 2. If the current moment corresponds to the end of the sequence, the most recent segment with a length of 2 is selected as the error persistence index. This index accurately reflects the persistence of recent prediction deviations, avoids interference from large instantaneous errors on long-term trend judgments, and significantly improves the stability of the power regulation elasticity coefficient calculation.
[0084] S2.3: Perform first-order difference calculation processing based on the residual values of adjacent time steps in the load forecast residual sequence set, extract the absolute value of the deviation change in adjacent time periods as the rate of change data, and generate an error mutation intensity index that characterizes the short-term drastic fluctuation characteristics of the forecast deviation.
[0085] Obtain the residual values at the current time t and the previous time t-1 from the load forecast residual sequence set, and use them as the input data source for calculating the error mutation intensity.
[0086] Perform first-order difference operations on the residual values of adjacent time steps to calculate the difference between the residual at the current time and the residual at the previous time. Take the absolute value as the rate of change of the deviation between adjacent time periods to quantify the magnitude of the change of the prediction deviation in the short time domain.
[0087] The absolute value of the calculated deviation change is taken to eliminate the influence of positive and negative direction differences on the fluctuation intensity assessment and ensure the non-negative characteristics of the mutation index.
[0088] The absolute value results are mapped to an error mutation intensity index, which directly reflects the degree of instability of the prediction model within a local time window.
[0089] By using first-order difference and absolute value processing, the residual sequence from the previous step is transformed into an error mutation intensity index that characterizes the short-term drastic fluctuations in prediction deviation. This achieves accurate quantification of the transient characteristics of prediction uncertainty and provides key vertical axis component data for the subsequent generation of error feature vectors.
[0090] For example, the real-time load prediction residual of a charging station at 10:00 is 5kW, and the real-time load prediction residual at 10:05 is 12kW. The system extracts the residual values of these two adjacent time steps and performs a first-order difference calculation: ΔR = 12 − 5 = 7kW. Taking the absolute value of the difference yields 7kW, which is used as the error mutation intensity index at 10:05. If the residual changes from 10kW to 8kW at another time, the change is -2kW, which is 2kW after taking the absolute value. The comparison shows that the mutation intensity of 7kW is significantly higher than that of 2kW, indicating that the prediction deviation fluctuates more drastically at the former time. The system will assign a higher adjustment elasticity coefficient response weight to it in subsequent steps to enhance its adaptability to sudden load fluctuations.
[0091] S2.4: Using the error persistence index as the horizontal axis component and the error mutation intensity index as the vertical axis component, perform vector encapsulation construction processing to generate an error feature vector that can fully characterize the prediction uncertainty state at the current moment.
[0092] Step S3: Based on the error persistence index and error mutation intensity index in the error feature vector, perform a nonlinear mapping operation to quantify the impact of the prediction error on the power regulation capability, and generate a power regulation elasticity coefficient to characterize the adaptive regulation capability of each charging station in topology iteration. Specifically, this includes:
[0093] S3.1: Obtain the error persistence index and error mutation intensity index from the error feature vector, construct a power regulation elasticity coefficient mapping model based on a preset nonlinear activation function family, input the error persistence index as the first input variable and the error mutation intensity index as the second input variable into the power regulation elasticity coefficient mapping model, generate an initial mapping weight matrix through bivariate coupling calculation, and establish the basic nonlinear correlation between the prediction error characteristics and the regulation capability.
[0094] Error persistence index and error mutation intensity index are separated and extracted from the error feature vector and used as bivariate input sources for the power regulation elasticity coefficient mapping model.
[0095] A family of nonlinear activation functions based on the hyperbolic tangent function is constructed. The dimension of the initial mapping weight matrix is set to 2×2, which corresponds to the coupling weights and bias terms of the persistence index and the mutation intensity index, respectively.
[0096] The error persistence index is input into the first channel of the mapping model, and the inhibitory effect of long-term deviation on the regulation capability is quantified by the exponential decay characteristic to generate the persistence basic response value.
[0097] The error mutation intensity index is input into the second channel of the mapping model, and the threshold cutoff characteristic is used to quantify the excitation effect of short-term fluctuations on the regulation sensitivity, generating a mutational basic response value.
[0098] Intermediate coupling states are generated through bivariate nonlinear coupling calculations. Matrix transpose and self-multiplication are performed on the intermediate coupling state vector to establish the nonlinear projection relationship from the prediction error feature space to the power regulation capability space, generating an initial mapping weight matrix.
[0099] Through the above nonlinear mapping operation, the two-dimensional error characteristics of the previous step are transformed into an initial mapping weight matrix that characterizes the adaptive adjustment potential of the nodes, thereby realizing the quantitative correlation between prediction uncertainty and adjustment weight.
[0100] S3.2: The initial mapping weight matrix is used to perform exponential decay processing on the error persistence index to generate a persistence suppression factor. At the same time, the initial mapping weight matrix is used to perform threshold truncation processing on the error mutation intensity index to generate a mutation amplification factor. By quantifying the negative impact of long-term cumulative error and the positive response requirements of short-term drastic fluctuations respectively, a persistence suppression factor containing time series stability characteristics and a mutation amplification factor containing transient sensitivity characteristics are obtained.
[0101] Obtain the error persistence index and error mutation intensity index from the initial mapping weight matrix and error eigenvector.
[0102] The initial mapping weight matrix is used to perform exponential decay processing on the persistent error features to generate a persistence suppression factor, and the threshold truncation processing on the abrupt error features to generate a mutation amplification factor.
[0103] By mapping the persistent inhibition factor and the amplification factor to the temporal stability dimension and the transient sensitivity dimension, respectively, an intermediate parameter set containing dual features is formed.
[0104] By employing exponential decay and threshold truncation techniques, the error characteristics of the previous step are transformed into persistent suppression factors and abrupt amplification factors, thereby achieving the expected technical effect of quantifying the long-term cumulative negative impact and responding positively to short-term drastic fluctuations.
[0105] For example, suppose a charging station has an error persistence characteristic value of 6 in the current scheduling cycle, exceeding a preset tolerance threshold of 5, and an error mutation characteristic value of 0.4, which does not exceed the mutation limit of 1.0. After applying exponential decay processing to the persistence characteristic value of 6, a lower persistence suppression factor is obtained, meaning that long-term deviation accumulation will weaken the station's regulation initiative. After threshold truncation processing to the mutation characteristic value of 0.4, it is retained as a mutation amplification factor, meaning that the station still maintains normal response sensitivity to sudden load changes. After weighted fusion of the above two factors, and after amplitude constraint and smoothing filtering, the final output power regulation elasticity coefficient is approximately 0.3, indicating that the station has a moderately weak regulation participation capability under the current prediction uncertainty. This processing effectively suppresses the regulation rigidity caused by long-term deviations, while retaining the ability to respond quickly to sudden load fluctuations, significantly improving the stability and flexibility of power allocation.
[0106] S3.3: Based on the continuous suppression factor and the abrupt amplification factor, a weighted fusion operation is performed. The maximum adjustable power range of the charging station is introduced as a normalization benchmark to perform amplitude constraint processing on the fusion result, eliminate dimensional differences and prevent the adjustment coefficient from exceeding the physical equipment limit, and generate intermediate adjustment coefficients after physical boundary correction to ensure that the generated adjustment parameters strictly conform to the physical operation constraints of the nodes in the joint topology graph.
[0107] Obtain the persistence suppression factor and abrupt amplification factor output from step S3.2, and use them as input variables for the weighted fusion operation. Read the local maximum adjustable power range parameter of the charging station, which characterizes the power adjustment limit boundary of the physical equipment.
[0108] The weighted fusion of the persistent suppression factor and the abrupt amplification factor is performed, and the maximum adjustable power range of each charging station is introduced as a normalization benchmark to perform amplitude constraint processing on the fusion result, eliminate dimensional differences and prevent the adjustment coefficient from exceeding the physical equipment limit, and generate intermediate adjustment coefficients after physical boundary correction.
[0109] Execute the physical boundary truncation logic to determine whether the intermediate adjustment coefficient exceeds the [0,1] interval. If the intermediate adjustment coefficient is greater than 1, it is forcibly set to 1; if it is less than 0, it is forcibly set to 0.
[0110] By using the above processing method, the two-factor characteristics of the previous step are transformed into intermediate adjustment coefficients that conform to the physical device limits, thereby realizing the physical feasibility verification and dimensional standardization of the adjustment parameters, and ensuring that the power commands in the subsequent distributed iteration process do not exceed the limits.
[0111] For example, a charging station has a persistence suppression factor of 0.6, a mutation amplification factor of 0.8, weighting coefficients w1 and w2 of 0.3, and epsilon of 0.001. The calculated initial adjustment coefficient is approximately 1.86. The station's maximum adjustable power range is 100kW. Using 1.86 as an intermediate adjustment coefficient for boundary judgment, a truncation operation is performed because 1.86 is greater than 1, resulting in a final intermediate adjustment coefficient of 1.0. This process eliminates the abnormal amplification of the coefficient caused by prediction error fluctuations, ensuring that the adjustment command is executed within the safe range of the equipment.
[0112] S3.4: Perform dynamic smoothing filtering on the intermediate adjustment coefficients to filter out coefficient oscillation noise caused by high-frequency jitter in the prediction residual sequence. Use the sliding window averaging algorithm to update the intermediate adjustment coefficients at the current time and combine them with the intermediate adjustment coefficients at historical time to generate smoothed power adjustment elastic coefficients. This ensures a fast response to prediction errors while improving the continuity and stability of the power adjustment elastic coefficients in the time dimension.
[0113] The intermediate adjustment coefficients, which have undergone physical boundary correction, are received from the output of step S3.3 and used as the initial input data sequence for dynamic smoothing filtering.
[0114] An exponentially weighted moving average filter model based on a time sliding window is constructed, and the sliding window length and forgetting factor parameters are set to balance the response speed and noise suppression capability.
[0115] Obtain the intermediate adjustment coefficient value at the current time t, and retrieve the smoothed power adjustment elastic coefficient sequence from historical times t-1 to tN to form a complete time domain data window.
[0116] An exponential weighting algorithm is used to assign decreasing weights to coefficients at different times within a window, with higher weights for recent data to preserve abrupt changes and lower weights for older data to filter out high-frequency fluctuations. A sliding window averaging algorithm is used to update the intermediate adjustment coefficients at the current time, and these are combined with the intermediate adjustment coefficients from historical times to generate a smoothed power adjustment elasticity coefficient. The power adjustment elasticity coefficient characterizes the adaptive adjustment capability of each charging station during topology iteration, and its value dynamically changes as the prediction error increases.
[0117] The calculated smoothing coefficients are subjected to monotonicity constraint checks. If the rate of change of the coefficients exceeds the preset physical response limit threshold, the coefficients are forcibly truncated to the maximum allowable change range.
[0118] The corrected smoothing coefficient is stored in the local status register, and the historical sequence cache is updated to provide a benchmark reference value for the iterative calculation of the next scheduling cycle.
[0119] By using dynamic smoothing filtering, the intermediate adjustment coefficients from the previous step are transformed into smooth power adjustment elastic coefficients with time continuity and noise resistance, thereby achieving the expected technical effect of eliminating high-frequency jitter interference in the prediction residuals and ensuring the stability of distributed consistency iterative convergence.
[0120] For example, the sliding window length N is set to 5 scheduling cycles, and the smoothing factor is set to 0.7. When a charging station's intermediate adjustment coefficient drops sharply from 0.8 to 0.2 at time t due to a predictive abrupt change, if the smoothing coefficient at the previous time was 0.75, the calculated smoothed power adjustment elasticity coefficient at time t is 0.365. This result avoids the violent oscillation of power commands from neighboring stations caused by a direct jump in the coefficient to 0.2, while preserving the downward trend. If the preset maximum rate of change is 0.2, then because |0.365-0.75|=0.385>0.2, the system corrects the coefficient to 0.55. This processing significantly improves the stability of power allocation commands in the time dimension and prevents invalid power adjustment actions caused by predictive noise.
[0121] S3.5: Encapsulate the smoothed power regulation elasticity coefficient into a standardized data format and bind it to the corresponding charging station node number. Output the final power regulation elasticity coefficient to the control interface of the distributed consensus iteration process, so that each charging station can dynamically adjust its own convergence speed when exchanging power allocation suggestions between topological adjacent stations according to the power regulation elasticity coefficient, and realize adaptive cooperative power allocation based on the perception of prediction uncertainty.
[0122] like Figure 2 As shown, step S4: Utilizing the adjacency relationships and power adjustment elasticity coefficients in the joint topology graph, a distributed consensus iteration process is initiated, ensuring that each charging station only exchanges current power allocation suggestions with its topology neighbors, and generates local power command intermediate variables through a weighted average update rule. Specifically, this includes:
[0123] S4.1: Based on the bidirectional edge connection relationship in the joint topology graph, the set of direct neighbor nodes of each charging station vertex is extracted to obtain a list of neighbor nodes containing neighbor station identifiers and corresponding edge weights, which serves as the communication routing basis for subsequent data exchange.
[0124] The adjacency matrix data of the joint topology graph is read, and the row or column vectors corresponding to the current charging station node are extracted. The connectivity states represented by the non-zero elements are identified. The identified non-zero elements are indexed and mapped to obtain the unique identifiers of the directly connected neighboring charging stations, constructing an initial set of neighbor node IDs. Simultaneously, the weight attribute data of the corresponding edges in the topology graph is read, and each neighboring node ID is bound to its associated edge weight value as a key-value pair. The bound data is validated, and invalid neighboring nodes marked as interrupted or timed out are removed, ensuring that the remaining nodes have real-time data interaction capabilities. The validated neighboring node IDs and corresponding edge weights are serialized and encapsulated according to a preset data structure to generate a list of neighboring nodes containing neighboring station identifiers and corresponding edge weights. Through the above processing, the static topology structure generated in the previous step is transformed into a dynamically traversable communication routing basis, achieving efficient indexing and accurate positioning of neighbor information during distributed iteration.
[0125] For example, a city's charging network contains 5 charging stations, numbered CS1 to CS5. The adjacency matrix row vector of CS1 is [0, 0.8, 0, 0.6, 0], indicating that CS1 is connected to CS2 and CS4. The non-zero element indices are extracted to obtain the neighbor ID set {CS2, CS4}. The edge weight data is read; the weight of CS1-CS2 is 0.8, and the weight of CS1-CS4 is 0.6. The communication status is verified, assuming the CS4 link is normal and the CS2 link has timed out. After removing CS2, the valid neighbor {CS4} is retained. A neighbor node list is generated: [{"id":"CS4","weight":0.6}]. This list serves as the data input for step S4.2, ensuring that power suggestion values are only sent to reachable neighbor stations, avoiding invalid communication overhead, and significantly improving the real-time performance and reliability of distributed iteration.
[0126] S4.2: Utilize the communication link status in the neighbor node list to encapsulate and package the current power allocation suggestion value and power adjustment elasticity coefficient of each charging station to generate a neighbor interaction data packet containing scalar power values and adaptive gain factors, which is used for low-bandwidth transmission between topological neighboring stations.
[0127] Based on the adjacent node list obtained in S4.1, the power allocation suggestion value of each charging station at the current time and the power adjustment elasticity coefficient calculated by S3 are read. The power allocation suggestion value is truncated to two decimal places to eliminate redundant bits in communication transmission.
[0128] The power regulation elasticity coefficient is normalized and mapped to an unsigned 8-bit integer range from 0 to 255, achieving extreme data compression.
[0129] Construct a standard frame structure for neighbor interaction data packets, including frame header identifier, source node ID, destination node ID, payload length, and checksum field.
[0130] The truncated power allocation recommendation is encapsulated as a 4-byte single-precision floating-point data block and filled into the first offset position of the data packet payload area.
[0131] The encoded power regulation elasticity coefficient is encapsulated into a 1-byte integer data block and filled into the second offset position of the data packet payload area. A cyclic redundancy check (CRC) algorithm is used to calculate the overall checksum of the data packet and fill it into the frame end check field to ensure data transmission integrity. The generated neighbor interaction data packets are sent to the receive buffer of the adjacent station in the topology via the underlying communication protocol stack.
[0132] Through the aforementioned lightweight encapsulation and encoding processing methods, the multi-dimensional state parameters of the previous step are transformed into standardized interactive data packets with low bandwidth consumption, enabling high-frequency, low-latency state synchronization among charging station groups and providing an efficient data exchange foundation for distributed consistency iteration.
[0133] For example, charging station A sends interactive data to neighboring station B. The suggested power value is 120.567kW, truncated to 120.57kW. The resilience factor is 0.85, mapped to the integer 2^17 (0.85 * 255 rounded down). The data packet payload contains 4 bytes of floating-point number 120.57 and 1 byte of integer 2^17. A CRC-16 checksum is generated. The final result is a compact data packet with a total length of 12 bytes. Compared to uncompressed JSON format data, communication overhead is reduced by more than 90%, significantly improving transmission success rate and real-time performance in weak network environments.
[0134] S4.3: Receive neighbor interaction data packets sent from all topological adjacent stations, and perform analytical separation processing on the scalar power values and adaptive gain factors in the packets to reconstruct the neighbor station power suggestion value sequence and the neighbor station elasticity coefficient sequence, providing a complete set of input variables for local consistency calculation.
[0135] Receive neighbor interaction data packets from all directly adjacent sites in the joint topology graph. These packets contain encrypted or verified scalar power recommendations and corresponding power adjustment resilience coefficients. Perform protocol parsing on the received raw data stream, extracting the source node identifier from the packet header according to a predefined data frame structure, and establishing a mapping index between data sources and topological adjacency relationships to ensure that subsequent calculations are based solely on data from valid connected links.
[0136] The extracted scalar power values are converted from the communication protocol format to floating-point physical quantity units, and the adaptive gain factor, i.e. the power adjustment elasticity coefficient, is parsed to form a structured tuple sequence containing neighbor station ID, power suggestion value, and elasticity coefficient.
[0137] Perform integrity checks on the structured tuple sequence, remove records with missing fields due to communication packet loss or timeout, and filter out the latest valid data within the current scheduling cycle based on the timestamp alignment mechanism to generate a cleaned neighbor station status dataset.
[0138] Based on the cleaned dataset, a power suggestion value sequence and a resilience coefficient sequence for neighboring stations are constructed respectively. The power suggestion value sequence reflects the expected contribution of neighboring stations to the global power balance, while the resilience coefficient sequence characterizes the sensitivity of neighboring stations to their own prediction errors.
[0139] Through the above-described analytical separation and reconstruction process, discrete network communication data is transformed into a set of collaborative control variables with physical meaning, providing accurate and synchronous input parameters for the local consistency update algorithm, and ensuring the accuracy and real-time performance of information interaction during distributed iteration.
[0140] For example, charging station A receives data packets from neighboring stations B and C. Parsing reveals that station B's recommended power output is 120kW with a flexibility factor of 0.8; station C's recommended power output is 150kW with a flexibility factor of 0.6. After verifying the consistency of the timestamps, the system constructs a sequence P. rec =[120,150] and alpha rec =[0.8,0.6]. This sequence is directly used for subsequent weighted average calculation, ensuring accurate alignment of the state information of each node in multi-station collaboration, eliminating data asynchrony problems caused by communication delays, and significantly improving the stability of distributed convergence.
[0141] S4.4: Based on the neighboring station power suggestion value sequence, the neighboring station elasticity coefficient sequence, and the local station's power regulation elasticity coefficient, a weighted average consensus update algorithm is executed to calculate the local power command intermediate variable that integrates global collaborative information and local regulation capability, thereby achieving autonomous convergence of the power command under topological constraints.
[0142] Obtain the parsed neighbor station power suggestion value sequence, neighbor station elasticity coefficient sequence, and local station power adjustment elasticity coefficient, and construct a distributed consensus update model based on topology weights.
[0143] Extract edge weight data from the list of adjacent nodes, and calculate the relative trust factor of each neighboring node by combining it with the neighboring station elasticity coefficient. This factor represents the influence weight of the neighboring station's suggestion value on the decision of this station.
[0144] A weighted average consensus algorithm is used to merge the current power allocation recommendation value of this station with the power recommendation values of each neighboring station. The weight of the recommendation value of this station is determined by the reciprocal of the power adjustment elasticity coefficient of this station, so as to reflect the low confidence under high uncertainty.
[0145] The intermediate variables for the local power command are calculated using the following formula:
[0146]
[0147] in, For the updated local power command intermediate variables, This is the current recommended power value for this site. For the set of adjacent nodes, For nodes and Topological edge weights between them For neighboring stations The power regulation elasticity coefficient, For neighboring stations Recommended power values.
[0148] Physical boundary pre-verification is performed on the intermediate variables of the calculated local power command to eliminate abnormal iterative values that exceed the maximum adjustable power range of the charging station, ensuring the physical feasibility of the iteration process.
[0149] By using a weighted average consensus update algorithm, the neighbor station interaction data from the previous step is transformed into an intermediate variable of local power command that integrates global collaborative information and local regulation capabilities, thereby enabling the autonomous convergence of power command under topological constraints.
[0150] For example, the current power recommendation value of charging station A is set to 100kW, and the power adjustment flexibility coefficient is 0.8. Its neighboring station B has a power recommendation value of 120kW and a flexibility coefficient of 0.9; its neighboring station C has a power recommendation value of 90kW and a flexibility coefficient of 0.7. The topology edge weight between A and B is 0.6, and the topology edge weight between A and C is 0.4. Substituting into the formula, the incremental term is 0.6×0.9×(120-100)+0.4×0.7×(90-100)=10.8-2.8=8.0kW. The updated local power command intermediate variable is 100+8.0=108.0kW. This result effectively integrates the suppressive effect of the high power recommendation and low flexibility coefficient of neighboring stations, achieving smooth convergence.
[0151] Step S5: Determine whether the potential over-limit risk of the local power command intermediate variable and the total capacity of the feeder meets the soft constraint condition of power balance. If an over-limit risk exists, trigger the local scaling mechanism, distribute the pressure to neighboring stations according to the topology edge weight ratio, and correct the local power command intermediate variable to generate a coordinated power command sequence that meets the feeder capacity constraint. Specifically, this includes:
[0152] S5.1: Obtain the local power command intermediate variable and the real-time total capacity data of the feeder, execute the over-limit risk discrimination logic based on the power balance soft constraint condition, and calculate the difference between the local power command intermediate variable and the feeder total capacity to generate the power over-limit risk flag.
[0153] Receive the local power command intermediate variable generated in step S4, which represents the preliminary power allocation recommendation value calculated by the charging station based on neighborhood consensus within the current iteration cycle.
[0154] The total capacity threshold data of the feeder at the current moment is obtained from the real-time monitoring system of the distribution network. This data serves as the physical hard constraint boundary for power distribution, reflecting the thermal stability limit of the transformer and the line.
[0155] A power over-limit risk judgment logic is constructed, which performs an algebraic subtraction operation between the intermediate variable of the local power command and the total capacity threshold of the feeder to calculate the remaining available capacity margin.
[0156] The power over-limit deviation is calculated using the following formula:
[0157]
[0158] in, This is the power over-limit deviation. This is an intermediate variable for local power commands. This represents the real-time total capacity of the feeder.
[0159] The calculated power over-limit deviation is subjected to sign judgment and threshold comparison. If the deviation is greater than zero, it is determined that there is an overload risk; if the deviation is less than or equal to the preset safety tolerance threshold, it is determined that it is in the safe operating range.
[0160] A binary power over-limit risk flag is generated based on the judgment result. The flag is set to 1 when there is an overload risk, and otherwise set to 0. This flag serves as a logical enable signal to trigger the subsequent local scaling mechanism.
[0161] Through the above-mentioned over-limit risk judgment logic, the intermediate variable of the power command in the previous step is transformed into a clear over-limit risk status identifier, realizing rapid pre-verification of feeder capacity constraints and providing an accurate trigger basis for subsequent distributed pressure sharing.
[0162] For example, after consistency iteration, the intermediate variable of the local power command obtained by a charging station is 850kW, while the real-time total capacity limit of its 10kV feeder is 800kW. The system performs a subtraction operation and obtains a power over-limit deviation of 50kW. Since 50kW is greater than 0, the system determines that there is an over-limit risk and generates a power over-limit risk flag bit of 1. This flag bit immediately triggers the local scaling mechanism in step S5.2, reads the topology edge weights for pressure distribution, avoids distribution network overload tripping caused by directly issuing commands, and significantly improves the physical feasibility of dispatch commands and grid security.
[0163] like Figure 3 As shown, Figure 3 This is the flowchart of the algorithm for generating the amortization coefficient matrix in step S5.2. S5.2: When the power over-limit risk flag indicates the existence of over-limit risk, read the topology edge weight data connected to this station in the joint topology map, and perform normalization operation based on the topology edge weight ratio to generate the amortization coefficient matrix.
[0164] In response to the power limit risk flag generated by S5.1, when the flag indicates a risk of exceeding the limit, the system immediately reads the weight data set of all topological edges directly connected to the local node from the local storage area of the joint topology graph. This weight data set contains multi-dimensional fused values reflecting spatial proximity, electrical coupling, and communication reachability, characterizing the physical potential and communication reliability of each neighboring station to participate in power coordinated regulation.
[0165] The read topology edge weight data is filtered for validity, removing invalid connection weights caused by communication link interruptions or node offline status, and retaining the weight values of adjacent edges in an active state to construct the original weight vector. This ensures that subsequent allocation calculations are based only on real and usable physical and logical connections, avoiding the distribution of adjustment instructions to unreachable nodes.
[0166] The sum of all effective weights in the original weight vector is calculated and used as the normalized baseline denominator. This sum reflects the base of the total collaborative capability that this station can reach under the current topology, providing a mathematical basis for converting absolute weights into relative proportions.
[0167] Using the normalized baseline denominator, a division operation is performed on each element of the original weight vector to generate the corresponding amortization coefficient. The specific calculation formula is as follows:
[0168]
[0169] in, Indicates site Towards neighboring stations The allocation coefficient for the pressure regulation. For the site With neighboring stations Topological edge weights between them For the site The set of valid adjacent stations.
[0170] The calculated allocation coefficients are arranged in order of neighboring station identifiers and assembled into an allocation coefficient matrix. Each element in this matrix strictly corresponds to a specific neighboring station, and the sum of all elements is always equal to 1, satisfying the probability distribution characteristics and ensuring that the regulation pressure is conserved and does not overflow across the entire network.
[0171] Through the above normalization operation, the risk of exceeding the limit identified in the previous step is transformed into quantitative pressure sharing ratio data based on topological potential, so as to achieve fair and accurate allocation of regulation responsibility among adjacent sites, and lay the data foundation for the subsequent generation of distributed regulation pressure vector.
[0172] For example, charging station A detects a risk of exceeding power limits. Its topology graph shows three valid neighboring stations B, C, and D. The edge weights are read as w. AB =0.6 (strong electrical coupling), w AC =0.3 (medium communication reachable), w AD =0.1 (weak spatial neighbor). The total weight is calculated as 0.6 + 0.3 + 0.1 = 1.0. Normalization is then performed: α AB =0.6 / 1.0=0.6, α AC =0.3 / 1.0=0.3, α AD =0.1 / 1.0=0.1. This generates the allocation coefficient matrix [0.6,0.3,0.1]. This matrix shows that 60% of the regulation pressure will be allocated to Station B, which has the strongest electrical coupling, significantly improving coordination efficiency and reducing communication latency.
[0173] S5.3: The excessive deviation of the intermediate variable of the local power command is weighted and decomposed using the apportionment coefficient matrix. Combined with the communication status of adjacent stations, a distributed regulation pressure vector is generated to achieve accurate mapping of regulation pressure to adjacent stations.
[0174] Obtain the allocation coefficient matrix generated in S5.2 and the power over-limit deviation calculated in S5.1. Perform element-by-element multiplication of each element in the allocation coefficient matrix with the over-limit deviation to generate the preliminary regulation pressure numerical sequence for each adjacent station.
[0175] The initial pressure numerical sequence is subjected to communication state weighting correction. The real-time communication link quality index of each station in the adjacent node list obtained by S4.1 is read, and the communication packet loss rate or delay jitter value is converted into communication reliability weighting factor.
[0176] The initial regulation pressure numerical sequence is attenuated by using a communication reliability weighting factor. The regulation pressure shared by neighboring stations with poor communication quality is reduced proportionally to ensure the executability of regulation commands under weak communication conditions.
[0177] The adjusted pressure values of each neighboring station after communication status correction are combined with the remaining deviation that the current station needs to absorb on its own to construct a distributed adjustment pressure vector that includes the current station and all valid communication neighboring stations.
[0178] Normalization verification is performed on the distributed regulation pressure vector to ensure that the sum of all elements in the vector is equal to the initially calculated power over-limit deviation, thus maintaining the power conservation constraint of the entire network.
[0179] Through the above weighted decomposition and communication state mapping process, the allocation coefficients of the previous step are transformed into a distributed regulation pressure vector with communication robustness, realizing accurate and reliable mapping of regulation pressure to neighboring stations.
[0180] For example, the calculated power over-limit deviation of charging station A is set to 50kW. The sharing coefficient matrix generated by S5.2 shows that the weight of neighboring station B is 0.6 and the weight of neighboring station C is 0.4. Initially, it is calculated that neighboring station B shares 30kW and neighboring station C shares 20kW. It is detected that the communication delay of neighboring station B is higher than the threshold, and the communication reliability weight factor is 0.8; the communication of neighboring station C is normal, and the weight factor is 1.0. After correction, the actual adjustment pressure of neighboring station B is adjusted to 30kW*0.8=24kW, and neighboring station C remains at 20kW. At this time, the total shared pressure is 44kW, and the remaining 6kW deviation is absorbed by station A itself. The final generated distributed adjustment pressure vector is [6kW, 24kW, 20kW], which ensures reasonable pressure distribution and total amount conservation under communication constraints.
[0181] S5.4: Based on the distributed regulation pressure vector, a local scaling mechanism is executed to dynamically correct the local power command intermediate variables, eliminate over-limit deviations, and generate a preliminary coordinated power command sequence that meets the feeder capacity constraints.
[0182] Obtain the apportionment coefficient and corresponding over-limit deviation of each adjacent station in the distributed regulation pressure vector, and use them as the calculation benchmark for local power command correction.
[0183] Based on the apportionment coefficient matrix and the over-limit deviation, a local scaling correction model is constructed to calculate the power reduction or increase that this station needs to bear.
[0184] The intermediate variables of the corrected local power command are calculated using the following formula:
[0185]
[0186] in, This is the revised local power command. This is an intermediate variable for local power commands. For the first The allocation coefficient for each charging station This represents the deviation of the total feeder capacity beyond the limit.
[0187] The calculated correction amount is subtracted from the intermediate variables of the local power command to eliminate the risk of feeder overload caused by the accumulation of prediction errors.
[0188] The corrected power command is physically truncated to ensure that its value is between the minimum and maximum adjustable power of the charging station.
[0189] If there are still remaining deviations after truncation, the remaining deviations are remapped to the topological adjacent nodes, triggering the next round of consistency iteration compensation.
[0190] By using a local scaling mechanism and boundary constraint processing, intermediate variables with the risk of exceeding limits are transformed into a preliminary coordinated power command sequence that satisfies feeder capacity constraints, thereby achieving dynamic power balance among multiple stations.
[0191] For example, a charging station's local power command intermediate variable is 120kW, and the total capacity limit of its feeder results in an over-limit deviation of 20kW. The station's allocation coefficient in the topology diagram is 0.4. According to the formula, the correction is 0.4 × 20kW = 8kW. The corrected power command is 120kW - 8kW = 112kW. If the station's minimum adjustable power is 10kW and its maximum is 150kW, then 112kW is within the legal range and is directly output. If the corrected result is 5kW, which is lower than the minimum value of 10kW, then it is truncated to 10kW, and the remaining 2kW deviation is redistributed to neighboring stations. This process significantly reduces the probability of single-station overload and improves the safety of grid operation.
[0192] S5.5: Perform consistency verification on the initial coordinated power command sequence to verify whether the corrected command simultaneously satisfies the single-station equipment boundary and the power conservation conditions of the entire network, and output the final coordinated power command sequence.
[0193] Obtain the preliminary coordinated power command sequence after correction by the local scaling mechanism, and extract the node identifiers of each charging station, the corrected power setting values, and the corresponding topological adjacency data contained therein.
[0194] Traverse each charging station node in the initial coordinated power instruction sequence, read the physical equipment technical parameter library to which the node belongs, and obtain the maximum charging power upper limit, minimum standby power lower limit, and transformer rated capacity constraint conditions.
[0195] The corrected power setting value is compared and verified with the technical parameters of the single station equipment one by one to determine whether there are any abnormal values that exceed the physical boundary, and a local equipment compliance flag is generated.
[0196] If the compliance flag of the local device indicates that the limit has been exceeded, the power setting value of the node will be forcibly clamped to the upper limit of the maximum charging power or the lower limit of the minimum standby power, and an intermediate power command truncated by the single station boundary will be generated.
[0197] The system aggregates intermediate power commands from all charging stations across the network after single-station boundary truncation, calculates the total power demand value of the entire network, and compares it with the regional total power quota issued by the upper-level dispatch center of the distribution network to calculate the deviation between the total power demand of the entire network and the regional total power quota.
[0198] If the absolute value of the total power deviation of the entire network is greater than the preset conservation tolerance threshold, the global proportional fine-tuning algorithm is activated to perform secondary amortization correction on the deviation based on the proportion of intermediate power commands of each station to the total commands of the entire network.
[0199] The power command, after global proportional fine-tuning, is remapped back to each charging station node, and the final consistency encapsulation is performed, adding timestamp tags and digital signature verification codes.
[0200] Through the aforementioned dual verification and dynamic correction processing methods, the initial coordinated power command is transformed into a final coordinated power command sequence that simultaneously satisfies the physical boundary of a single station and the power conservation conditions of the entire network, thereby achieving the physical executability of the dispatch command and the security of the power grid.
[0201] For example, a charging station cluster contains 5 nodes. In the initial coordinated power command sequence, node 3's command value is 120kW, but its transformer's rated capacity limit is 100kW. The system identifies this over-limit risk and clamps node 3's command to 100kW. At this point, the total command for the entire network changes from 500kW to 480kW, while the regional total power quota is 500kW, resulting in a 20kW deviation. The system calculates the weight ratio of the remaining 4 nodes and distributes the 20kW deviation according to the weights of nodes 1, 2, 4, and 5, increasing node 1 by 5kW, node 2 by 5kW, node 4 by 5kW, and node 5 by 5kW. In the final output command sequence, node 3 is 100kW, and the other nodes increase accordingly, with the sum strictly equal to 500kW. Furthermore, all nodes do not exceed their respective equipment limits, ensuring the safe issuance of commands.
[0202] Step S6: Monitor changes in the charging station group topology or the status of new station access. Based on the updated joint topology graph, reinitialize the boundary conditions of the distributed consensus iteration process, drive the cooperative power command sequence to automatically reconverge to the new global coordination solution, and output the final dynamic power allocation command sequence. Specifically, this includes:
[0203] S6.1: Real-time acquisition and anomaly detection processing of heartbeat messages and distribution network switch status signals in the communication link of charging station groups to identify charging station offline fault events or new station access registration events, and generate topology change event identifiers containing topology change type and affected node number.
[0204] The S6.1 sub-step aims to perceive the dynamic changes in the charging station cluster topology in real time, providing accurate event triggering basis for subsequent map reconstruction. This step achieves accurate identification of offline faults and new station access through multi-source data fusion and state machine logic.
[0205] Heartbeat packets in the communication link of the charging station group are collected periodically. The heartbeat timeout threshold is set to 3 times the communication cycle. The number of consecutively lost heartbeat packets is counted to construct a communication connectivity time series.
[0206] Synchronously acquire the status signals of the distribution network side switches, including the opening and closing status of circuit breakers and the position signals of isolating switches, and map the discrete binary status to electrical connection on / off indicators.
[0207] The system performs a logical AND operation based on the communication connectivity time series and the electrical connection on / off flags. When the heartbeat times out and the electrical connection is disconnected, it is determined as a charging station offline fault event; when a new device registration request is detected and the electrical connection is established, it is determined as a new station access registration event.
[0208] For the identified topology change event, the unique identifier ID of the affected node is extracted, and an event type code is generated by combining the timestamp of the event, where 01 represents offline and 02 represents connected.
[0209] The node ID, event type code, and timestamp are encapsulated into a standardized topology change event identifier object, which serves as the input parameter to drive subsequent graph updates.
[0210] By using a multi-source state joint discrimination processing method, the original communication and electrical signals are transformed into structured topology change event identifiers, achieving millisecond-level response and precise location of network structure changes, and ensuring that the distributed iterative process can adapt to changes in physical topology in a timely manner.
[0211] For example, the heartbeat period is set to 1 second, and the timeout threshold is set to 3 seconds. When station A fails to send a heartbeat three times consecutively and the distribution network switch is open, the system generates an event identifier {NodeID:"Station_A",Type:01,Timestamp:1715668800}. If a new station B initiates registration and the switch is closed, an event identifier {NodeID:"Station_B",Type:02,Timestamp:1715668805} is generated. This mechanism effectively avoids misjudgments caused by a single communication jitter and significantly improves the robustness of topology awareness.
[0212] S6.2: Based on the topology change event identifier, perform adjacency matrix reconstruction and edge weight update operations on the original joint topology graph to remove invalid node connections or add valid node connections, and generate an updated joint topology graph that reflects the current physical connection status.
[0213] Receive the topology change event identifier generated by S6.1, and parse out the change type and the set of affected node numbers. Retrieve the adjacency matrix data block in the original joint topology graph based on the change type. If the change type is a node offline, locate the corresponding row and column index of that node in the adjacency matrix.
[0214] Set all elements in the selected row and column to zero, severing the logical connection between this node and all other nodes in the network. Simultaneously clear all bidirectional edge weight data associated with this node, releasing the corresponding storage resources. If the change type is new site access, obtain the new site's geographical coordinates, electrical parameters, and communication status. Calculate the electrical distance and communication reachability indicators between the new site and all existing sites.
[0215] Based on a preset threshold, a connection is established, generating a new adjacency vector. This vector is appended to the end of the adjacency matrix, expanding the matrix dimension to accommodate the new node. For the retained valid connections, the edge weights affected by topology changes are recalculated. A weighted fusion algorithm based on spatial proximity, electrical coupling, and communication reachability is used to update the weight values. All active nodes are traversed to verify the symmetry and consistency of adjacency relationships. Isolated subgraphs or broken paths caused by node additions or removals are corrected to ensure graph connectivity.
[0216] Through the above reconstruction and update operations, topology change events are transformed into an updated joint topology graph that reflects the current physical connection state, thereby achieving dynamic real-time mapping with topology awareness.
[0217] For example, when charging station 3 is detected to be offline, the system reads the original 10x10 adjacency matrix. The third row and third column are set to 0, disconnecting it from stations 1, 2, and 4. The original edge weight W(3,4)=0.8 is cleared. If station 11 is newly connected, its electrical distance to station 4 is calculated to be 50 meters, less than the threshold of 100 meters, and the communication latency is less than 20ms. An 11th row / column is added to the matrix, A(11,4)=1 is set, and a new weight W(11,4)=0.9 is calculated. The weights between other unaffected nodes remain unchanged. Finally, a new 11x11 adjacency matrix is generated, accurately reflecting the current network topology and ensuring the correct foundation for subsequent iterations to converge.
[0218] S6.3: Utilize the node degree distribution information in the updated joint topology graph to perform boundary condition reset and normalization processing on the initial power proposal value and power adjustment elasticity coefficient of the distributed consensus iteration process, so as to eliminate the interference of the old topology residual state on the convergence process and generate the iteration process initialization boundary conditions that satisfy the new topology constraints.
[0219] Read the adjacency matrix data of each node in the updated joint topology graph, count the number of neighboring nodes directly connected to each charging station node, and calculate the locality value of each node.
[0220] Based on the set of local degree values, the degree of all points in the site is standardized using the maximum-minimum normalization algorithm to generate a node centrality index with a value between 0 and 1, which is used to characterize the degree of connectivity of the site in the current topology.
[0221] The system retrieves historical cached power recommendation values from each station at the end of the previous iteration, weights and fuses them with the corresponding node centrality index, and calculates the initial power recommendation value after reset by combining this with the station's maximum adjustable power. This approach reduces the impact of offline node historical data on the new topology, while leveraging centrality indicators to accelerate the convergence speed of power allocation for newly connected nodes.
[0222] Extract the current power regulation elasticity coefficient of each station, and perform secondary smoothing filtering in combination with the node centrality index to eliminate coefficient jump noise caused by topological abrupt changes, thus ensuring the continuity of regulation capability.
[0223] The reset initial power recommendation value and the smoothed power adjustment elasticity coefficient are encapsulated into a standardized boundary condition data packet, timestamped, and broadcast to the local consistency iteration module.
[0224] The initial boundary conditions for the iterative process include at least the initial power recommendation value after each charging station is reset and the smoothed power adjustment elastic coefficient.
[0225] Through the above chain derivation process, the topology update result of the previous step is transformed into iterative initialization parameters with physical meaning and no historical bias, thereby achieving the expected technical effect of fast reconvergence of distributed algorithms under dynamic topology.
[0226] For example, in a charging station cluster, a new station A is added. It has 3 neighboring nodes, and the maximum number of neighboring nodes in the entire network is 10. The calculated node centrality index is 0.3. Station A's historical power recommendation value is 0 (new stations have no history), its maximum adjustable power is 200kW, and the forgetting factor α is set to 0.6. According to the boundary condition reset mechanism, the initial power recommendation value for station A is set to 24kW. This initial value serves as the iteration starting point, avoiding convergence delays caused by searching from zero, and preventing grid impact caused by directly assigning high power. For the existing station B, if its number of neighbors remains unchanged, the node centrality index remains at 0.5, and its historical power recommendation value is 100kW, then the initial power recommendation value is 100kW, maintaining scheduling continuity. Actual measurements show that after adopting this reset mechanism, the average number of convergence rounds for the system after topology changes is reduced from 8 rounds to 4 rounds, significantly improving dynamic response efficiency.
[0227] S6.4: Based on the boundary conditions initialized by the iterative process, multiple rounds of distributed weighted average iterative calculations are initiated, so that each charging station exchanges local power command intermediate variables and performs local scaling correction according to the adjacency relationship in the updated joint topology graph, thereby driving the coordinated power command sequence to automatically reconverge to a new global coordinated solution that satisfies the power conservation of the entire network within a finite number of steps.
[0228] Based on the iterative process generated by S6.3, the boundary conditions are initialized, and each charging station node starts the distributed weighted average consensus iterative calculation according to the updated topology graph adjacency relationship.
[0229] Read the power suggestion value and elasticity coefficient from the intermediate variables of the local power command and the data packets of the neighboring station, and construct a local consistent update state vector.
[0230] A weighted average algorithm is used to integrate local regulation capabilities and neighbor cooperation information to calculate the power command estimate for the next iteration.
[0231] The calculated power command estimate is locally scaled to determine whether it exceeds the feeder capacity safety threshold.
[0232] If there is a risk of exceeding the limit, a distributed adjustment pressure vector is generated based on the topological edge weight ratio, and the estimated value is dynamically decayed.
[0233] The revised power command is used as input for the next iteration, and the weighted averaging and local scaling steps are repeated.
[0234] Monitor the norm of the power command difference between two adjacent iterations. When the difference is less than the preset convergence threshold or the maximum number of iterations is reached, terminate the iteration process.
[0235] Through multiple rounds of distributed weighted average iteration and local scaling correction, the initial boundary conditions are transformed into a new globally coordinated solution that satisfies the power conservation of the entire network and the feeder capacity constraint, thereby achieving fast automatic reconvergence under topology changes.
[0236] For example, a convergence threshold of 0.5kW is set, and the maximum number of iterations is 5. A charging station node i has an initial power command of 100kW, a neighboring station power of 120kW, an edge weight of 0.8, and a neighboring station elasticity coefficient of 0.9. The estimated value calculated in the first iteration is 114.4kW. If this value does not exceed the feeder limit, it is directly used as the input for the next round; if it exceeds the limit, it is scaled proportionally. After 3 iterations, the power difference between nodes decreases to 0.3kW, below the convergence threshold, and the final coordinated power command sequence is output, significantly improving the system response speed and stability.
[0237] S6.5: Perform device boundary verification and feeder capacity compliance encapsulation on the converged new global coordinated solution to generate a final dynamic power allocation instruction sequence with timestamps that can be directly sent to the edge controller.
[0238] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.
[0239] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the element or object preceding “comprising” or “including” encompasses the element or object listed following “comprising” or “including” and its equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.
[0240] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. A dynamic power allocation method for charging station groups based on load forecasting, characterized in that, include: Obtain topological association data of multiple charging stations in the charging network, and construct a joint topological map based on the topological association data to characterize the spatial proximity, electrical coupling and communication reachability between charging stations; Obtain load forecast residual sequences at multiple time scales, extract error persistence features and error mutation features from the load forecast residual sequences, and generate an error feature vector characterizing forecast uncertainty. Based on the error feature vector, the influence of the prediction error on the power regulation capability of the multiple charging stations is quantified, and a power regulation elasticity coefficient is generated. Using the adjacency relationships in the joint topology graph and the power adjustment elasticity coefficient, through distributed consensus iteration, the charging station exchanges power allocation suggestion values with neighboring stations and performs weighted updates to generate intermediate power command variables. The intermediate variables of the power command are checked for over-limit feeder capacity. If there is a risk of over-limit, the pressure is distributed according to the topological adjacency relationship to generate a cooperative power command sequence that meets the feeder capacity constraint. In response to changes in the charging station cluster topology, the joint topology graph is updated and the distributed consensus iterative reconvergence is triggered, outputting a dynamic power allocation instruction sequence.
2. The dynamic power allocation method for charging station groups based on load forecasting according to claim 1, characterized in that, Utilizing the adjacency relationships in the joint topology graph and the power adjustment elasticity coefficient, through distributed consensus iteration, charging stations exchange power allocation suggestion values with neighboring stations and perform weighted updates to generate intermediate power command variables, including: Based on the bidirectional edge connection relationship in the joint topology graph, the set of direct neighbor nodes of each charging station vertex is extracted to obtain a list of neighbor nodes containing neighbor station identifiers and corresponding edge weights, which serves as the basis for communication routing for subsequent data exchange. Using the communication link status in the neighbor node list, the current power allocation suggestion value and power adjustment elasticity coefficient of each charging station are encapsulated and packaged to generate a neighbor interaction data packet containing a scalar power value and an adaptive gain factor, which is used for low-bandwidth transmission between topological neighboring stations. It receives neighbor interaction data packets sent from all topological adjacent stations, performs analytical separation processing on the scalar power values and adaptive gain factors in the packets, and reconstructs the neighbor station power suggestion value sequence and the neighbor station elasticity coefficient sequence to provide a complete set of input variables for local consistency calculation; Based on the neighboring station power suggestion value sequence, the neighboring station elasticity coefficient sequence, and the local station's power regulation elasticity coefficient, a weighted average consensus update algorithm is executed to calculate the local power command intermediate variable that integrates global collaborative information and local regulation capability, thereby achieving autonomous convergence of the power command under topological constraints.
3. The dynamic power allocation method for charging station groups based on load forecasting according to claim 1, characterized in that, The intermediate variables of the power command are checked for over-limit feeder capacity. If over-limit risk exists, the adjustment pressure is distributed according to the topological adjacency relationship to generate a cooperative power command sequence that satisfies the feeder capacity constraint, including: Obtain the local power command intermediate variable and the real-time total capacity data of the feeder. Execute the over-limit risk judgment logic based on the power balance soft constraint condition, and calculate the difference between the local power command intermediate variable and the feeder total capacity to generate the power over-limit risk flag. When the power over-limit risk flag indicates that there is an over-limit risk, read the topology edge weight data connected to this station in the joint topology graph, and perform normalization operation based on the topology edge weight ratio to generate the amortization coefficient matrix. The excessive deviation of the intermediate variable of the local power command is weighted by using the apportionment coefficient matrix, and a distributed regulation pressure vector is generated by combining the communication status of adjacent stations to achieve accurate mapping of regulation pressure to adjacent stations. Based on the distributed regulation pressure vector, a local scaling mechanism is executed to dynamically correct the local power command intermediate variables, eliminate over-limit deviations, and generate a preliminary coordinated power command sequence that meets the feeder capacity constraints. The initial coordinated power command sequence is subjected to consistency verification to verify whether the corrected command simultaneously satisfies the single-station equipment boundary and the power conservation conditions of the entire network, and the final coordinated power command sequence is output.
4. The dynamic power allocation method for charging station groups based on load forecasting according to claim 1, characterized in that, The topology association data includes at least four of the following: geographical coordinate data of the multiple charging stations, data on the node numbers connected to the power distribution network, transformer capacity data, and real-time communication link status data.
5. The dynamic power allocation method for charging station groups based on load forecasting according to claim 1, characterized in that, The construction of the joint topology graph specifically includes: Based on the topology association data, electrical distance calculation and feeder affiliation matching are performed. The electrical coupling strength between stations is determined by using a preset electrical distance threshold and combined with feeder homogeneity analysis to generate an electrical proximity association matrix. The electrical proximity correlation matrix is subjected to communication reachability verification. After three-dimensional weighted fusion of the spatial proximity, electrical coupling and communication reachability of the electrical proximity correlation matrix, it is assembled into the joint topology map.
6. The dynamic power allocation method for charging station groups based on load forecasting according to claim 5, characterized in that, The weighting coefficients of the three-dimensional weighted fusion are dynamically adjusted according to the scheduling scenario.
7. The dynamic power allocation method for charging station groups based on load forecasting according to claim 1, characterized in that, The load forecast residual sequence is generated by differentiating the day-ahead, intraday, and real-time load forecast data with the actual monitoring data.
8. The dynamic power allocation method for charging station groups based on load forecasting according to claim 1, characterized in that, The error persistence feature is the number of time periods that continuously exceed a preset error tolerance threshold, and the error mutation feature is the rate of change of deviation between adjacent time periods.
9. A dynamic power allocation method for charging station groups based on load forecasting according to claim 1, characterized in that, The generation of the power regulation elasticity coefficient includes: The error persistence feature is exponentially decayed, the error mutation feature is thresholded, the two processed feature values are weighted and fused, and the power adjustment elasticity coefficient is output after amplitude constraint and smoothing filtering.
10. A dynamic power allocation method for charging station groups based on load forecasting according to claim 1, characterized in that, The events responding to changes in topology include: Collect heartbeat messages and switch signals to generate change event identifiers, reconstruct the joint topology map based on the change event identifiers, and reset the iterative boundary conditions.
