New energy vehicle aggregation prevention and control system based on clustering identification and regional risk modeling

Abstract

The invention provides a new energy vehicle aggregation prevention and control system based on clustering identification and regional risk modeling, and aims to solve the potential safety hazards such as thermal runaway caused by high-density vehicle aggregation. The system comprises a state acquisition module, a clustering identification module, a regional risk modeling module, a risk grade determination module, a response control scheduling module and a thermal risk prediction module. The system collects running state data such as vehicle position, battery temperature and charge state, uses a density-based spatial clustering algorithm to identify clusters, calculates risk score values of the clusters, and implements current limiting and path guiding control according to risk levels. Meanwhile, the risk change trend of a future area is predicted through the trained long-short-term memory neural network model, and early warning and response are achieved in advance. The system is suitable for safety management scenes of new energy vehicles in high-density areas such as charging stations and underground garages.

Description

Technical Field

[0001] The present invention relates to the technical field of the intersection of intelligent traffic control and artificial intelligence algorithms, and particularly relates to a new energy vehicle aggregation prevention and control system based on clustering recognition and regional risk modeling. Background Art

[0002] With the large-scale popularization of new energy vehicles, centralized parking and charging scenarios have gradually formed in cities. Especially in areas such as underground garages, bus terminals, commercial complexes, and residential communities, a large number of new energy vehicles are highly aggregated in space. The intensive parking and charging of multiple new energy vehicles in the physical space may lead to the following safety problems: thermal runaway chain propagation; local temperature rise, inducing battery failure; accumulation of smoke, high temperature, and gas causing secondary disasters.

[0003] The existing technologies mainly implement thermal safety monitoring through the battery management system (BMS) of individual vehicles, or rely on traditional fire-fighting equipment for post-disposal. There are also some platforms that attempt to control local temperature rise or charging power based on fixed thresholds. Relevant research mostly focuses on individual vehicle risk assessment or global charging load optimization.

[0004] However, the above technologies have multiple deficiencies in practical applications. On the one hand, current methods generally lack the ability to spatially identify the aggregation behavior of multiple vehicles and cannot dynamically identify the aggregation form of vehicle clusters and their associated risks. On the other hand, the risk assessment mechanism usually ignores the comprehensive influence of vehicle status, temperature rise changes, and local environmental load, making it difficult to conduct hierarchical management of potential risks. At the same time, most of the existing response strategies are based on static rules or manual scheduling, making it difficult to form a real-time and efficient linkage disposal mechanism, and it is also difficult to meet the requirements of early warning and active control in high-density scenarios.

[0005] Therefore, it is urgent to propose a risk prevention and control technical path for new energy vehicle aggregation scenarios to achieve full-process dynamic management from identification, assessment, determination to response. Summary of the Invention

[0006] The present application provides a new energy vehicle aggregation prevention and control system based on clustering recognition and regional risk modeling to achieve real-time identification and early intervention of the risks of high-density aggregation of new energy vehicles.

[0007] The present application provides a new energy vehicle aggregation prevention and control system based on clustering recognition and regional risk modeling, including: A status acquisition module for acquiring the operation status data of new energy vehicles in the target area, where the operation status data includes the geographical position coordinates, battery temperature data, state of charge data, and vehicle operation status label of the vehicle; A clustering recognition module, which is used to identify aggregation clusters according to the operation status data by using a density-based spatial clustering algorithm, and obtain the aggregation characteristics of the aggregation clusters; A regional risk modeling module, which is used to calculate the risk score value of each aggregation cluster according to the aggregation cluster characteristics provided by the clustering recognition module; the risk score value is obtained according to the vehicle density per unit area, the average temperature value of the aggregation cluster, the proportion value of high-temperature vehicles, the proportion value of abnormally state vehicles, and the charging load ratio of the target area; A risk level determination module, which is used to determine the risk level of the corresponding aggregation cluster according to the risk score value; A response control and scheduling module, which is used to execute current-limiting control and path guidance measures according to the risk level, and provide response results to an external platform; A thermal risk prediction module, which is used to construct a time series feature data set based on a sliding window according to the operation status data, and use a pre-trained long short-term memory neural network model to predict the potential risk score trend of each target area within a preset future time range; output the predicted potential risk score trend to the risk level determination module and the response control and scheduling module to execute current-limiting or guidance strategies in advance.

[0008] The beneficial effects of this application mainly include: (1) It can identify the spatial aggregation behavior of new energy vehicles in real time, effectively extract high-risk aggregation clusters through density clustering algorithms, realize the technological leap from individual monitoring to group recognition, and significantly improve the perception ability of vehicle aggregation risks in high-density scenarios. (2) By introducing a multi-factor risk modeling method, factors such as vehicle distribution density, temperature level, abnormal state ratio, and regional charging load are integrated into a risk score value, improving the comprehensiveness and scientificity of risk assessment, and avoiding misjudgment and missed judgment problems caused by relying on a single indicator in the prior art. (3) Dynamic current limiting and path guidance control are realized by means of the determination result of the risk level, enabling the response measures to have hierarchical linkage capabilities, effectively reducing the risk of thermal runaway in high-risk areas, and improving the response efficiency and safety guarantee level of the system. (4) By introducing a thermal risk prediction module based on a sliding window and an LSTM network, the prediction of future risk evolution trends is realized, so as to intervene in potential aggregation risks in advance, enhancing the forward-looking and proactive prevention and control capabilities of the system. Description of the Drawings

[0009] Figure 1 It is a schematic diagram of a new energy vehicle aggregation prevention and control system based on clustering recognition and regional risk modeling provided by the first embodiment of this application. Detailed Embodiments

[0010] In the following description, numerous specific details are set forth to provide a thorough understanding of the present application. However, the present application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the connotation of the present application. Therefore, the present application is not limited by the specific implementations disclosed below.

[0011] The first embodiment of the present application provides a new energy vehicle aggregation prevention and control system based on clustering recognition and regional risk modeling. Please refer to Figure 1 , which is a schematic diagram of the first embodiment of the present application. The following combines Figure 1 to elaborate in detail on a new energy vehicle aggregation prevention and control system based on clustering recognition and regional risk modeling provided by the first embodiment of the present application.

[0012] The new energy vehicle aggregation prevention and control system based on clustering recognition and regional risk modeling includes a status acquisition module 101, a clustering recognition module 102, a regional risk modeling module 103, a risk level determination module 104, a response control scheduling module, and a thermal risk prediction module 106.

[0013] The status acquisition module 101 is used to collect the operation status data of new energy vehicles in the target area. The operation status data includes the geographical location coordinates, battery temperature data, state of charge data, and vehicle operation status labels of the vehicles.

[0014] The status acquisition module 101 is used to continuously and real-time collect data of all new energy vehicles in the target area that are in operation, parked, or charging states, so as to provide complete and accurate operation status data for subsequent processing modules. To achieve this purpose, the status acquisition module 101 is deployed in designated monitoring areas of the city, such as underground parking lots, electric vehicle centralized charging areas, battery swapping stations, bus terminals, etc., which have the characteristics of high vehicle density aggregation. This module can integrate multiple acquisition means, including fixed cameras, parking space detectors, charging pile interfaces, battery T-Boxes, in-vehicle OBD terminals, LBS positioning devices, Bluetooth or ultra-wideband RTLS tag systems, etc.

[0015] In the specific data acquisition process, the status acquisition module 101 needs to complete information perception in the following several dimensions: First, collect the geographical location coordinate information of each new energy vehicle. The coordinate information can be obtained by using high-precision GPS, RTK, UWB, or base station fusion positioning methods, and the unit can be accurate to the meter level or sub-meter level, which is used to describe the static parking position or dynamic movement trajectory of the vehicle in two-dimensional or three-dimensional space. The acquisition frequency can be configured according to the scenario, and generally, it is recommended not to be less than 1 time per 5 seconds to ensure the continuity and real-time nature of the aggregation state judgment.

[0016] Secondly, obtain the temperature data of the battery system of each vehicle. The battery temperature data can be read in real time through the temperature sensors in the on-vehicle battery management system (BMS), including the cell temperature, the temperature of the battery pack housing, or the thermal runaway warning status. The data can be transmitted in degrees Celsius and uploaded in association with the vehicle unique identifier and location data. Abnormal temperature may indicate risks such as thermal runaway, local short circuit, and cooling failure, which are key indicators in the aggregation risk analysis.

[0017] In addition, the State of Charge (SOC) data is also an important part of the collection scope. SOC reflects the ratio of the current remaining battery charge to the fully charged state, usually expressed as 0% to 100%. The SOC value is collected to judge the current charging behavior of the vehicle and the degree of battery load. An overly high or low SOC state may increase the safety risk in the group aggregation, especially continuous charging when approaching the fully charged state is more likely to cause thermal anomalies.

[0018] Finally, it is also necessary to collect the vehicle operation status labels. The labels are used to identify the typical status of the vehicle currently, including at least five types: "charging", "not charging", "standby", "offline", and "abnormal status". Among them, the abnormal status can be further divided into thermal runaway alarm, BMS failure, communication interruption, etc. The label can be reported by the vehicle end or obtained by the platform through intelligent judgment in combination with temperature, SOC, and location changes. The operation status label not only reflects the status of a single vehicle but also serves as the basis for forming the weighted factor of the aggregation cluster status.

[0019] The status collection module 101 needs to encapsulate all the above data in a unified structured format, including fields such as vehicle unique identifier, timestamp, geographical location, battery temperature, SOC value, and status label, to form a standard data record. These data records are uploaded to the platform in real time through wireless communication (such as 4G, NB-IoT, WiFi, or LoRa) and stored in the dynamic status database or cache for simultaneous access and processing by the subsequent clustering recognition module 102 and thermal risk prediction module 106.

[0020] To improve the collection efficiency and data quality, the status collection module 101 can also be configured with data preprocessing functions, including position data denoising, temperature sampling averaging, status identification correction, outlier removal, etc. The module can also set up a time synchronization mechanism to ensure the alignment of data time sequences across devices and regions, supporting global aggregation cluster evolution analysis within a sliding time window.

[0021] In summary, the status collection module 101 not only provides the basic data support for the assessment of the aggregation status of new energy vehicles, but also provides executable input conditions for functional modules such as clustering recognition, risk modeling, and prediction response of the entire prevention and control system through real-time, high-dimensional, and structured data output, ensuring that the system has an accurate, continuous, and scalable data source.

[0022] The clustering recognition module 102 is used to identify aggregation clusters according to the operation status data and obtain the aggregation characteristics of the aggregation clusters by using a density-based spatial clustering algorithm.

[0023] The clustering recognition module 102 is used to identify groups of new energy vehicles that gather in a specific area and at a specific moment from the operation status data obtained by the status collection module 101, and extract the aggregation characteristics of each group. Its core goal is to identify a set of vehicles with spatial density and similar operation status from a large amount of spatial location data, that is, the so-called "aggregation clusters", so as to provide reliable input for subsequent risk modeling and response control.

[0024] In practical applications, this module first receives data such as the real-time position coordinates, battery temperature, state of charge (SOC), and operation status label of each new energy vehicle. These data may be uploaded in real time by in-vehicle systems, charging piles, ground positioning devices, or communication modules, and enter the clustering recognition module after being standardized. The module uses a density-based spatial clustering algorithm to analyze the vehicles. The basic idea of this algorithm is: if a vehicle has enough neighboring vehicles within a preset distance range, it can be regarded as a "core point", and an aggregation cluster is formed based on this.

[0025] For example, the system can set a fixed radius range (such as 100 meters) and require that at least 5 vehicles exist within this range to be considered as forming an aggregation. This method can automatically adapt to the shape of the vehicle distribution, without the need to artificially set the number of clusters, and can also effectively identify "isolated" vehicles, that is, those individuals that do not belong to any aggregation cluster, and these isolated vehicles will be marked as noise points.

[0026] After identifying multiple aggregation clusters, the clustering recognition module continues to extract a series of key features for each cluster to describe the structure and state of the aggregation cluster. First, calculate the number of vehicles within the aggregation cluster to reflect the scale of the cluster. Then, according to the position of each vehicle, determine the physical space area covered by the cluster, such as enclosing all vehicles in the form of a minimum bounding rectangle. Next, divide the number of vehicles by this area to obtain the unit density of the vehicles, thereby reflecting the spatial congestion degree of the aggregation cluster.

[0027] In addition, the clustering recognition module also performs statistical processing on the temperature data. For example, it calculates the average value of the battery temperatures of all vehicles within the aggregated cluster to judge the overall heat load level of the area. At the same time, the system also counts the proportion of vehicles whose temperature exceeds a preset threshold. The higher this proportion, the greater the likelihood of thermal anomalies in this area. Similarly, the proportion of vehicles with abnormal operating states is also counted, indicating that there may be risk factors such as operating failures or communication interruptions within the cluster.

[0028] All these features, including the number of vehicles, the occupied area, the unit density, the average temperature, the proportion of high-temperature vehicles, and the proportion of vehicles in abnormal states, will jointly constitute a complete description of the aggregated cluster. The clustering recognition module will output these feature data in a structured manner and transmit them to the subsequent regional risk modeling module 103.

[0029] To adapt to the dynamically changing urban traffic environment, the clustering recognition module usually runs automatically at fixed intervals, for example, refreshing the recognition results every 5 seconds or 10 seconds. The module can set a sliding time window to track the states of the same cluster at different time points and identify whether the aggregated cluster is expanding, shrinking, or moving. In addition, the module can also integrate a post-processing mechanism to merge or split the aggregated clusters near the boundary to prevent fragmentation of recognition due to position errors and improve the overall recognition stability and accuracy.

[0030] Through the above mechanism, the clustering recognition module 102 can accurately and timely identify the aggregation behavior of new energy vehicles in high-density urban areas and output aggregated cluster data with physical and thermal state characteristics, providing clear and definite data support for the risk modeling and control strategies of the system, ensuring that the system can achieve intelligent aggregation situation recognition capabilities for actual application scenarios.

[0031] Furthermore, the clustering recognition module is specifically used for: Construct a four-dimensional state space that combines vehicle geographical location coordinates, battery temperature data, state of charge data, and vehicle operating state labels. The four-dimensional state space is normalized through a state similarity function guided by state labels, enabling different-dimensional features to have a unified measurement basis in spatial distance calculation; In the four-dimensional state space, execute a dynamic density estimation mechanism. The dynamic density estimation mechanism dynamically adjusts the neighborhood search radius and minimum sample number of the clustering algorithm according to the density fluctuations of the proportion of high-temperature vehicles and abnormal states in the local space, making the clustering results more sensitive to potential thermal runaway areas; After the initial clustering is completed, introduce a structural stability determination process for each aggregated cluster, identify the boundary unstable areas based on the fluctuation degree of the temperature gradient at the cluster boundary, and perform boundary correction to supplement the edge vehicles missed in the clustering process due to critical states. Output the clustering characteristics of each obtained clustering cluster. The clustering characteristics include the number of vehicles, the spatial coverage, the vehicle density per unit area, the average temperature value of the clustering cluster, the proportion value of high-temperature vehicles, the proportion value of vehicles in abnormal states, and the boundary temperature fluctuation index of the clustering cluster, which are used to characterize the thermal risk sensitivity and evolution tendency of the clustering cluster.

[0032] In this embodiment, it is first necessary to make full use of the multi-source operation state data obtained by the state acquisition module. Specifically, the data collected by this module at least includes the geographical location coordinates, battery temperature data, state of charge data, and operation state labels of each new energy vehicle. These data are both time-sensitive and have differences, respectively reflecting the distribution of vehicles in the physical space, the thermal stability level of the electrochemical system, the remaining energy state, and the real-time operation state (such as: charging, not charging, abnormal, etc.). To achieve a more refined clustering cluster recognition mechanism that is more in line with the actual evolution process of risks, the clustering recognition module fuses these multi-dimensional data to construct a four-dimensional state space.

[0033] The establishment of this four-dimensional state space is not a simple splicing, but by introducing a state similarity function guided by state labels, the position, temperature, state of charge, and state labels are unified into the measurement framework. This similarity function first performs a normalization transformation on the data of different dimensions. For example, the geographical location coordinates are mapped to a unified unit according to the urban scale, the temperature range is compressed to the standard interval of 0 to 1, the state of charge is converted into the relative charging rate or the remaining capacity ratio, and the operation state labels are given weighted factors through numerical mapping, such as "abnormal state" is 1, "charging" is 0.75, "idle state" is 0.5, etc. In this way, in subsequent spatial distance or density estimation, each dimension feature can be comparable and have a unified measurement basis, thus avoiding the imbalance problem that one dimension of information dominates and other dimensions are ignored during clustering.

[0034] After constructing the four-dimensional state space, the clustering recognition module will enter the core density estimation and clustering stage. Different from the traditional DBSCAN or OPTICS algorithms where the neighborhood search radius and minimum number of points are fixedly set, this implementation plan adopts a dynamic density estimation mechanism to cope with the thermal risk sensitivity characteristics of new energy vehicle aggregations. Specifically, in each candidate aggregation area, the system first counts the proportion of high-temperature vehicles in this local space and the density of vehicles with abnormal operation status labels. If the high-temperature proportion in a certain area is significantly higher than that in the surrounding areas, or there are multiple abnormal vehicle aggregations, the system will automatically reduce the clustering neighborhood radius of this area and increase the clustering density threshold, making it easier to identify the core area of risk aggregation behavior. On the contrary, for areas with stable states and slow temperature changes, the system can appropriately relax the clustering conditions to improve the recognition coverage rate and fault tolerance. This mechanism ensures that the system has a higher recognition accuracy for thermally risky areas and avoids over-clustering of normal areas.

[0035] After completing the preliminary clustering, to further improve the accuracy of recognition and the stability of the aggregation cluster structure, the system will perform a structural stability determination for each identified aggregation cluster. This determination process focuses on the aggregation cluster boundary as the key analysis object and mainly examines the temperature distribution gradient of the vehicles at the cluster edge position. If there is a sharp rise or fall in temperature at the boundary area, or the temperature fluctuation amplitude is greater than the set threshold, the system will determine that there is a structural instability phenomenon at this boundary. At this time, the module will execute the boundary correction logic to re-incorporate some vehicles with critical states at the boundary into the aggregation cluster, thus compensating for the edge omissions caused by the clustering parameter settings and ensuring that the aggregation cluster completely represents the real physical aggregation situation.

[0036] Finally, the aggregation features will be extracted from each adjusted aggregation cluster. These features cover multiple dimensions, specifically including: the number of vehicles in the aggregation cluster, which is used to reflect the aggregation scale; the spatial range covered by the aggregation cluster, which can be calculated by the minimum bounding rectangle or convex hull method; the vehicle density per unit area, which reflects the vehicle congestion degree; the average temperature value, which represents the overall thermal environment level; the proportion of high-temperature vehicles, which is used to indicate potential local thermal anomalies; the proportion of vehicles with abnormal states, which reflects the prevalence of system anomalies; and the boundary temperature fluctuation index, which reflects the stability degree and evolution trend of the aggregation cluster structure. All these aggregation features will be uniformly structured and output, providing complete, rich, and quantifiable input data support for the subsequent regional risk modeling module, so that the entire new energy vehicle aggregation risk prevention and control system has the ability of dynamic and accurate recognition, real-time thermal risk assessment, and intelligent intervention control. The following is the reference implementation code of the clustering recognition module: import numpy as np import pandas as pd from sklearn.cluster import DBSCAN from scipy.spatial import ConvexHull, distance_matrix # Example data preparation (simulating vehicle status) # Simulated vehicle data, including geographical coordinates, temperature, SOC, status labels (numerical mapping) data = pd.DataFrame( [30.1234, 120.5678, 45.0, 0.8, 'charging'], [30.1235, 120.5679, 50.5, 0.9, 'abnormal'], [30.1240, 120.5685, 42.0, 0.7, 'idle'], [30.1300, 120.5700, 39.0, 0.6, 'idle'], [30.1232, 120.5672, 49.0, 0.85, 'abnormal'], [30.1236, 120.5676, 48.5, 0.8, 'charging'], [30.1233, 120.5673, 46.0, 0.75, 'idle'] , columns=['lat', 'lon', 'temp','soc','status']) # Status label mapping status_weight = {'abnormal': 1.0, 'charging': 0.75, 'idle': 0.5} data['status_val'] = data['status'].map(status_weight) # Status similarity normalization function def normalize(df, col, min_v=None, max_v=None): if min_v is None: min_v = df[col].min() if max_v is None: max_v = df[col].max() return (df[col]- min_v) / (max_v - min_v + 1e-6) data['lat_n'] = normalize(data, 'lat') data['lon_n'] = normalize(data, 'lon') data['temp_n'] = normalize(data, 'temp', 30, 60) data['soc_n'] = normalize(data,'soc', 0.0, 1.0) data['status_n'] = data['status_val'] # The status has been mapped, so it can be used directly # Construct a four-dimensional state space: latitude and longitude + temperature + status feature_vector = data[['lat_n', 'lon_n', 'temp_n','status_n']].values # Dynamic density estimation mechanism (adjust eps and min_samples) high_temp_threshold = 47 abnormal_density = (data['status'] == 'abnormal').sum() / len(data) hot_density = (data['temp'] > high_temp_threshold).sum() / len(data) # Automatically adjust clustering parameters based on heat risk sensitivity if abnormal_density + hot_density > 0.6: eps = 0.02 min_samples = 2 else: eps = 0.05 min_samples = 2 # Perform clustering (DBSCAN) clustering = DBSCAN(eps=eps, min_samples=min_samples).fit(feature_vector) data['cluster'] = clustering.labels_ # Post - processing of clustering: Structural stability analysis and boundary correction clusters = [] for cluster_id in sorted(data['cluster'].unique()): if cluster_id == -1: continue # Skip noise cluster_data = data[data['cluster'] == cluster_id].copy() coords = cluster_data[['lat', 'lon']].values if len(coords)<3: area = 0 boundary_std = 0 else: hull = ConvexHull(coords) area = hull.volume # Used as two - dimensional area boundary_idx = np.unique(hull.vertices) boundary_temps = cluster_data.iloc[boundary_idx]['temp'].values boundary_std = np.std(np.diff(sorted(boundary_temps))) # Temperature fluctuation gradient # If the boundary temperature fluctuates too much, try to merge edge vehicles if boundary_std>2.5: # Calculate the Euclidean distance to the boundary points and try to retrieve missing vehicles border_points = cluster_data.iloc[boundary_idx][['lat_n','lon_n']].values other_points = data[data['cluster']== -1][['lat_n', 'lon_n']].values dist = distance_matrix(other_points,border_points) close_points_idx = np.where(np.min(dist,axis=1)<eps)[0] new_indices = data[data['cluster']== -1].iloc[close_points_idx].index data.loc[new_indices, 'cluster']= cluster_id cluster_data = data[data['cluster']== cluster_id] # 更新 # 重新计算聚集簇特征 updated_coords = cluster_data[['lat', 'lon']].values if len(updated_coords)>= 3: area = ConvexHull(updated_coords).volume else: area = 0.0001 # 防止除零 vehicle_count = len(cluster_data) density = vehicle_count / area avg_temp = cluster_data['temp'].mean() high_temp_ratio = (cluster_data['temp']>high_temp_threshold).sum() / vehicle_count abnormal_ratio = (cluster_data['status']== 'abnormal').sum() / vehicle_count cluster_summary = { 'cluster_id': cluster_id, 'vehicle_count': vehicle_count, 'spatial_area': round(area, 5), 'vehicle_density': round(density, 2), 'avg_temp': round(avg_temp, 2), 'high_temp_ratio': round(high_temp_ratio, 2), 'abnormal_ratio': round(abnormal_ratio, 2), 'boundary_temp_fluctuation': round(boundary_std, 2) } clusters.append(cluster_summary) The regional risk modeling module 103 is used to calculate the risk score value of each aggregation cluster according to the aggregation cluster features provided by the clustering identification module; the risk score value is obtained according to the vehicle density per unit area, the average temperature value of the aggregation cluster, the high-temperature vehicle ratio value, the abnormal state vehicle ratio value, and the charging load ratio of the target area.

[0037] The regional risk modeling module 103 is used to conduct quantitative risk assessment on each aggregation cluster output by the clustering identification module 102. Its core function is to establish a computable risk scoring mechanism based on the statistical features of the aggregation cluster, so as to provide a basic basis for subsequent risk level division and response decision-making. The implementation of this module is based on the comprehensive modeling of multi-dimensional factors such as spatial distribution, vehicle status, and environmental load, ensuring that the output score not only reflects the scale and density of the current aggregation cluster but also can accurately capture its potential thermal safety hazards.

[0038] In actual operation, the regional risk modeling module first receives the basic statistical characteristics of the aggregation clusters. These characteristics include the number of vehicles within each aggregation cluster, the occupied space range, the average battery temperature, the proportion of vehicles with a temperature exceeding the safety threshold, the proportion of vehicles with abnormal operating states, and the current charging load ratio of the target area. Among them, the vehicle density per unit area is obtained by the ratio of the number of vehicles to the occupied area, reflecting the degree of spatial compression of the aggregation cluster. The average temperature of the aggregation cluster is obtained by performing a weighted average on the battery temperature data of all vehicles in the cluster, representing the overall thermal state of the aggregation cluster. The proportion of high-temperature vehicles refers to the proportion of the number of vehicles with a temperature exceeding the preset upper limit (such as 50°C) in the cluster, reflecting the potential triggering risk of thermal runaway. The proportion of vehicles in abnormal states refers to the proportion of the number of vehicles with status labels marked as abnormal (such as communication failures, thermal warnings, hardware failures, etc.). The charging load ratio of the target area is the ratio between the used charging power and the maximum bearable power in this area, used to characterize the energy load pressure of the environment.

[0039] Based on these input factors, the regional risk modeling module assigns a risk score value to each aggregation cluster. The scoring process can adopt a linear weighted model, where each indicator has a corresponding importance weight. These weights can be obtained through training with historical accident samples or set empirically based on domain knowledge. Before system deployment, aggregation cluster samples can be collected based on multiple typical scenarios, and record whether a thermal warning or safety event has occurred, which is used as a sample label, and a set of stable and effective weight combinations can be obtained through methods such as least squares fitting or logistic regression. The calculation result of each score value corresponds to a numerical interval, usually a real number, such as a value between 0 and 1 or 0 and 100. The higher the value, the higher the risk level.

[0040] To improve the interpretability and consistency of the scoring, the system normalizes various types of input data to ensure that the scoring output is not affected by absolute dimensions. For example, in the calculation of vehicle density, the normalization range can be determined based on the historical average and maximum values of the current area; in terms of temperature processing, a reference temperature (such as 45°C) can be set as the critical point for segment weight adjustment. The final scoring result will reflect the comprehensive risk degree of each aggregation cluster at the current moment, providing a direct input for the subsequent risk level determination module 104.

[0041] In addition, to adapt to different scenarios and dynamic environmental changes, the regional risk modeling module supports online updates of model parameters. For example, when the monitoring platform detects that some high-scoring aggregation clusters do not trigger any abnormal events, or low-scoring clusters have thermal runaway alarms, the system can automatically adjust the parameter weights to optimize the accuracy and sensitivity of the scoring mechanism.

[0042] The output results of the module are provided in a structured form, and the content includes at least information such as the cluster number, score value, scoring timestamp, main parameter values involved in the calculation, and risk explanation. These results are not only used by internal modules but also can be used for logging, trend analysis, or platform visualization to help operation and maintenance personnel intuitively understand the current regional thermal safety risk situation.

[0043] Through the above design and process, the regional risk modeling module 103 plays a crucial role of connecting the upper and lower levels in the system. It not only converts complex multi-dimensional vehicle aggregation information into quantifiable risk signals but also provides direct and clear input parameters for the scheduling and prediction module, ensuring that the entire prevention and control system has good response efficiency and accuracy.

[0044] Furthermore, the regional risk modeling module is specifically used for: Performing a difference calculation on the vehicle density per unit area and the average temperature value of each cluster within a continuous time period to obtain the density change rate and the temperature rise rate; Performing a sliding window statistics on the boundary temperature data of the cluster within a specified time period to extract the boundary temperature fluctuation amplitude, which is used to characterize the stability of the cluster structure; According to the magnitudes of the density change rate, the temperature rise rate, and the boundary temperature fluctuation amplitude, performing a ratio correction on the vehicle density per unit area, the average temperature value of the cluster, the proportion value of high-temperature vehicles, and the proportion value of abnormal state vehicles; Using the corrected vehicle density per unit area, the average temperature value of the cluster, the proportion value of high-temperature vehicles, the proportion value of abnormal state vehicles, and the charging load ratio of the target area together to calculate the risk score value of each cluster.

[0045] In the actual application process, it is often difficult to accurately capture the change trend of the risk situation only relying on the characteristics of the cluster at a certain moment. Therefore, in this embodiment, by introducing the evolution characteristics of the cluster, the scoring parameters are dynamically corrected to enhance the time-series responsiveness and the ability to identify the structural stability of the regional risk modeling, so as to achieve a more accurate risk assessment.

[0046] First, when calculating the risk score in each round, the system needs to extract the vehicle density per unit area and the average temperature value of each aggregation cluster within a continuous time period from historical data. The vehicle density per unit area is calculated by dividing the number of vehicles in the aggregation cluster by the spatial area covered by the aggregation cluster. The area can be calculated by the convex hull algorithm or the minimum bounding rectangle method. The average temperature value is the arithmetic mean of the battery temperatures of all vehicles in the aggregation cluster. The system performs a difference calculation on these values recorded at multiple consecutive time points, that is, the change amount between two adjacent time points is divided by the time interval, so as to obtain the density change rate and the temperature rise rate respectively. These two parameters respectively reflect the change trend of the aggregation intensity and the heat accumulation evolution speed during the aggregation process. For example, if the density per unit area of a certain aggregation cluster increases rapidly and the temperature rise rate increases significantly in a short period of time, it indicates that this aggregation cluster may be rapidly evolving into a high-risk area.

[0047] Secondly, the system further extracts the thermal stability index of the boundary region of the aggregation cluster. To this end, sliding window statistical processing is performed on the vehicle temperature data at the boundary of each aggregation cluster. The boundary can be determined by sorting the distances of vehicles from the geometric center in the aggregation cluster and selecting a certain proportion of samples close to the boundary. A time window with a fixed length (such as 5 minutes or 10 minutes) is set, and the range, variance or maximum gradient of the boundary vehicle temperatures are calculated within this window to quantify the amplitude of their temperature fluctuations. This amplitude of fluctuations reflects whether there is a significant thermal instability phenomenon at the boundary. If the fluctuations are significant, there may be a risk that the aggregation structure is about to break or expand. Therefore, the system should increase the degree of attention to this aggregation cluster.

[0048] Combining the above two sets of time evolution parameters, the system uses them to perform proportional correction on the original scoring indicators. Specifically, the original scoring indicators include the vehicle density per unit area, the average temperature value of the aggregation cluster, the proportion value of high-temperature vehicles, and the proportion value of abnormal-state vehicles. If the density change rate of a certain aggregation cluster is relatively high or the temperature rise rate is extremely rapid, the scoring weights corresponding to the vehicle density per unit area and the average temperature value can be appropriately increased, so as to enhance the system's sensitivity to rapidly aggregating hotspots. Similarly, if the boundary temperature fluctuation index exceeds the stability threshold, the weights of the proportion value of high-temperature vehicles and the proportion value of abnormal-state vehicles can be correspondingly increased to reflect the potential structural risks in this area.

[0049] Finally, the system inputs the corrected scoring metrics and the charging load ratio of the target area into the scoring calculation logic together to comprehensively evaluate the risk level of each clustering cluster. The scoring logic can be linearly aggregated by setting the weighting coefficients of each scoring factor, or a multi-level segmented determination method can be adopted to ensure that the scoring results can reflect the dynamic evolution characteristics of heat, density, and structural risks at the same time. The finally output risk score value can not only accurately reflect the static risk state of the current clustering cluster, but also has a certain degree of forward-looking and trend judgment ability, providing a more reliable input data basis for the subsequent risk level determination module and response control scheduling module. Through the method described in this embodiment, high-sensitivity dynamic modeling and intelligent quantitative evaluation of new energy vehicle aggregation risks can be achieved, significantly improving the thermal safety management efficiency and precise intervention ability of new energy aggregation areas.

[0050] The following is an example to make it easier for those skilled in the art to understand. For example, during the actual operation of the new energy vehicle aggregation prevention and control system, a certain city monitoring system identified a new energy vehicle clustering cluster with typical aggregation characteristics from 10:00 to 10:10 am. The real-time monitoring data of this clustering cluster shows that at 10:00, there were a total of 50 vehicles in this area, the vehicle distribution range was 2000 square meters, and the average battery temperature was 45.0 °C. By 10:05, the number of vehicles increased to 60, the distribution area shrank to 1800 square meters, and the average temperature increased to 46.5 °C. By 10:10, the number of vehicles further increased to 70, the coverage area shrank to 1650 square meters, and the average battery temperature rose to 48.0 °C.

[0051] Based on the above time series data, the system first calculates the vehicle density per unit area of this clustering cluster in three consecutive time slices, which are 0.025 vehicles per square meter, 0.0333 vehicles per square meter, and 0.0424 vehicles per square meter respectively. Further calculate the density change rate between two adjacent time slices, which are 0.00166 vehicles per square meter per minute and 0.00182 vehicles per square meter per minute respectively, and take the average to obtain a density change rate of approximately 0.00174 vehicles per square meter per minute. At the same time, the change rate of temperature is also calculated within the same time period, and both changes are 0.3 °C per minute, indicating that there is a continuous upward trend in temperature inside this clustering cluster.

[0052] Subsequently, the system focuses on the boundary region of the aggregation cluster and conducts temperature fluctuation analysis on the boundary vehicles in the outermost 10%. Within the set sliding window range, the range of temperature changes collected for this part of the boundary vehicles are 2.5°C, 3.0°C, 2.8°C, 3.1°C, 2.9°C, 3.2°C, and 3.0°C respectively. It can be calculated that the average fluctuation amplitude of the boundary temperature is 2.93°C. Considering that the stability determination threshold set by the system is 2.5°C, and this value significantly exceeds the stable range, the system determines that there is strong instability in the current aggregation cluster boundary based on this, and triggers the correction process for the scoring parameters accordingly.

[0053] Dynamically correct the original aggregation cluster eigenvalue according to the aggregation evolution characteristics. The vehicle density per unit area is increased by 10% due to the rapid growth of density, that is, it is corrected from 0.0424 vehicles per square meter to 0.0466 vehicles per square meter; the average temperature is increased by 5% due to the obvious temperature rise trend, from 48.0°C to 50.4°C; the proportion of high-temperature vehicles was originally 30% (that is, 21 out of 70 vehicles have a battery temperature exceeding 50°C), and it is increased by 15% due to the influence of boundary fluctuations, and is corrected to 34.5%; the proportion of vehicles in abnormal state was originally 10%, and after increasing by 15%, it is 11.5%. In addition, the charging load ratio of the target area obtained by the system is 0.80.

[0054] Finally, the system comprehensively evaluates the above corrected parameters and calculates the risk score using the set linear weighting method. The weights of each feature are 0.2 for the vehicle density per unit area, 0.25 for the average temperature value, 0.2 for the proportion of high-temperature vehicles, 0.15 for the proportion of abnormal state, and 0.2 for the charging load ratio. The calculation results are as follows: the item of vehicle density per unit area is 0.2×0.0466 = 0.00932, the item of average temperature is 0.25×50.4 = 12.6, the item of high-temperature proportion is 0.2×0.345 = 0.069, the item of proportion of abnormal state is 0.15×0.115 = 0.01725, and the item of charging load ratio is 0.2×0.8 = 0.16. Adding the above values together, the final risk score value of this aggregation cluster is approximately 12.86.

[0055] This score value is higher than the high-risk threshold set by the system (for example, 10.0). Therefore, this aggregation cluster will be automatically recognized as a high-risk area by the system, and corresponding intervention measures will be triggered, such as warning notifications, enhanced thermal runaway monitoring, and charging power adjustment strategies.

[0056] Furthermore, the regional risk modeling module is also used for: Based on the spatial boundary form of each aggregation cluster, identify the position sequence of boundary vehicles within a continuous time period and construct a boundary trajectory set, and the boundary trajectory set is used to retain the geographical distribution path of boundary vehicles changing with time; Using the set of boundary trajectories, the battery temperature of each boundary vehicle within the corresponding time slice is extracted to construct a set of boundary temperature sequences with time indices, which is used to record the dynamic evolution process of the thermal state at the boundary of the aggregation cluster; Based on the set of boundary temperature sequences, the temperature variation amplitude, temperature rise rate, and length of the continuous temperature difference section within each window are extracted in a time-sliding window manner to form a sequence of local temperature perturbation feature vectors; Perform boundary directional aggregation on the sequence of local temperature perturbation feature vectors. According to the arrangement order of each boundary segment in the spatial trajectory, construct an overall boundary temperature perturbation characteristic curve, and extract the boundary temperature fluctuation amplitude based on its fluctuation amplitude as the output, which is used to characterize the thermal stability risk of the aggregation cluster structure.

[0057] In the new energy vehicle aggregation prevention and control system, in order to more accurately evaluate the thermal stability risk of the aggregation cluster structure, the regional risk modeling module needs to further extract and analyze the dynamic evolution process of the thermal state at the boundary of the aggregation cluster.

[0058] First, to construct a representation of the thermal feature evolution at the boundary of the aggregation cluster, it is necessary to identify the vehicle individuals on this boundary based on the spatial boundary shape of each aggregation cluster within a specific time period. The identification of boundary vehicles is not completed by static spatial distance judgment, but based on the outer contour topological structure of the aggregation cluster after clustering. By calculating the outer contour envelope boundary of the aggregation cluster and matching the extreme points within the boundary neighborhood, the vehicle trajectory points that meet the boundary determination rules are selected as the set of boundary candidate points. On this basis, considering the factor of time continuity, for each aggregation cluster, the boundary identification process is executed on multiple consecutive time slices respectively to form a set of boundary vehicle trajectory points under time indices. To retain the evolution path of the boundary over time, the trajectories should be further connected according to the vehicle identification, and the boundary positions of the same vehicle in different time slices are formed into a trajectory sequence to construct a boundary trajectory set.

[0059] After obtaining the set of boundary trajectories, it is necessary to extract the battery temperature data of these boundary vehicles at the corresponding time slices. Since the positions of boundary vehicles have the attribute of spatial outer edge, their temperature changes are more sensitive to environmental changes and vehicle aggregation effects. Therefore, the temperature data extracted from the set of boundary trajectories can more effectively reflect the thermal state of the overall edge of the aggregation cluster. The specific operation is as follows: According to the time stamp index order, for each time point in each boundary trajectory, obtain the battery temperature value of the corresponding vehicle at that moment, and accumulate them item by item to form a time-temperature mapping sequence with the vehicle trajectory as the unit. To support the subsequent extraction of temperature perturbation features, these time-temperature sequences need to be uniformly stored in the set of boundary temperature sequences with time indices, so that the boundary thermal state of each aggregation cluster within each time slice has a clear structural expression and time relationship.

[0060] After the set of boundary temperature sequences is constructed, the operation of extracting dynamic disturbance features needs to be carried out. The temperature sequence set is locally sliced in the time dimension by means of a sliding window. Three main disturbance feature indicators are sequentially extracted within each window: First, the temperature variation amplitude, which is the difference between the maximum temperature and the minimum temperature within the window, is used to characterize the fluctuation intensity of the boundary temperature within the window; Second, the temperature rising rate, which is defined as the average growth amount of the backend value of the temperature sequence relative to the frontend value within the window, is used to identify the trend of boundary heat accumulation; Third, the length of the continuous temperature difference section, which is the maximum time period length within the window where the continuous temperature difference change exceeds the preset fluctuation threshold, is used to capture the phenomenon of continuous temperature rise. These three disturbance features together constitute the local temperature disturbance feature vector, and the disturbance feature vectors of each sliding window form a sequence of local temperature disturbance feature vectors in chronological order, laying the foundation for the generation of the subsequent overall boundary disturbance pattern.

[0061] After the local disturbance feature extraction is completed, it is necessary to aggregate the local disturbance feature vectors on all boundary trajectories to construct the overall boundary temperature disturbance characteristic curve. The construction of this curve not only involves the time series splicing of the feature vectors, but also needs to introduce a spatial directionality sorting mechanism to ensure that the disturbance features of each section of the boundary are combined according to the spatial arrangement order of the vehicle on the boundary. The spatial directionality sorting is based on the geometric continuity of the boundary trajectory. By calculating the included angle of the direction vectors between adjacent trajectory segments, the natural connection direction of the spatial path is determined, and then an ordered disturbance sequence from the starting point to the ending point of the boundary is constructed. The boundary temperature disturbance characteristic curve generated in this way has directionality and continuity, and can truly reflect the temperature fluctuation trend and change trend of the aggregation cluster boundary in space.

[0062] Finally, it is necessary to extract the key indicator representing the thermal stability risk of the aggregation cluster structure based on the constructed boundary temperature disturbance characteristic curve, that is, the boundary temperature fluctuation amplitude. The extraction method of this indicator is to identify the maximum temperature difference between the local peaks and valleys in the complete boundary temperature disturbance characteristic curve, and combine the spatial span information to filter out abnormal isolated fluctuation points, extract the representative continuous fluctuation amplitude section, and then output the temperature difference of this section as the final boundary temperature fluctuation amplitude. This indicator can effectively identify the thermal disturbance caused by the instability of the aggregation cluster boundary or the environmental coupling effect, and can be used as an important basis for the output of the regional risk modeling module to support the subsequent determination of the risk level, calibration of the warning threshold, and adjustment of the response scheduling strategy.

[0063] The risk level determination module 104 is used to determine the risk level of the corresponding aggregation cluster according to the risk score value.

[0064] The risk level determination module 104 is used to receive the risk score values of each clustering cluster output by the regional risk modeling module 103, and based on a set of preset multi-level risk level classification criteria, map these score values to specific risk level labels, so as to provide clear and definite classification instructions for the subsequent response control scheduling module. The core function of this module is to convert the continuous score results into executable classification signals, and ensure that the boundary settings between different levels are scientific, stable and adaptable.

[0065] In specific implementation, the risk level determination module first defines several risk level intervals, which are commonly three to four-level structures, such as "low risk", "medium risk", "high risk" or extended to levels such as "normal", "warning", "high alert", "emergency" etc. in the actual deployment scenario. Each level corresponds to a range of risk score values, and the higher the score value, the greater the likelihood of potential thermal runaway or safety incidents. To set these intervals, the module will conduct statistical modeling and threshold optimization based on a large amount of historical data, and use methods such as ROC curve (Receiver Operating Characteristic Curve) analysis or maximum F1 value method to find the optimal discrimination point, ensuring the maximum classification accuracy of the score value division between different levels.

[0066] When the risk level determination module is running, it compares the risk score value of each clustering cluster with the set threshold. If the score value of a clustering cluster is lower than the first threshold, it is classified as the low risk level; if the score value is between two intermediate thresholds, it is classified as the medium risk level; if the score value exceeds the highest threshold, it is determined as the high risk level. After each level determination is completed, a structured data record will be generated, including the clustering cluster number, the current score value, the corresponding risk level, the determination timestamp and the specific index items triggering the level change. This record will be synchronously transmitted to the response control scheduling module for triggering the decision execution of current limiting, guiding or linkage strategies.

[0067] Considering that the actual operating environment may change dynamically, the risk level determination module also supports an adaptive threshold adjustment mechanism. For example, in a specific period or specific area, if it is found that the overall system score value is generally high but no actual accident has occurred, the high risk threshold can be automatically increased to avoid misjudgment; on the contrary, in the area where an accident has occurred or under the condition of detecting the deterioration of the external thermal environment, the threshold can be temporarily lowered to improve sensitivity. In addition, this module can also access the output results of the thermal risk prediction module 106 to adjust the current level determination boundary in advance, making the risk level judgment more in line with the future situation.

[0068] To ensure the consistency and stability of system responses, the risk level determination module uses a sliding time window mechanism to update and judge levels, avoiding frequent triggering of level changes due to short-term fluctuations in scoring values. For example, if the scoring values of a certain aggregation cluster exceed the high-risk threshold in two consecutive periods, the system will determine it as a continuously high-risk cluster and trigger a high-priority response. At the same time, to prevent delayed recognition, the module also sets up a fast response channel. Once the scoring value instantaneously exceeds the limit threshold by a large margin, it can be immediately marked as the emergency level, and the platform is notified to execute emergency strategies such as power-off and dispersion.

[0069] In summary, the risk level determination module 104 not only serves as a bridge from continuous risk scoring values to hierarchical response control signals, but also ensures the accuracy, stability, and responsiveness of risk assessment results through a multi-level threshold system, an adaptive optimization mechanism, and time consistency constraints, thereby providing a reliable guarantee for risk management in the new energy vehicle aggregation scenario.

[0070] Furthermore, the risk level determination module is specifically used for: Obtain the risk scoring values output by the regional risk modeling module for the same aggregation cluster in multiple consecutive time slices, and construct a scoring time series in chronological order; Based on the scoring time series, calculate the risk scoring increment values and their average change rates between time slices, which are used to characterize the short-term scoring growth trend of this aggregation cluster; According to the short-term scoring growth trend, combined with the current scoring value and static aggregation cluster characteristics including vehicle density per unit area and average temperature value of the aggregation cluster, predict the range of risk scoring changes within a preset future time window, and output the future scoring prediction value interval; Conduct an interval overlap analysis between the future scoring prediction value interval and the static risk level threshold to identify whether there is an early warning signal of a trend crossing the high-risk threshold; In the case where the early warning signal is established, combine the current scoring value and the upper limit of the future scoring prediction value interval to construct a trend strengthening factor and correct the current scoring value, and then output the final scoring result; Compare the final scoring result with the risk level determination rule to obtain the current risk level corresponding to this aggregation cluster.

[0071] In this embodiment, the risk level determination module undertakes the key task of converting the risk score value into a specifically recognizable risk level, and by introducing a time series trend analysis mechanism, enables the risk determination to have the characteristics of dynamic evolution. This module first needs to obtain the risk score values output by the regional risk modeling module for the same aggregation cluster within multiple consecutive time slices. Generally speaking, this score value is calculated based on unified clustering recognition and updated at fixed time intervals. Therefore, by collecting the score results of several recent time slices, a complete score time series can be constructed, which is arranged in chronological order and reflects the risk evolution trajectory of the aggregation cluster in the near future.

[0072] Subsequently, the system performs a short-term growth trend analysis on this score time series. Specifically, the score difference between adjacent time slices is calculated to obtain the risk score increment value at each moment. On this basis, the average value of all score increment values is further calculated to obtain the average risk score growth rate of the current aggregation cluster in the most recent period. This growth rate can effectively quantify the change speed of the risk score and thus characterize the risk rising trend of the aggregation cluster, providing support for subsequent prediction.

[0073] Next, combining the score value of the current time slice and several static characteristic indicators of this aggregation cluster, such as parameters like vehicle density per unit area and average temperature value of the aggregation cluster, an input feature vector for future score prediction is constructed. Based on the collaborative analysis results of the score growth trend and static characteristics, the system uses a regression algorithm or an interval growth model to predict the possible change range of the risk score within a preset future time window and outputs it as an interval of future score prediction values. This interval can be represented in the form of upper and lower limits to reflect the fluctuation boundaries of the score within this time window.

[0074] To determine whether the evolution of the future score will trigger a change in the risk level, the system introduces an interval overlap analysis mechanism. Here, the static risk level threshold has been preset as a fixed segmented interval. For example, different score ranges correspond to low risk, medium risk, and high risk. When the future score prediction interval partially or completely overlaps with the high-risk threshold interval, the system determines this as an early warning signal of a potential trend to cross the high-risk threshold. This signal is used to pre-sense the sudden risk growth that the aggregation cluster may face.

[0075] If the early warning signal holds, the system will further construct a trend strengthening factor using the current score value and the upper limit value of the future score prediction interval. This factor reflects the potential jump trend of the risk score through a certain functional relationship and performs a correction operation on the current score value. The correction method can adopt linear amplification, exponential weighting, or threshold-driven methods to adjust the current score to a final score result that can better reflect the trend warning effect, ensuring that the system has the ability to respond in a timely manner when perceiving the evolution trend.

[0076] Finally, the system inputs the above final scoring results into the risk level determination rules for level comparison. The determination rules are generally a multi-segment scoring level division mechanism. By the scoring value falling into different intervals, the risk level of the current aggregation cluster is determined. For example, if it falls within a certain high-risk scoring threshold range, it is classified as a high-risk level. This final determination result constitutes the output of the risk level determination module and provides a direct instruction basis for the risk response control of the downstream system of the system.

[0077] Through the above method, the risk level determination module can not only make static determination based on the current scoring value, but also integrate the scoring change trend and future prediction results to construct a risk level judgment mechanism facing evolutionary dynamics, enabling the system to have foresight and high sensitivity in dealing with the aggregation heat risk of new energy vehicles. This mechanism effectively improves the system's ability to identify and respond to sudden risks, breaking through the technical path of making level determination only based on the scoring value at a single moment in the traditional way.

[0078] In a central area of a certain city, the new energy vehicle monitoring system identifies a high-density aggregation cluster numbered C1 through the clustering recognition module. The recognition time of this aggregation cluster is 15:00 on June 1, 2025, and it contains 47 vehicles. The preliminary aggregation characteristics include a vehicle density of 16.5 vehicles per thousand square meters in the unit area, an average temperature of the aggregation cluster of 50.2 °C, a high-temperature vehicle ratio of 19.1%, an abnormal state vehicle ratio of 12.8%, and a charging load ratio of 0.72. The regional risk modeling module successively outputs the risk scoring values of C1 within six consecutive time slices (each slice is 5 minutes) between 15:00 and 15:25: 0.42, 0.46, 0.51, 0.57, 0.64, 0.72.

[0079] The risk level determination module first constructs a scoring time series S = [0.42, 0.46, 0.51, 0.57, 0.64, 0.72], and arranges it in chronological order. The system then performs a difference processing on adjacent scoring values to obtain a scoring increment sequence ΔS = [0.04, 0.05, 0.06, 0.07, 0.08], and calculates the average value of this sequence to obtain an average risk scoring growth rate of 0.06. This value is used as a quantitative index to reflect the growth trend of the risk scoring of the current aggregation cluster in a short time period.

[0080] Next, the system obtains the static aggregation cluster features at the current time point (15:25), including parameters such as vehicle density per unit area of 16.5, average temperature of the aggregation cluster of 50.2 °C, proportion of high-temperature vehicles of 19.1%, and proportion of vehicles in abnormal states of 12.8%. Through the set scoring growth trend prediction model (a linear or non-linear regression function that can be trained based on historical regression samples), the scoring growth rate of 0.06 and the above static features are jointly input into the model to predict the possible risk scoring interval within the next time window (set to the next 15 minutes, i.e., from 15:30 to 15:45). For example, according to the output of this model, the predicted scoring value of C1 within the future window is in the interval [0.74, 0.86].

[0081] After obtaining the scoring prediction interval, the system compares it with the statically set risk level threshold. The high-risk level threshold is set to 0.80. The system conducts an interval overlap analysis of the prediction interval [0.74, 0.86] and the threshold interval [0.80, 1.00], and finds that there is a significant intersection between the two, and the prediction upper limit has exceeded the high-risk threshold. Therefore, the system records that this aggregation cluster has a warning signal of trending through the risk level.

[0082] On the premise of confirming the warning signal, the system constructs a trend strengthening factor. Set the weight parameter of the trend strengthening factor as α = 0.75, and use the current scoring value Scurrent = 0.72 and the upper limit of the prediction interval Spred = 0.86 to calculate the corrected scoring value Sfinal = α × Scurrent + (1 – α) × Spred = 0.75 × 0.72 + 0.25 × 0.86 = 0.755 by weighted calculation. This value is used as the final scoring result that fuses the current state and future predictions.

[0083] Finally, the system compares Sfinal = 0.755 with the risk level rules. The risk level rules are set as follows: a score less than 0.6 is a low risk, 0.6 to 0.8 is a medium risk, and 0.8 and above is a high risk. Since 0.755 falls at the upper limit of the medium-risk interval and there is a predicted trend of evolving towards a high risk, the system finally determines C1 to be in a medium-high risk critical state and classifies it as a high-risk level in advance through the trend strengthening mechanism. This level will be output to the intervention decision-making module for triggering the regional charging limit strategy and scheduling optimization.

[0084] Furthermore, the risk level determination module is also used for: Decompose the future scoring prediction value interval into three indicators: the lower limit value, the upper limit value, and the interval span, and calculate the distance between the interval span and the current scoring value based on the short-term scoring growth trend of the scoring time series, which is used to characterize the possible growth fluctuation range of the future score; According to the fluctuation range, all level cutoff values greater than the current score value in the static risk level threshold are retrieved, a set of cutoff values with overlapping values is screened out, and the score difference interval corresponding to the set of cutoff values is output; Compare the score difference interval with the fluctuation range, calculate the coverage ratio of each boundary value within the fluctuation range, and use the ratio as a trend crossing strength indicator to determine whether the future score interval has a tendency to migrate to a high risk level; When the trend crossing strength index reaches a preset judgment threshold, an early warning signal is generated.

[0085] In the new energy vehicle gathering prevention and control system, the risk level determination module performs a series of processing steps to determine whether to generate an early warning signal based on the analysis of the future score prediction value interval. First, after obtaining the future score prediction value interval, the module does not directly use the overall information of the interval for judgment, but instead disassembles the interval and decomposes it into three indicators: lower limit value, upper limit value and interval span. The lower limit value and upper limit value represent the possible minimum and maximum values of the future estimated score value, respectively, and the interval span is the difference between the upper and lower limits. This disassembly process ensures that subsequent processing can identify risk trends based on a more detailed structure.

[0086] Next, the module obtains the short-term score growth trend of the score time series corresponding to the score interval. This trend extracts the score growth range and speed information over a continuous time slice by calculating the change in the score value between adjacent time slices. On the basis of obtaining the short-term score growth trend, the module further calculates the numerical distance between the interval span and the current score value to form a measurement indicator of the fluctuation range. This numerical distance is used to quantify the extent to which the score prediction value interval may move up in the current state, reflecting the potential range of score growth.

[0087] After the calculation of the above fluctuation range is completed, the module enters the matching process of the static risk level threshold. The system presets multiple static risk level thresholds as the boundaries of risk level division. The module traverses all risk level cutoff values higher than the current score value, compares the values of each cutoff value, and determines whether the cutoff value overlaps with the future score prediction value interval. The judgment is based on whether the upper and lower limits of the future score prediction interval cover the cutoff value. If there is an overlap, it means that the risk level threshold may be crossed by the score value in the future.

[0088] For all overlapping demarcation values, the module continues to perform the comparison between the score difference interval and the fluctuation range. Specifically, the score difference interval refers to the difference range between the demarcation value and the current score value, and the fluctuation range is the difference between the upper limit of the score interval and the current score value. The module performs a ratio operation on the score difference interval and the fluctuation range to obtain the proportion of the demarcation value covered in the fluctuation range, which is the trend crossing strength indicator. The calculation of this ratio must not only ensure the consistency of the upper and lower limits, but also unify the data source and time synchronization of the score value to avoid inconsistencies caused by differences in scoring time.

[0089] After completing the calculation of the trend crossing strength index of all demarcation values, the module compares the index result of each demarcation value with the judgment threshold preset by the system to determine whether the triggering condition of the early warning is met. If the trend crossing strength index of any demarcation value exceeds the threshold, it means that the score value has a trend of crossing the level threshold, and the module generates an early warning signal accordingly. This signal is then used as the status identifier of the current cluster for use by subsequent scheduling and guidance modules.

[0090] The response control scheduling module 105 is used to execute flow control and path guidance measures according to the risk level and provide response results to the external platform.

[0091] The role of the response control scheduling module 105 is to execute the corresponding intervention control strategy according to the risk level of the cluster output by the risk level determination module 104, so as to reduce or avoid the safety risks such as thermal runaway, fire spread or regional load overload caused by the gathering of new energy vehicles. This module not only contains the strategy formulation logic, but also undertakes the linkage and information exchange tasks between the terminal execution equipment and the external city management platform, and is a key component for realizing system closed-loop control and practical deployment.

[0092] When the risk level determination module outputs that a certain cluster is at a medium risk or high risk level, the response control scheduling module first starts the current limiting control mechanism. Specifically, the system calculates the acceptable maximum vehicle load threshold of the area based on the spatial boundaries of the area where the cluster is located and the current number of vehicles. If the actual number of vehicles exceeds the threshold, the module will issue an access restriction instruction to the charging platform or parking management system to prohibit new vehicles from entering the specified range of the area. This current limiting instruction can directly act on the charging pile, gate system or parking lot management system through an API interface or communication protocol, and prioritizes limiting the access of vehicles with high temperature, high SOC or long-term residence. At the same time, the current limiting strategy has the ability to adjust dynamically. The system will periodically update the access upper limit value according to changes in the cluster status to avoid waste of resources caused by long-term rigid thresholds.

[0093] Under the high-risk level determination, the response control and scheduling module will also synchronously activate the path guidance strategy. The core purpose of this strategy is to timely divert the vehicles that are already in the risk area or approaching the area to a relatively safe area. The system will automatically generate the optimal diversion path according to factors such as the road structure, the number of available parking spaces, the temperature distribution in the area, and the traffic flow density in the current area. The guidance information will be sent to the vehicle owner or the vehicle control system through multiple terminals such as the in-vehicle T-Box module, the mobile application (APP), the navigation terminal, and the charging pile display screen, prompting them to immediately leave the current area and indicating them to go to the recommended alternative area to continue parking or charging. For vehicles in a non-networked state, the system can also link with on-site induction screens or broadcast systems for area announcements.

[0094] The response control and scheduling module not only includes the functions of flow limiting and guidance, but also supports multi-level linkage with other subsystems. After a high-level risk determination occurs, the module can actively send linkage signals to the fire protection system to pre-start the ventilation device or sprinkler device to reduce local heat accumulation; it can also control the lighting system to turn on the high-brightness flashing mode to prompt people's attention; at the same time, it can also call the cameras of the area monitoring system to perform AI video analysis on the core area of the aggregation cluster to identify whether there are visual risk signals such as smoke, abnormal behavior, or hardware damage, and feedback to the monitoring background.

[0095] In addition, the response control and scheduling module will transmit the execution results back to the system platform in real time and synchronously upload them to external interfaces such as the urban management platform, the energy scheduling platform, or the emergency command center. The uploaded content includes whether the flow limiting control is effective, the execution situation of the vehicle diversion path, the response status of the linked devices, and the change trend of the aggregation cluster, providing a complete response log and basis for subsequent disposal for the supervision department. All execution actions and status changes are recorded in the system in the form of an event chain and can be exported as operation logs, charts, or alarm reports to meet the traceability and visualization requirements of urban-level risk management.

[0096] To ensure high reliability, the response control and scheduling module supports a multi-threaded scheduling mechanism and priority queue management during deployment. It can automatically schedule resources, allocate channels in the case of multiple clusters existing simultaneously or sudden emergency aggregations, and prioritize the control requirements of the most serious areas according to the risk level. In addition, to avoid false triggering and resource waste, the module also sets a delay confirmation mechanism and the shortest execution interval to ensure that the control behavior has sufficient response strength after being triggered by actual needs without being executed too frequently.

[0097] Through the above mechanism, the response control and scheduling module 105 realizes the closed-loop logic from risk identification to on-site control. It can not only respond in real time based on the current risk level, but also cooperate with the prediction module to deploy early warning measures in advance, enabling the entire system to have strong dynamic intervention and linkage control capabilities, thereby effectively ensuring the operation safety and urban management efficiency in the scenario of intensive new energy vehicles.

[0098] The thermal risk prediction module 106 is used to construct a time series feature dataset based on a sliding window according to the operation status data, and use a pre-trained long short-term memory neural network model to predict the potential risk score trend of each target area within a preset future time range; and output the predicted potential risk score trend to the risk level determination module and the response control scheduling module to execute the current limiting or guiding strategy in advance.

[0099] The design purpose of the thermal risk prediction module 106 is to further achieve a forward-looking judgment of the future risk trend of the new energy vehicle aggregation area on the basis of the system completing the current state monitoring and risk level determination, so as to provide a scientific basis for the response control scheduling module to deploy the current limiting or guiding strategy in advance. Based on the time series modeling method and combined with artificial intelligence algorithms, especially the long short-term memory neural network structure (LSTM), this module can effectively capture the dynamic characteristics of the regional aggregation situation evolving over time, and is suitable for dealing with the nonlinearity, hysteresis and short-term mutations existing in the new energy vehicle aggregation process.

[0100] In the specific implementation process, the thermal risk prediction module first continuously receives the operation status data from the status acquisition module 101 and organizes it with time as the main axis through a sliding window. The so-called sliding window means that the system selects a time series with a fixed length as the input sample for the current prediction at each time step. For example, if the step length is set to 5 minutes, the data of the past 30 minutes is taken to form a window, and the window advances forward by 5 minutes each time to form a continuous time series dataset. The feature items included in each window include but are not limited to the total number of vehicles in the target area, the vehicle density per unit area, the average battery temperature, the proportion of high-temperature vehicles, the proportion of vehicles in abnormal states, the charging load ratio of the current area, the temperature rise rate, etc. These feature items are combined into an input vector in a structured manner and arranged in chronological order to form training samples, which are input into the trained prediction model.

[0101] This module uses LSTM as the main prediction structure. LSTM belongs to a type of recurrent neural network and can effectively retain long-term dependence information in time series data, avoiding the problems of gradient disappearance or explosion that may occur in traditional neural networks when processing sequences. The internal structure of LSTM includes an input gate, a forget gate and an output gate. Through these gating mechanisms, it dynamically selects which historical information to retain and which to discard, so as to enhance the sensitivity to new state changes while maintaining the memory ability of the model. The system establishes a corresponding LSTM model instance for each target area and completes offline training according to the historical operation data of the deployment area. In the training stage, the system will use the real aggregation state of the known time period and the subsequent risk score values as training labels, and optimize the model parameters through backpropagation to finally form the model weights that can independently predict the future risk trend.

[0102] After the deployment is completed, the heat risk prediction module can call this model in real time during the system operation. In each prediction cycle, the module extracts time series features from the current sliding window, inputs them into the LSTM model, and outputs the prediction results within one or more future time steps. The prediction results are usually presented as continuous risk score values, indicating the risk score trend of the target area 5 minutes, 10 minutes, or 15 minutes later. The prediction results can not only be used for trend display, but also be linked with the risk level determination module. The predicted score value is compared with the determination threshold. If the future trend shows that a high risk is approaching, the system can mark this cluster as "predicted high risk" in advance, thereby triggering the early intervention of the response control and scheduling module to execute measures such as flow limiting, route guidance, or warning release.

[0103] To improve the stability and reliability of the prediction results, the heat risk prediction module can also perform weighted correction by combining the measured score value at the current time point, that is, using the combined value of the current score and the predicted score as the final reference, and setting a confidence interval to determine whether the prediction result significantly deviates from the existing situation. During the continuous operation of the model, the system can also periodically update the model weights through an online learning mechanism to make it adapt to seasonal changes, regional feature changes, or changes in user behavior patterns.

[0104] Finally, the results output by the heat risk prediction module will be provided to other modules of the system in a structured form for use, and simultaneously recorded in the risk management platform for trend analysis, historical backtracking, and strategy evaluation. The prediction results can not only be visualized as a time series curve, but also be overlaid on the regional map to form a heat map, highlighting the areas that may form risks in the future period, and guiding managers and the scheduling platform to take intervention measures in advance.

[0105] In this way, by combining time series modeling and LSTM intelligent prediction, the heat risk prediction module 106 endows the system with the ability of proactive warning, enabling the identification of new energy vehicle aggregation risks to no longer stay at the current situation level only, but to be able to achieve dynamic prediction extending into the future, providing key technical support for building a more efficient, intelligent, and safe urban new energy operation environment.

[0106] To facilitate those skilled in the art to better understand and implement the LSTM model used in the heat risk prediction module, a specific example is provided below.

[0107] For example, this LSTM model consists of three main parts, namely the input layer, the LSTM processing layer, and the output layer.

[0108] The role of the input layer is to receive multi-dimensional feature data continuously collected within a time window. The data at each time step includes the total number of vehicles, vehicle density per unit area, average temperature of the aggregation cluster, proportion of high-temperature vehicles, proportion of vehicles in abnormal state, regional charging load ratio, and average temperature rise rate, a total of 7 features. Assuming that the length of the time window is set to 6 time steps, corresponding to the past 30 minutes (5 minutes per step), the input received by the input layer is a two-dimensional matrix with a shape of 6×7, where 6 represents the number of time steps and 7 represents the feature dimension per step.

[0109] The output of the input layer is directly passed to the LSTM processing layer. This processing layer includes a standard single-layer LSTM structure, and the number of its units can be set according to computing power, and it is recommended to set it to 32 hidden units. The LSTM processing layer gradually reads each time step of the input sequence and updates the internal state vector and output state at each moment. The final output of this processing layer is a one-dimensional vector with a length of 32, representing the representative state representation extracted by the model after comprehensively considering the historical time series features.

[0110] The output layer receives the final output vector of the LSTM processing layer and maps this vector to a scalar through a fully connected (fully connected means dense) neural network layer, that is, to predict the risk score value for the next time step (such as 5 minutes later). This score value will be sent to the risk level determination module during deployment and compared with a preset threshold to assist in judging whether to take response control in advance.

[0111] This model can be implemented using mainstream deep learning frameworks (such as TensorFlow or PyTorch). The model parameters can be trained through historical labeled datasets. The loss function uses mean squared error (MSE), and it is recommended to use the Adam algorithm as the optimizer. During the training stage, a labeled sample sequence (the input is time series features and the label is the actual future score value) needs to be provided to optimize the model weights in a supervised manner. During the inference stage, the model runs once every prediction period (such as 5 minutes) and continuously outputs the score prediction results for the next time step.

[0112] This structure has good real-time performance and deployment efficiency while ensuring prediction accuracy, and is suitable for dynamic thermal risk prediction of the aggregation situation of new energy vehicles at the urban level.

[0113] Furthermore, the thermal risk prediction module is specifically used for: Based on the operation state data, construct a multi-dimensional time series feature dataset including vehicle geographical location, battery temperature, state of charge, and vehicle operation state labels, and use a sliding window method to segment the feature dataset in chronological order; Using a long short-term memory neural network model adjusted by introducing a dynamic attention mechanism to encode the time series features within each sliding window, obtaining a sequence of hidden state vectors for capturing the temporal patterns and spatial evolution trends of vehicle aggregation behaviors; According to the spatial distribution of the sequence of hidden state vectors and its relationship with the boundary structure of the aggregation clusters, generating the potential risk score trend for each target area within a preset future time range, and outputting the potential risk score trend in the form of a structured tensor; Performing spatial mapping processing on the output in the form of a structured tensor, converting it into a heat map image format at the regional level according to the geographical grid encoding rules of the target area, and projecting it onto the target area map to form a predictive high-risk aggregation area layer; Outputting the risk score trend of each grid in the high-risk aggregation area layer to a risk level determination module and a response control scheduling module to complete the scheduling planning of flow-limiting control, vehicle guidance, and load dispersion strategies within the area in advance.

[0114] In the new energy vehicle aggregation prevention and control system based on clustering recognition and regional risk modeling described in the present invention, the heat risk prediction module is the core component of the prediction and early warning functions of the entire system. Its role is to construct a prediction model with time dependence and spatial evolution laws by analyzing historical and current new energy vehicle operation status data, so as to judge the future heat risk evolution trend in advance and project the prediction results visually onto the target area map to assist in the pre-identification and intervention planning of high-risk aggregation areas.

[0115] The heat risk prediction module first uses the operation status data as the basis. The operation status data includes vehicle geographical location, battery temperature, state of charge, and vehicle operation status labels. These data are all indexed by timestamps, constituting a continuous original time series data set. To effectively analyze these time series data, the module adopts a sliding window processing strategy. That is, a time window with a fixed length is selected, and the entire original time series data set is sequentially traversed according to the set sliding step. The data segment within each window constitutes an independent subsequence, and this subsequence contains the four-dimensional features of each vehicle at multiple moments: geographical coordinates, battery temperature, state of charge, and operation status label. Since the dimensions and variation ranges of different features vary greatly, after the sliding window segmentation is completed, the system first normalizes all features to give them a unified scale basis for the subsequent input encoding of the neural network model.

[0116] After completing the feature normalization process, the time series data within each sliding window will be used as input and fed into a long short-term memory neural network model introduced with a dynamic attention mechanism for adjustment. This model is constructed based on the basic structure of the long short-term memory network (LSTM) and contains multiple stacked LSTM cell layers inside. Each layer receives the state at the previous moment as input and outputs the hidden state vector at the current moment. To enhance the model's response ability to critical moments and critical vehicle states, a dynamic attention mechanism is introduced. In each time window, the attention mechanism dynamically evaluates the importance of different time slices and feature dimensions, and weights the hidden state to strengthen the expression weights of the regions with severe high-temperature fluctuations, vehicles with abnormal state of charge, and nodes with abnormal operating states in the hidden state vector, ensuring that the model focuses on the signal information that plays a key role in the evolution of thermal risks.

[0117] After processing each window, the model outputs the corresponding hidden state vector sequence, which carries the evolution trajectory information of the vehicle state within the entire time window. To convert the time series coding result into a spatially interpretable thermal risk trend output, the system further analyzes the aggregation distribution characteristics of the hidden state vector in the spatial dimension and combines the aggregation cluster boundary structure information provided by the clustering recognition module to determine whether the current hidden state vector forms a high-risk trend aggregation. If the hidden state is densely concentrated in a certain spatial region and contains a large number of high-temperature or high-risk state vehicles during the corresponding period, this region will be marked as a potential risk hot spot.

[0118] After obtaining the hidden state sequences of all sliding windows, the system statistically aggregates all time windows of each target region, calculates the potential risk score trend of this region within a preset future time period, and organizes it in the form of a structured tensor. The structured tensor is constructed according to the spatial dimension and the time dimension. Its spatial dimension is determined by the grid coding rule of the target region, and each grid corresponds to a fixed geographical location. The time dimension is arranged in the order of the sliding window time segments, so that the tensor value of each grid reflects the risk score prediction value of this geographical unit at different time slices. This structure not only retains the spatial differences but also presents the time evolution trend, which is convenient for subsequent modules to use.

[0119] After the output of the potential risk score trend in the above structured tensor form is completed, the system performs spatial mapping processing. According to the pre-constructed target area map and geographical grid division scheme, each tensor unit is mapped to a grid block in the actual map. The system maintains a unique association between the tensor value and the geographical grid during each mapping process, that is, each data point in the tensor can only correspond to a specific geographical location unit on the map. After the mapping is completed, the system generates a heat map in the form of an image at the regional level. Each pixel block represents a spatial unit, and the future risk score change trend of the unit is intuitively represented by a color gradient. This heat map can be updated in real time and dynamically overlaid on the map visualization interface to realize the dynamic display of the predictive high-risk aggregation area layer.

[0120] After generating the high-risk aggregation area layer, the system further processes the risk score trend corresponding to each grid in the layer and transmits it as a prediction result to the risk level determination module and the response control scheduling module. Among them, the risk level determination module uses the score trend result to determine whether there is a trend that the risk score of a certain grid will rise significantly and break through the high-risk threshold in the future time period. If so, the corresponding aggregation cluster will be promoted to a higher risk level in advance, triggering the early warning response mechanism. The response control scheduling module uses the score trend prediction result, combines the actual position of the vehicle and the current load status, and formulates a flow-limiting strategy, a path guidance strategy and a load dispersion plan in advance to realize the active intervention and dynamic optimization control of the regional thermal risk.

[0121] In the above manner, the thermal risk prediction module of the present invention not only constructs a complete time series feature processing and deep neural prediction path, but also realizes a highly integrated process from prediction to map layer display through the enhanced representation of the spatial structure, ensuring the feasibility, timeliness and forward-looking of the system in engineering deployment. At the same time, the dynamic attention mechanism and the structured spatial tensor output introduced in the module make the prediction process highly sensitive and interpretable, significantly superior to the traditional static regression model and the single-point scoring prediction-based scheme, and improving the recognition accuracy and response efficiency of the evolution process of the aggregation thermal risk of new energy vehicles.

[0122] Although this application is disclosed above with preferred embodiments, it is not intended to limit this application. Any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of this application. Therefore, the protection scope of this application should be determined by the scope defined by the claims of this application.

Claims

1. A new energy vehicle aggregation prevention and control system based on clustering recognition and regional risk modeling, characterized in that Including: A status acquisition module, which is used to acquire the operation status data of new energy vehicles in the target area. The operation status data includes the geographical location coordinates of the vehicle, battery temperature data, state of charge data, and vehicle operation status tags; A clustering and recognition module, which is used to identify aggregation clusters according to the operation status data by using a density-based spatial clustering algorithm, and obtain the aggregation characteristics of the aggregation clusters; A regional risk modeling module, which is used to calculate the risk score value of each aggregation cluster according to the aggregation cluster characteristics provided by the clustering and recognition module; The risk score value is obtained according to the vehicle density per unit area, the average temperature value of the aggregation cluster, the proportion value of high-temperature vehicles, the proportion value of abnormal-state vehicles, and the charging load ratio of the target area; A risk level determination module, which is used to determine the risk level of the corresponding aggregation cluster according to the risk score value; A response control and scheduling module, which is used to execute current-limiting control and path guidance measures according to the risk level, and provide response results to an external platform; A thermal risk prediction module, which is used to construct a time series feature data set based on a sliding window according to the operation status data, and use a pre-trained long short-term memory neural network model to predict the potential risk score trend of each target area within a preset future time range; output the predicted potential risk score trend to the risk level determination module and the response control and scheduling module to execute current-limiting or guidance strategies in advance.

2. The new energy vehicle aggregation prevention and control system based on clustering recognition and regional risk modeling according to claim 1, wherein, The clustering and recognition module is specifically used for: Construct a four-dimensional state space combining the geographical location coordinates of the vehicle, battery temperature data, state of charge data, and vehicle operation status tags. The four-dimensional state space is normalized through a state similarity function guided by state tags, so that different-dimensional features have a unified measurement basis in spatial distance calculation; In the four-dimensional state space, execute a dynamic density estimation mechanism. The dynamic density estimation mechanism dynamically adjusts the neighborhood search radius and minimum sample number of the clustering algorithm according to the proportion of high-temperature vehicles and the density fluctuation of abnormal-state vehicles in the local space, so that the clustering result has higher sensitivity to potential thermal runaway areas; After completing the preliminary clustering, introduce a structural stability determination process for each aggregation cluster, identify the boundary unstable area based on the fluctuation degree of the temperature gradient at the cluster boundary, and perform boundary correction to supplement the edge vehicles missed by the critical state in the clustering process; Output the aggregation characteristics of each obtained aggregation cluster. The aggregation characteristics include the number of vehicles, the spatial coverage, the vehicle density per unit area, the average temperature value of the aggregation cluster, the proportion value of high-temperature vehicles, the proportion value of abnormal-state vehicles, and the aggregation cluster boundary temperature fluctuation index, which are used to characterize the thermal risk sensitivity and evolution tendency of the aggregation cluster.

3. The new energy vehicle aggregation prevention and control system based on clustering recognition and regional risk modeling according to claim 1, characterized in that, The regional risk modeling module is specifically used for: Perform differential calculation on the vehicle density per unit area and the average temperature value of each aggregation cluster within a continuous time period to obtain the density change rate and the temperature rise rate; Perform sliding window statistics on the boundary temperature data of the aggregation cluster within a specified time period, and extract the boundary temperature fluctuation amplitude, which is used to characterize the structural stability of the aggregation cluster; Proportionally correct the vehicle density per unit area, the average temperature value of the aggregation cluster, the proportion value of high-temperature vehicles, and the proportion value of abnormal-state vehicles according to the density change rate, the temperature rise rate, and the magnitude of the boundary temperature fluctuation. Use the corrected vehicle density per unit area, the average temperature value of the aggregation cluster, the proportion value of high-temperature vehicles, the proportion value of abnormal-state vehicles, and the charging load ratio of the target area together to calculate the risk score value of each aggregation cluster.

4. The new energy vehicle aggregation prevention and control system based on clustering recognition and regional risk modeling according to claim 1, characterized in that The risk level determination module is specifically used for: Obtain the risk score values output by the area risk modeling module for the same aggregation cluster within a continuous number of time slices, and construct a score time series in chronological order. Based on the score time series, calculate the risk score increment value and its average change rate between each time slice, which are used to characterize the short-term score growth trend of this aggregation cluster. According to the short-term score growth trend, combined with the current score value and the static aggregation cluster characteristics including the vehicle density per unit area and the average temperature value of the aggregation cluster, predict the risk score change range within a preset future time window, and output the future score prediction value interval. Conduct an interval overlap analysis between the future score prediction value interval and the static risk level threshold to identify whether there is an early warning signal of trend crossing the high-risk threshold. In the case where the early warning signal is established, combine the current score value and the upper limit of the future score prediction value interval, construct a trend strengthening factor, and correct the current score value and then output the final score result. Compare the final score result with the risk level determination rule to obtain the current risk level corresponding to this aggregation cluster, and use it as the output of the risk level determination module.

5. The new energy vehicle aggregation prevention and control system based on clustering recognition and regional risk modeling according to claim 1, characterized in that The thermal risk prediction module is specifically used for: Based on the operation state data, construct a multi-dimensional time series feature dataset including vehicle geographical location, battery temperature, state of charge, and vehicle operation state labels, and use a sliding window method to segment the feature dataset in chronological order. Use a long short-term memory neural network model with a dynamically introduced attention mechanism to adjust to encode the time series features within each sliding window, and obtain a sequence of hidden state vectors, which are used to capture the temporal pattern and spatial evolution trend of vehicle aggregation behavior. According to the spatial distribution of the sequence of hidden state vectors and its relationship with the boundary structure of the aggregation cluster, generate the potential risk score trend of each target area within a preset future time range, and output the potential risk score trend in the form of a structured tensor. Perform spatial mapping processing on the output in the form of a structured tensor, convert it into a regional-level heat map image format according to the geographical grid coding rule of the target area, and project it onto the target area map to form a predictive high-risk aggregation area layer. Output the risk score trend of each grid in the high-risk aggregation area layer to the risk level determination module and the response control scheduling module to complete the scheduling planning of current-limiting control, vehicle guidance, and load dispersion strategies within the area in advance.

6. The new energy vehicle aggregation prevention and control system based on clustering recognition and regional risk modeling according to claim 3, wherein The area risk modeling module is also used for: Based on the spatial boundary morphology of each aggregation cluster, identify the position sequence of boundary vehicles within a continuous time period, and construct a boundary trajectory set, which is used to retain the geographical distribution path of boundary vehicles changing over time; Using the boundary trajectory set, extract the battery temperature of each boundary vehicle within the corresponding time slice, and construct a boundary temperature sequence set with time indices, which is used to record the dynamic evolution process of the thermal state at the boundary of the aggregation cluster; Based on the boundary temperature sequence set, extract the temperature variation amplitude, temperature rise rate, and length of the continuous temperature difference section within each window in a time-sliding window manner, and form a local temperature perturbation feature vector sequence; Perform boundary direction aggregation on the local temperature perturbation feature vector sequence, construct an overall boundary temperature perturbation characteristic curve according to the arrangement order of each boundary segment in the spatial trajectory, and extract the boundary temperature fluctuation amplitude based on its fluctuation amplitude as the output, which is used to characterize the thermal stability risk of the aggregation cluster structure.

7. The new energy vehicle aggregation prevention and control system based on clustering recognition and regional risk modeling according to claim 4, wherein The risk level determination module is also used for: Decompose the future score prediction value interval into three indicators: a lower limit value, an upper limit value, and an interval span, and calculate the distance between the interval span and the current score value based on the short-term score growth trend of the score time series, which is used to characterize the possible growth fluctuation range of the future score; According to the fluctuation range, retrieve all level boundary values greater than the current score value in the static risk level thresholds, screen out the set of boundary values with numerical overlap relationships, and output the score difference interval corresponding to the set of boundary values; Compare the score difference interval with the fluctuation range, calculate the coverage ratio of each boundary value within the fluctuation range, and use the ratio as the trend crossing intensity index to determine whether the future score interval has a tendency to migrate to a high risk level; Generate an early warning signal when the trend crossing intensity index reaches a preset judgment threshold.

Images