Risk grading and classifying method for charging station

Abstract

The invention relates to the technical field of electric vehicle charging safety, and discloses a charging station risk grading and classifying method, which comprises the following steps: acquiring safety quantitative evaluation indexes of different vehicle types in each SOC stage, integrating multi-source data, extracting risk correlation characteristics, identifying a periodic risk mode by using time sequence analysis, obtaining the charging safety coefficient of the SOC in each stage, and classifying the charging safety coefficient of the SOC in each stage. According to the method, accurate acquisition and fusion analysis of multi-dimensional data are realized, a risk early warning model is trained through supervised learning LightGBM and 3-sigma multi-level screening and a hybrid neural network algorithm, an unknown abnormal mode is identified in combination with abnormal detection of an isolated forest and an auto-encoder, operation fault reasons are divided according to risk grade scores, and a risk early warning result is obtained. Graded response and automatic control are achieved, a multi-site data updating model is aggregated through federal learning, the generalization ability is improved while privacy is protected, and dynamic optimization of model parameters and scenarized adjustment of risk thresholds are achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of electric vehicle charging safety, and particularly to a method for classifying and grading risks of charging stations. Background Art

[0002] Due to the complexity and diversity of accidents, how to scientifically and reasonably sort out and classify various risk factors, establish accurate and appropriate evaluation levels, and thus quickly and effectively formulate and implement the most suitable countermeasures according to the danger levels of different risk levels, minimizing the possible harm caused by accidents to the greatest extent, has become a key problem that the current electric vehicle charging industry urgently needs to overcome, and its urgency and importance are self-evident. This not only concerns the stable operation of charging stations, but also directly relates to the life and property safety of the general users and the healthy development of the entire industry.

[0003] Chinese patent document CN117937700A discloses an "Internet-based lithium battery charge and discharge safety warning system". Specifically, it is an Internet-based lithium battery charge and discharge safety warning system, including an interference analysis unit, a battery supervision unit, a surrounding comprehensive analysis unit, a battery analysis unit, and a risk warning unit; in the present invention, by monitoring the charge and discharge process of lithium batteries, the safety level classification of lithium batteries is realized according to the monitoring results, and combined with the comprehensive situation of the levels of all batteries in the charging area, the probability of an accident occurring during charging is determined, and then combined with the number of lithium batteries in the charging area, the consequences of the accident are estimated. Through the comprehensive analysis of the accident consequences and accident probability, and at the same time by combining the interference factors of the external environment, the lithium batteries that have been classified according to the safety level are further comprehensively analyzed to achieve the overall monitoring of the lithium battery charging process. The above technical solution depends on preset weight coefficients and thresholds and lacks universality in the face of different battery types, usage scenarios, and environmental conditions. Summary of the Invention

[0004] The present invention mainly solves the technical problem that the original technical solution depends on preset weight coefficients and thresholds and lacks universality in the face of different battery types, usage scenarios, and environmental conditions. It provides a method for classifying and grading the risks of charging stations, obtains safety quantification evaluation indicators for each SOC stage of different vehicle models, integrates multi-source data and extracts risk correlation features, uses time series analysis to identify periodic risk patterns, obtains the charging risk coefficients of each stage of SOC, realizes accurate collection and fusion analysis of multi-dimensional data, trains a risk warning model through supervised learning LightGBM and 3σ multi-level screening and hybrid neural network algorithms, labels historical fault data as labels, and at the same time combines the anomaly detection of isolation forest and autoencoder to identify unknown anomaly patterns, divides the causes of operating faults according to the risk level score, realizes hierarchical response and automatic control, and aggregates multi-site data through federated learning to update the model, improves the generalization ability while protecting privacy, and realizes the dynamic optimization of model parameters and the scenario-based adjustment of risk thresholds.

[0005] The above technical problems of the present invention are mainly solved by the following technical solutions: The present invention includes the following steps: S1. Collect data and obtain safety quantification evaluation indicators for each SOC stage of different vehicle models; S2. Integrate multi-source data and extract risk correlation features, use time series analysis to identify periodic risk patterns, obtain the charging risk coefficients of each stage of SOC, and conduct a quantitative evaluation of the risk level of the charging station; S3. Build a risk model and dynamically evaluate and divide the risk levels of the causes of operating faults; S4. Respond hierarchically according to the risk level and conduct automatic control; S5. Aggregate multi-site data to update the model.

[0006] Preferably, collect data from different sensors, devices, and systems, and perform cleaning, preprocessing, and storage, and dynamically adjust the sampling frequency and threshold according to historical extreme climate data.

[0007] Preferably, obtain the voltage risk coefficient and the early warning threshold fm of the maximum charging risk coefficient during the charging process of electric vehicles according to the collected data, and set the early warning voltage risk coefficient threshold Vfm and the early warning current risk coefficient threshold Afm for safety quantification evaluation.

[0008] Preferably, different vehicle models have independent safety quantification evaluation indicators, calculate safety indicators at each SOC stage of different vehicle models, and if a SOC stage has multiple different required voltages, required currents, measured voltages, and measured currents, calculate according to the average value.

[0009] Preferably, the calculated risk coefficient of the electric vehicle follows a normal distribution. First, the standard risk coefficient is obtained according to the specific vehicle model, and then the real-time detected data is substituted to obtain the charging risk coefficient of each stage of the SOC.

[0010] Preferably, compare the early warning threshold fm of the maximum charging risk coefficient with the risk coefficients of each SOC stage. If the risk coefficient of a certain SOC stage is greater than the early warning threshold, a charging safety warning is given and the corresponding risk level score is obtained.

[0011] Preferably, if the risk coefficient of a certain SOC stage is greater than or equal to 0.75fm and less than fm, it is determined to be passing; if the risk coefficient of a certain SOC stage is greater than or equal to 0.5fm and less than 0.75fm, it is determined to be medium; if the risk coefficient of a certain SOC stage is greater than or equal to 0.25fm and less than 0.5fm, it is determined to be good; if the risk coefficient of a certain SOC stage is greater than or equal to 0 and less than 0.25fm, it is determined to be excellent.

[0012] Preferably, risk modeling specifically includes training a risk early warning model, labeling historical fault data as labels, combining the anomaly detection of the isolation forest and the autoencoder to identify unknown anomaly patterns, and classifying the causes of operation failures into four levels: normal, low risk, medium risk, and high risk according to the risk level score.

[0013] Preferably, for high risk: trigger automatic power-off + push an alarm to the operation and maintenance personnel and the supervision platform + send a text message to notify the user of the abnormal information + display the specific abnormal information of the user in a page pop-up window; for medium risk: send a text message to notify the user of the abnormal information + display the specific abnormal information of the user in a page pop-up window + generate a work order and dispatch it to the operation and maintenance APP for processing within a limited time; for low risk: send a text message to notify the user of the abnormal information + display the specific abnormal information of the user in a page pop-up window + include it in the periodic inspection plan.

[0014] Preferably, aggregate multi-site data through federated learning to update the model, improve the generalization ability while protecting privacy. The operation and maintenance personnel's processing results are reversely labeled to the system to form a closed loop of "risk identification - disposal - verification - optimization", and the historical risk events and disposal plans are constructed into a knowledge graph to assist in new scenario decision-making.

[0015] The beneficial effects of the present invention are as follows: realizing the accurate acquisition and fusion analysis of multi-dimensional data, dynamically adapting to the risk prediction in the high-density urban charging station environment under extreme climate conditions; quantifying in real time the impact of environmental mutations on the overall risk level, introducing an adaptive learning mechanism or edge computing technology to realize the dynamic optimization of model parameters and the scenario-based adjustment of risk thresholds; quickly locating high-risk batteries, predicting the trend of group risk diffusion, and realizing cross-platform data interconnection and instruction coordination. Description of the Drawings

[0016] Figure 1 It is a flow chart of the present invention.

[0017] Figure 2 It is a system architecture diagram of the present invention.

[0018] Figure 3 It is a feature extraction framework diagram of the present invention. Detailed implementation manners

[0019] To make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the technical solutions of this application will be further described in detail below through embodiments and in conjunction with the accompanying drawings. It should be understood that the specific implementation manners described herein are only the best embodiments of this application, which are only used to explain this application and do not limit the protection scope of this application. All other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the protection scope of the present invention.

[0020] The prior art Internet-based lithium battery charge and discharge safety warning system includes an interference analysis unit, a battery supervision unit, a peripheral comprehensive analysis unit, a battery analysis unit and a risk warning unit. The battery supervision unit can obtain the charging safety performance and discharge safety performance of the lithium battery during the charge and discharge process of the lithium battery, and send the charging safety performance and discharge safety performance to the battery analysis unit; The battery analysis unit conducts comparative analysis according to the charging safety performance, classifies the lithium battery into a normal lithium battery, a low-risk lithium battery and a high-risk lithium battery according to the analysis result, and sends the classification of the lithium battery to the peripheral comprehensive analysis unit. The battery analysis unit classifies the discharge safety level of the lithium battery into a normal discharge battery and an abnormal discharge battery, and the battery analysis unit sends the charging safety level classification and the discharge safety level classification to the peripheral comprehensive analysis unit; The interference analysis unit can obtain the charge and discharge environment of the lithium battery, generate charging interference information according to the charge and discharge environment of the lithium battery, and send the charging interference information to the peripheral comprehensive analysis unit; After obtaining the charging interference information and the charging safety level classification of the lithium battery, the peripheral comprehensive analysis unit conducts charging statistics on the classification of the lithium battery, creates a charging statistics set, comprehensively analyzes the charging statistics set and the charging interference information, generates a charging risk warning, and sends the charging risk warning to the risk warning unit; The peripheral comprehensive analysis unit obtains the discharge safety level of the lithium battery. When the discharge safety level of the lithium battery is an abnormal discharge battery, it generates a discharge risk warning and generates an alarm reminder through the risk warning unit.

[0021] The prior art has problems with the universality and adaptability of algorithms: The algorithms in the prior art (such as the calculation of charging risk coefficients and discharging risk coefficients) rely on preset weight coefficients and thresholds. These parameters may need to be adjusted according to different battery types, usage scenarios, and environmental conditions, lacking universality. If the algorithm cannot adapt to different scenarios, it may cause the system to perform poorly or even fail in certain specific environments.

[0022] Failure to consider battery aging and lifespan issues: The patent does not clearly mention how to consider the impact of battery aging on safety performance. As the battery usage time increases, its internal chemical properties will change, which may lead to an increase in safety risks. Ignoring the battery aging factor may cause the system to be unable to accurately evaluate the safety risks during long-term use.

[0023] In the risk classification method for charging stations, the core technical problems to be solved include how to achieve precise collection and fusion analysis of multi-dimensional data, and how to dynamically adapt to risk prediction in a high-density urban charging station environment under extreme climate conditions. First, the charging and discharging processes of lithium batteries in a charging station involve real-time monitoring of multiple parameters such as temperature, voltage, current, environmental temperature and humidity. The collection of these data needs to solve problems such as insufficient sensor accuracy, signal interference, and data transmission delay. Especially in large-scale charging scenarios, extreme climates such as high temperature, high humidity, and extreme cold occur frequently, and environmental parameters fluctuate violently. The simultaneous operation of a large number of batteries may lead to an increase in the pressure of concurrent data processing. Traditional single-dimensional monitoring methods are difficult to effectively distinguish the risk correlation between individual batteries and groups. For example, when the temperatures of multiple batteries are abnormal at the same time, the system needs to quickly determine whether it is a local environmental interference or a battery self-fault. However, the prior art often relies on static threshold judgment and lacks the comprehensive analysis ability of dynamically coupling the battery's historical state, operation duration, and environmental factors, which is prone to misjudgment or missed judgment.

[0024] Secondly, the dynamic adaptability and scalability of the risk classification model face challenges. The types, aging degrees, and usage scenarios of lithium batteries vary significantly. A general classification algorithm may not be able to adapt to the characteristics of different brands or models of batteries. At the same time, in a high-density charging scenario, the demand for real-time transmission and processing of massive data surges. The temperature rise laws of new high-energy-density batteries are different from those of conventional batteries. If the algorithm is not optimized for their characteristics, it may lead to deviation in risk assessment. In addition, the charging station environment is complex and changeable (such as extreme weather, electromagnetic interference, etc.). Existing methods have deficiencies in the coupling analysis of environmental interference and internal battery risks. For example, the decline in battery heat dissipation efficiency in a high-temperature environment may accelerate risk accumulation. However, if the system only relies on preset environmental correction coefficients, it is difficult to quantify the impact of environmental mutations on the overall risk level in real time. There is an urgent need to introduce an adaptive learning mechanism or edge computing technology to achieve dynamic optimization of model parameters and scenario-based adjustment of risk thresholds.

[0025] Finally, the real-time response and collaborative early warning capabilities of the system still need to be improved. The risk grading of charging stations not only requires quickly locating high-risk batteries but also predicting the diffusion trend of group risks, which poses higher requirements for data processing speed and algorithm efficiency. The traditional centralized architecture may cause early warning delays due to data transmission bottlenecks, while distributed computing can improve efficiency but faces technical challenges in multi-node collaborative consistency. For example, when a battery cluster in a certain area triggers an early warning due to local overheating, the system needs to synchronously adjust the charging strategies of surrounding batteries to avoid chain reactions, but there is currently no mature solution for multi-unit collaborative decision-making and dynamic resource allocation. In addition, the risk grading results need to be seamlessly connected to the charging station management system, fire protection facilities, and emergency response mechanisms. How to achieve cross-platform data interconnection and instruction collaboration remains an important obstacle to the implementation of the technology.

[0026] Before discussing the exemplary embodiments in more detail, it should be noted that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts depict operations (or steps) as sequential processes, many of the operations (or steps) can be implemented in parallel, concurrently, or simultaneously. In addition, the order of the operations can be rearranged. The process can be terminated when its operations are completed, but it can also have additional steps not included in the drawings; the process can correspond to a method, function, procedure, subroutine, subprogram, and so on.

[0027] The technical solution of the present invention will be further specifically described below through embodiments in conjunction with the accompanying drawings.

[0028] Embodiment 1: A method for classifying and grading risks of a charging station in this embodiment, as Figure 1 shown, includes the following steps: S1. Collect data and obtain safety quantification evaluation indicators for each SOC stage of different vehicle models. Collect data from different sensors, devices, and systems, clean, preprocess, and store it, and dynamically adjust the sampling frequency and threshold according to historical extreme climate data. Obtain the voltage risk coefficient during the electric vehicle charging process and the early warning threshold fm of the maximum charging risk coefficient from the collected data, and set the early warning voltage risk coefficient threshold Vfm and the early warning current risk coefficient threshold Afm for safety quantification evaluation.

[0029] Different vehicle models have independent safety quantification evaluation indicators. Calculate safety indicators for each SOC stage of different vehicle models. If a SOC stage has multiple different required voltages, required currents, measured voltages, and measured currents, calculate according to the average value.

[0030] S2. Integrate multi-source data and extract risk correlation features, use time series analysis to identify periodic risk patterns, obtain the charging risk coefficients of each stage of SOC, and conduct quantitative evaluation of the charging station risk level.

[0031] The calculated risk coefficient of the electric vehicle follows a normal distribution. First, the standard risk coefficient is obtained according to the specific vehicle model, and then the real-time detected data is substituted to obtain the charging risk coefficient of each stage of SOC. Compare the maximum charging risk coefficient warning threshold fm with the risk coefficients of each SOC stage. If the risk coefficient of a certain SOC stage is greater than the warning threshold, a charging safety warning is given and the corresponding risk level score is obtained.

[0032] If the risk coefficient of a certain SOC stage is greater than or equal to 0.75fm and less than fm, it is determined to be passing; if the risk coefficient of a certain SOC stage is greater than or equal to 0.5fm and less than 0.75fm, it is determined to be medium; if the risk coefficient of a certain SOC stage is greater than or equal to 0.25fm and less than 0.5fm, it is determined to be good; if the risk coefficient of a certain SOC stage is greater than or equal to 0 and less than 0.25fm, it is determined to be excellent.

[0033] S3. Risk modeling and dynamically evaluating and dividing the risk levels of the causes of operating failures. Risk modeling specifically includes training a risk warning model, labeling historical fault data as labels, combining the anomaly detection of the isolation forest and the autoencoder to identify unknown anomaly patterns, and dividing the causes of operating failures into four levels: normal, low risk, medium risk, and high risk according to the risk level score.

[0034] S4. Classify responses according to the risk level and automate control. High risk: Trigger automatic power-off + push an alarm to the operation and maintenance personnel and the supervision platform + send a text message to notify the user of the abnormal information + display the specific abnormal information of the user in a page pop-up window; Medium risk: Send a text message to notify the user of the abnormal information + display the specific abnormal information of the user in a page pop-up window + generate a work order and dispatch it to the operation and maintenance APP for processing within a limited time; Low risk: Send a text message to notify the user of the abnormal information + display the specific abnormal information of the user in a page pop-up window + include it in the periodic inspection plan.

[0035] S5. Aggregate multi-site data to update the model. Aggregate multi-site data to update the model through federated learning, improve the generalization ability while protecting privacy. The operation and maintenance personnel's processing results are reversely labeled to the system to form a closed loop of "risk identification - disposal - verification - optimization", and the historical risk events and disposal plans are constructed into a knowledge graph to assist in new scenario decision-making.

[0036] Embodiment 2 S1: Data collection and preprocessing S2: Risk feature extraction and integration S3: Risk modeling and dynamic evaluation S4: Hierarchical response and automated control S5: Knowledge precipitation and continuous optimization The above-mentioned S1 data collection and preprocessing includes the following steps: The main tasks of data collection and processing are to collect data from different sensors, devices and systems in real time and accurately, and perform necessary cleaning, preprocessing and storage. Through efficient data processing, ensure that the data can provide accurate information in subsequent analysis, query, certification and traceability. Dynamically adjust the sampling frequency and threshold according to historical extreme climate data (such as continuous high temperature or extreme cold records). When the temperature exceeds 40°C, increase the temperature monitoring frequency from 1 time per minute to 1 time per 10 seconds, and introduce an environmental temperature and humidity coupling correction coefficient to eliminate the interference of extreme climate on the sensor accuracy.

[0037] Based on the real-time data obtained by the electric vehicle charging safety warning device, a quantitative evaluation of the risk level of the charging station is carried out. Different vehicle models have different required voltages and required currents. When some electric vehicles are charging, the data read during the charging stage shows that the required voltage and required current do not change with the progress of the charging process, while for some electric vehicles, the required voltage and required current change continuously with the progress of the charging process. However, the calculation principle of the quantitative index is the same. Let the required voltage for electric vehicle charging be Vr, the measured voltage of the electric vehicle be Vm, the voltage risk coefficient during the electric vehicle charging process be Vf, and the maximum allowable voltage Vb be obtained during the parameter configuration stage of the electric vehicle charging. Then the risk coefficient calculation formula is as follows:

[0038] For the quantitative evaluation of charging safety warning, it is necessary to set the threshold value Vfm of the warning voltage risk coefficient. Therefore, the maximum allowable voltage Vb and the required voltage Vr for electric vehicle charging are combined to form the calculation formula as follows:

[0039] Let the required current for electric vehicle charging be Ar, the measured current of the electric vehicle be Am, and the current risk coefficient during the electric vehicle charging process be AϜ. Then the risk coefficient calculation formula is as follows:

[0040] For the quantitative evaluation of charging safety warning, it is necessary to set the threshold value Afm of the warning current risk coefficient. Therefore, the maximum allowable total charging current Ab and the required current Ar for electric vehicle charging are combined to form the calculation formula as follows:

[0041] The voltage risk coefficient and current risk coefficient calculated according to the formula can be used for safety quantitative evaluation based on this coefficient. Each vehicle model has its own safety quantitative evaluation index, and the safety index needs to be calculated at each SOC stage. If there are multiple different required voltages, required currents, measured voltages, and measured currents in one SOC stage, the average values of the required voltage, required current, measured voltage, and measured current are calculated and substituted into the risk coefficient calculation formula for calculation.

[0042] The S2 risk feature extraction and integration include the following steps: Integrate multi-source data (charging duration, measured current, measured voltage, output current, output voltage, highest voltage of single battery, lowest voltage of single battery, highest temperature of single battery, SOC, lowest temperature of single battery), and store them uniformly in the big data platform. Extract risk-related features (overcurrent of the whole battery pack, overvoltage of the whole battery pack, overvoltage of single battery, abnormal charging SOC, over-difference of charger output current, over-difference of charger output voltage, battery overtemperature, charging SOC out of range, required current exceeding the maximum allowable total charging current, required voltage exceeding the maximum allowable total charging voltage, output short-circuit startup, short-circuit during operation, output reverse polarity, output adhesion, overtemperature protection, input overvoltage, input undervoltage, communication abnormality), and use time series analysis to identify periodic risk patterns.

[0043] The calculated electric vehicle risk coefficient follows a normal distribution. First, obtain the standard risk coefficient according to the specific vehicle model, and then substitute the real-time detected data into the calculation formula to obtain the charging risk coefficient of each stage of SOC. Given that the early warning threshold of the maximum charging risk coefficient is Vfm, and the voltage risk coefficient of each SOC stage is Vsoc, the formula for quantitative scoring of the charging station risk level is:

[0044] If the risk coefficient of a certain SOC stage is greater than the early warning threshold, charging safety warning is carried out and the corresponding risk level score is obtained. If the calculated result is greater than or equal to 0.75*Vfm and less than Vfm, it is judged as passing (60 - 69 points), but the user will also be reminded to pay attention to battery maintenance.

[0045]

[0046] If the calculated result is greater than or equal to 0.5*Vfm and less than 0.75*Vfm, it is judged as medium (70 - 79 points), but the user will also be reminded to pay attention to battery maintenance.

[0047]

[0048] If the calculated result is greater than or equal to 0.25*Vfm and less than 0.5*Vfm, it is determined as good (80 - 89 points), and the warning device does not give a warning.

[0049]

[0050] If the calculated result is greater than or equal to 0 and less than 0.25*Vfm, it is determined as excellent (90 - 100 points), and the battery state is very healthy.

[0051]

[0052] The detection of the current risk coefficient is similar to that of the voltage. Given that the warning threshold of the maximum charging risk coefficient is Afm and the voltage risk coefficient at each SOC stage is Asoc, the warning threshold alarm formula for quantifying and scoring the risk level of the charging station is:

[0053] If the risk coefficient at a certain SOC stage is greater than the warning threshold, a charging safety warning is given. If the calculated result is greater than or equal to 0.75*Afm and less than Afm, it is determined as passing (60 - 69 points), but the user will also be reminded to pay attention to the battery maintenance.

[0054]

[0055] If the calculated result is greater than or equal to 0.5*Afm and less than 0.75*Afm, it is determined as medium (70 - 79 points), but the user will also be reminded to pay attention to the battery maintenance.

[0056]

[0057] If the calculated result is greater than or equal to 0.25*Afm and less than 0.5*Afm, it is determined as good (80 - 89 points), and the warning device does not give a warning.

[0058]

[0059] If the calculated result is greater than or equal to 0 and less than 0.25*Afm, it is determined as excellent (90 - 100), the battery state is very healthy, and no warning is required.

[0060]

[0061] According to the above judgment criteria, we can conduct a quantitative evaluation of the risk level of the charging station.

[0062] The S3 risk modeling and dynamic assessment includes the following steps: It mainly focuses on electrical safety, equipment status, and environmental factors. In risk modeling, key parameters include voltage, current, temperature, power, etc. A risk warning model is trained through supervised learning LightGBM and a 3σ multi-level screening and hybrid neural network algorithm, and historical fault data is labeled as tags. At the same time, anomaly detection combining Isolation Forest and Autoencoder is used to identify unknown anomaly patterns. According to the risk level score, the reasons for operating faults are divided into four levels: normal, low risk, medium risk, and high risk.

[0063] The S4 hierarchical response and automated control include the following steps: High risk (below 60 points): Trigger automatic power-off + Push an alarm to the operation and maintenance personnel and the supervision platform + Send a text message to notify the user of the abnormal information + Pop up a window on the page to display the specific abnormal information of the user. For high-risk scenarios, secondary confirmation is required, and after verifying the feasibility, it can be executed. Medium risk (60 - 79 points): Send a text message to notify the user of the abnormal information + Pop up a window on the page to display the specific abnormal information of the user + Generate a work order and dispatch it to the operation and maintenance APP for processing within a limited time. Low risk (80 - 100 points): Send a text message to notify the user of the abnormal information + Pop up a window on the page to display the specific abnormal information of the user + Incorporate it into the periodic inspection plan.

[0064] The S5 knowledge precipitation and continuous optimization include the following steps: Aggregate multi-site data through federated learning to update the model, improving the generalization ability while protecting privacy. The processing results of the operation and maintenance personnel are reversely labeled in the system to form a closed loop of "risk identification - disposal - verification - optimization". Construct historical risk events and disposal plans into a knowledge graph to assist in decision-making for new scenarios.

[0065] 1) The overall framework diagram is as Figure 2 shown.

[0066] 2) The system business architecture is as Figure 3 shown.

[0067] Data acquisition layer: As the front-end perception unit of the intelligent monitoring system, it captures the key operating parameters (including voltage, current, SOC, etc.) of charging equipment and environmental monitoring indicators (such as temperature, humidity, combustible gas concentration, etc.) in real time through a distributed sensing network, synchronously accesses the monitoring information and equipment operating status data, and completes the reliable conversion from physical signals to digital information. This layer integrates equipment operating parameters, environmental perception data, and video stream information to form a real-time monitoring network covering the operating parameters of the entire life cycle of the equipment.

[0068] Transmission and Storage Layer: As the core hub of the system, the data transmission and storage layer realizes multi-channel transmission by adopting heterogeneous communication protocols. It obtains the vehicle charging dynamic data of the charging pile (such as charging power, battery status) in real time through the CAN bus, collects the environmental parameters collected by sensors through the RS-485 bus, and builds a remote transmission channel relying on the 4G wireless network to synchronously upload multi-source heterogeneous data to the cloud platform. After the data is standardized and cleaned, it is structurally encapsulated according to the GB / T 27930 protocol, and invalid information is filtered through an anomaly detection algorithm. Finally, a dual-mode storage architecture is realized - the local database executes hot data caching, and the cloud server completes cold data archiving, forming a hierarchical storage and multiple backup mechanism.

[0069] Feature Extraction Layer: First, it collects diverse operation data of the charging pile and battery (including parameters such as voltage, current, temperature, SOC, etc.) in real time through sensors and Internet of Things devices, and uniformly stores them in the central data platform after data cleaning and format standardization; then, through an intelligent analysis module, it converts this data into easily understandable risk signals, and counts the number of occurrences and duration of abnormal situations; finally, it introduces a time series analysis engine to perform sliding window statistics and periodic pattern mining on the feature data, and finally outputs a feature vector with time for the upper-layer risk warning system to call.

[0070] Modeling and Evaluation Layer: The system architecture of this modeling and evaluation layer includes four core links: First, it integrates the real-time monitoring information from electrical parameters (voltage / current fluctuations), device status (battery temperature, SOC value), and environmental data (temperature and humidity), and inputs it into the analysis system after removing obvious outliers through statistical screening methods; then it uses historical fault-labeled data to train an intelligent warning model using supervised learning LightGBM and 3σ multi-level screening and hybrid neural network algorithms, predicts potential risks by comparing the deviation degree of actual parameters from normal thresholds and combining the time series change trend; at the same time, it sets up an anomaly detection module, and for new fault modes that are difficult to predict, it identifies abnormal signals through feature recombination and pattern comparison; finally, it performs fusion analysis on the model prediction results and anomaly detection data, dynamically divides four levels of early warnings according to the risk occurrence probability and impact degree, and regularly optimizes the model parameters with newly generated operation and maintenance data to form a continuously evolving evaluation system.

[0071] Hierarchical Response Layer: When the system determines a high risk, it first automatically triggers the device protection program (such as emergency power-off), and at the same time sends a red alert to the operation and maintenance personnel and the supervision platform, and details the abnormal type to the user through text messages and pop-up windows on the charging pile screen. For medium-risk situations, while sending a warning message to the user, the system automatically generates a maintenance work order containing the fault description and location, and pushes it to the mobile application of the operation and maintenance personnel for the staff to complete the processing feedback. Low-risk events are informed to the user to pay attention to the abnormality through information prompts, and at the same time, the device is marked as a key attention object and arranged to be preferentially inspected in the next routine inspection cycle. After all response actions are executed, the system continuously monitors the change of the device status and feeds back the disposal effect to the risk assessment model to achieve dynamic optimization of the strategy.

[0072] Precipitation and Optimization Layer: A double-loop feedback mechanism is designed in the operation and maintenance link - the outer loop automatically labels the disposal results as training samples through reinforcement learning, driving the model to complete a complete iteration of "risk feature extraction → disposal strategy generation → protection effect verification → parameter dynamic tuning"; the inner loop introduces an active learning algorithm to initiate a targeted data collection request for high-uncertainty risk scenarios, improving the confidence of model decision-making. On this basis, through ontology modeling of a large number of historical events, a multi-dimensional associated knowledge graph is constructed. Finally, a closed-loop protection ecosystem driven by data, inheriting experience, and evolving dynamically is formed.

[0073] User Interaction Layer: The user interaction center, as the main operation interface of the system, integrates three major practical functions: information query, data acquisition, and instant alarm notification. Through a unified operation platform, the staff can quickly view the device operation status, retrieve historical records, and immediately start the emergency handling procedure when problems occur. This design makes team collaboration more efficient and helps managers make accurate judgments faster.

[0074] Beneficial Effects 1. Improve dynamic protection ability: Through multi-level risk modeling of LightGBM and hybrid neural networks, the system can real-time analyze complex working conditions such as battery pack voltage fluctuations and temperature gradient changes. For example, when the charging SOC exceeds the range, the threshold interval is corrected by combining historical fault labels, increasing the overvoltage / overtemperature warning accuracy by more than 30% and simultaneously reducing the false alarm rate, solving the unique problems in extreme climate and high-density scenarios.

[0075] 2. Improve the operation and maintenance collaboration efficiency: The hierarchical response mechanism compresses the disposal time of high-risk events to the second level (such as output short circuit triggering automatic power-off + three-terminal alarm linkage), and medium- and low-risk work orders are distributed through the APP and intelligently allocated in the inspection plan, reducing 70% of the redundant operations of manual verification.

[0076] 3. Realize cross-scenario adaptive protection: Based on multi-site data aggregation and collection, the system can adapt to the parameter differences of charging piles of different brands and battery models. For example, according to the difference in charging and discharging characteristics between lithium iron phosphate and ternary lithium batteries, the system can dynamically adjust the overcurrent protection threshold to achieve "one platform, multiple compatibility" security coverage.

[0077] 4. Extension of risk prediction capabilities: The knowledge graph's attribution analysis of historical faults (such as the correlation between communication anomalies and sudden temperature rise) can predict potential derivative risks, greatly enhancing the charging station's emergency response capabilities and accident prevention level when a fire occurs, forming a full-chain protection closed loop of "monitoring-handling-prevention".

[0078] This application intends to build a comprehensive charging pile risk warning system covering the entire chain of "data collection-risk assessment-dynamic response-knowledge accumulation". By effectively grading and handling the causes of failure risks of high-density urban charging stations under extreme climate environments and taking timely response measures, it can effectively avoid resource and manpower losses caused by improper handling of abnormal situations. By adopting real-time processing of multi-source heterogeneous data, using sensors to collect dynamic parameters such as voltage, current, and temperature, and combining data cleaning and standardization to ensure information reliability; the early warning system ensures the integration of multi-dimensional risk characteristics, integrating 18 types of risk characteristics such as battery cell status (such as maximum voltage, SOC abnormality), charger output indicators (current / voltage tolerance) and environmental factors (over-temperature protection), and using time series analysis to mine periodic risk patterns; using hybrid modeling and dynamic classification, LightGBM supervised learning is used to annotate historical fault labels, combined with the 3σ rule to screen steady-state anomalies, and isolated forests and autoencoders to detect unknown anomalies, to achieve a "normal-low-medium-high" four-level risk assessment; adopting a risk grading and classification closed-loop response mechanism, high-risk scenarios trigger automatic power-off + multi-terminal alarms (requires secondary confirmation to prevent misoperation), medium and low-risk systems and inspection plans; the entire system is continuously optimized, based on federated learning to aggregate multi-site data update models, to form a "risk identification-handling-verification" self-evolution capability.

[0079] Example 3 1. Scenario Overview A high-density DC fast charging station (model: State Grid Smart Charging SGC-8000) was deployed in the core business district of a coastal city in Hainan, integrating 150 multi-brand charging piles, serving more than 1,800 vehicles per day. The climate characteristics of this area are an average summer temperature of 40°C and humidity of 85%, and the charging piles are densely spaced (1.2 meters). The equipment is operating at high load for a long time, facing the following challenges: Extreme climate interference: high temperature and high humidity cause abnormal battery temperature rise (cell temperature > 60°C) and degradation of insulation performance (resistance < 100MΩ); poor compatibility among multiple brands and heterogeneous data formats; high-concurrency data processing: during peak charging periods, the peak data volume reaches 120,000 items per second, and the response delay of the traditional architecture exceeds 8 seconds.

[0080] 2. Implementation Details of the Technical Solution 2.1 Data Acquisition and Dynamic Calibration 2.1.1 Sensor Deployment Temperature and Humidity Monitoring: The Sensirion SHT45 sensor (accuracy ±0.2°C) is used, the sampling frequency is dynamically adjusted (1 time / minute at normal temperature → 1 time / 10 seconds at high temperature), and a compensation formula is introduced:

[0081] (H is the humidity percentage, compensating for the superposition effect of high temperature and high humidity).

[0082] Voltage / Current Acquisition: The HIOKI BT3563 high-precision module (range 0 - 1000V / 0 - 500A) is integrated to monitor the voltage of individual battery cells (threshold 4.2V) and the output current fluctuation (±5% tolerance) in real time.

[0083] 2.1.2 Multi-Protocol Data Normalization Through the Moxa UC-3112 protocol converter, the data is uniformly converted into the standardized JSON format, and a feature mapping table is established: Table 1 Feature Mapping Table of Example 3

[0084] 2.2 Risk Modeling and Real-Time Assessment 2.2.1 Model Training and Deployment Dataset: Based on historical failure data (50,000 labeled samples), 18 types of risk features are extracted (such as abnormal SOC deviation, temperature rise rate ΔT / Δt > 3°C / min).

[0085] Algorithm Architecture: The LightGBM model (learning rate 0.05, tree depth 8) and the hybrid neural network (3-layer LSTM + 2-layer fully connected) are used to train the dynamic risk assessment model.

[0086] Edge Computing Optimization: The NVIDIA Jetson Xavier NX module is deployed locally at the charging pile to run the lightweight isolation forest algorithm to detect input undervoltage (<200V) and communication anomalies (packet loss rate > 5%) in real time.

[0087] 2.2.2 Risk Classification Example A certain charging pile detects in a 45°C environment: Temperature of individual battery cell: 58°C (threshold 55°C); SOC Deviation: 8% (threshold 5%); Temperature Rise Rate: 4°C / min.

[0088] The system predicts that the probability of overheating within 10 minutes is 88% through the model and marks it as a high risk (score: 52 points).

[0089] 2.3 Hierarchical Response and Cooperative Control 2.3.1 High - risk Disposal Automatic power - off: Trigger the Schneider Easypact MCCB circuit breaker (model: EZD400) to cut off the power within 1 second; Multi - terminal alarm: Push the alarm to the operation and maintenance platform (Alibaba Cloud IoT platform), the fire protection system (Honeywell Xtralis VESDA), and the user side (7 - inch industrial display and control technology GTC - 700); Environmental linkage: Start the liquid - cooling system (Danfoss MGX - 20 magnetic levitation compressor, flow rate 40L / min) to force the temperature down below 50°C.

[0090] 2.3.2 Medium - risk Handling It is detected that the output current fluctuation during charging exceeds the tolerance (±12%). The system: Generate a maintenance work order (fault description: "Abnormal current fluctuation") and push it to the operation and maintenance APP (Enterprise WeChat); Limit the charging power to 70% and include it in the 4 - hour priority maintenance queue.

[0091] 2.3.3 Low - risk Optimization For the charging pile positions with slightly abnormal SOC (deviation 3%), automatically mark them as "periodic inspection objects" and arrange for key inspection of the battery health (trigger replacement suggestion when SOH < 85%) the next day.

[0092] 2.4 Knowledge Deposition and Continuous Optimization 2.4.1 Federal Learning Model Iteration Aggregate data from 10 surrounding stations to optimize the over - temperature threshold adaptive algorithm. For example, for lithium iron phosphate batteries, dynamically lower the over - charge protection threshold from 4.2V to 4.05V (compensate for the capacity decay effect in high - temperature environments).

[0093] 2.4.2 Knowledge Graph Construction Associate the historical event "humidity > 80% + salt spray corrosion → insulation resistance decline → leakage risk" to generate an operation and maintenance strategy: Prioritize replacing the charging guns (model: Yonggui Electric Appliance YGC - 65) at coastal stations every month; Mark the "extreme climate inspection list" in the knowledge base to guide operation and maintenance personnel to focus on checking terminal corrosion and tightness.

[0094] 3. Verification of Implementation Effect Table 2 Experimental Result Table of Example 3

[0095] 4. Conclusion In this case, through core technologies such as extreme climate dynamic calibration, multi-protocol data normalization, and edge-cloud collaborative computing, accurate risk classification and efficient response have been achieved in high-density urban charging stations. The deep combination of specific devices (such as SHT45 sensors and EZD400 circuit breakers) and algorithm parameters (such as federated learning weights and temperature rise rate thresholds) verifies the practicality and innovation of the patented technology. Implementation data shows that the system performs significantly in reducing false alarm rates, shortening response times, and improving cross-brand compatibility, providing a standardized solution for the safety management of charging stations under high load and extreme environments, and having broad industry promotion value.

[0096] Example 4 1. Scenario Overview A high-density DC fast charging station (model: State Grid Smart Charging NGC-9000) is deployed at a transportation hub in a cold northern city, integrating 200 multi-brand charging piles and serving more than 2,000 vehicles per day on average. The average winter temperature in this area is -25°C, and the extreme low temperature reaches -35°C. Moreover, the charging piles are densely arranged (spacing 1.0 meter), and the equipment has been facing the following challenges for a long time: Influence of extreme low temperature: The increase in battery internal resistance leads to a decrease in charging efficiency (SOC growth rate < 1% / min), and the risk of electrolyte solidification (temperature < -30°C); poor multi-protocol compatibility and significant differences in data formats; high concurrency and low-temperature latency: During the peak charging period, the data volume peak reaches 150,000 pieces / second, and the traditional system has a data transmission delay of more than 10 seconds due to low temperature.

[0097] 2. Implementation Details of Technical Solutions 2.1 Data Acquisition and Dynamic Calibration 2.1.1 Sensor Deployment Temperature monitoring: Omega PT100 platinum resistance sensors (accuracy ±0.1°C) are used to dynamically adjust the sampling frequency (from 1 time / minute at normal temperature to 1 time / 5 seconds at low temperature), and a low-temperature compensation formula is introduced:

[0098] (Tenv is the ambient temperature, compensating for the deviation of the compensating wire resistance with temperature change).

[0099] Voltage / current acquisition: Integrate Keysight 34470A high-precision modules (range 0 - 1500V / 0 - 600A) to monitor the voltage of each battery cell (threshold 3.8V) and charging efficiency (trigger an alarm when the SOC growth rate < 0.8% / min) in real time.

[0100] 2.1.2 Multi-protocol Data Normalization Through the Advantech ADAM-3600 protocol gateway, the data is uniformly converted into the standardized MQTT format, and a feature mapping table is established: Table 3 Feature Mapping Table of Example 4

[0101] Risk Modeling and Real-time Assessment 2.2.1 Model Training and Deployment Dataset: Based on historical fault data (60,000 labeled samples), 18 types of risk features are extracted (such as sudden increase in internal resistance at low temperature, stagnant growth of SOC, communication packet loss rate > 8%).

[0102] Algorithm Architecture: Based on historical fault data (60,000 labeled samples), 18 types of risk features are extracted (such as sudden increase in internal resistance at low temperature, stagnant growth of SOC, communication packet loss rate > 8%).

[0103] Edge Computing Optimization: Deploy the Renesas RZ / V2L edge AI module locally at the charging pile, run the lightweight autoencoder algorithm, and detect voltage dips (ΔV > 5%) and electrolyte solidification risks (temperature < -30°C for 5 minutes) in real time.

[0104] 2.2.2 Risk Grading Example A certain CATL CTP-600 pile detects in an environment of -30°C: Battery internal resistance: 25 mΩ (threshold 20 mΩ); SOC growth rate: 0.5% / min (threshold 0.8% / min); Ambient temperature: -32°C (for 6 minutes).

[0105] The system predicts through the model that the probability of electrolyte solidification within 15 minutes is 75%, marked as high risk (score 45 points).

[0106] 2.3 Hierarchical Response and Cooperative Control 2.3.1 High-risk Disposal Preheating Start: Trigger the in-vehicle PTC heating system (model: BorgWarner HVH250) to raise the battery temperature above -20°C; Reduce Charging Power: Limit the charging current to 300 A (rated 600 A), and prompt the user "Under low-temperature protection" through a 7-inch industrial screen (Xiankong Technology GTC-800); Multi-terminal Alarm: Push the alarm to the operation and maintenance platform (Huawei Cloud IoT), the emergency management center (Siemens Desigo CC), and the user APP (Teld).

[0107] 2.3.2 Medium-risk Handling The communication packet loss rate of a certain charging pile is detected to be 12%. The system generates a maintenance work order (fault description: "CAN FD bus anomaly") and pushes it to the operation and maintenance APP (DingTalk); switches to the redundant communication channel (4G network) to ensure data continuity.

[0108] 2.3.3 Low-risk optimization For the pile positions with slow SOC growth (0.6% / min), they are automatically marked as "low-temperature optimization objects", and a heating film installation plan (model: DuPont Kapton HN) is arranged.

[0109] 2.4 Knowledge precipitation and continuous optimization 2.4.1 Federal learning model iteration Aggregate data from 8 surrounding stations to optimize the low-temperature internal resistance compensation algorithm. For example, for ternary lithium batteries, the internal resistance threshold is dynamically adjusted from 20 mΩ to 22 mΩ (to compensate for the Ohmic polarization effect caused by low temperature).

[0110] 2.4.2 Knowledge graph construction Associate the historical event "temperature < -30°C + sudden increase in internal resistance → decrease in charging efficiency → user complaints", and generate operation and maintenance strategies: start the preheating mode (preheat 1 hour in advance) on extremely cold days; mark the "low-temperature operation and maintenance checklist" in the knowledge base to guide key inspections of the heating system and insulation performance.

[0111] 3. Implementation effect verification Table 4 Experimental result table of Example 4

[0112] 4. Conclusion In this case, through core technologies such as extreme low-temperature dynamic calibration, multi-protocol data normalization, and edge intelligent computing, accurate risk classification and efficient response have been achieved in extremely cold and high-density charging stations. The deep combination of specific devices (such as PT100 sensors, HVH250 heating systems) and algorithm parameters (such as federal learning compensation coefficients, internal resistance dynamic thresholds) verifies the reliability and innovation of the patented technology in extreme environments. The implementation data shows that the system performs outstandingly in reducing accident rates, improving data integration efficiency, and reducing user complaints, providing a reusable solution for the safety management of charging stations in low-temperature and high-load scenarios, and having broad industry promotion value.

[0113] The specific embodiments described herein are merely illustrative of the spirit of the present invention. The above embodiments only express several implementation manners of the present application, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the protection scope of the present application. It should be noted that those skilled in the art to which the present application pertains can make various modifications or supplements to the described specific embodiments or use similar ways to replace them, but they will not deviate from the spirit of the present application or exceed the scope defined by the appended claims. For those of ordinary skill in the art, without departing from the concept of the present application, multiple deformations and improvements can also be made. Therefore, the protection scope of the present application shall be subject to the appended claims.

Claims

1. A charging station risk classification method, characterized in that: The following steps are involved: S1. Collect data and obtain safety quantitative evaluation indicators at each SOC stage for different vehicle models; S2. Integrate multi-source data and extract risk-related features, use time series analysis to identify periodic risk patterns, obtain the charging safety factor of SOC at each stage, and conduct quantitative evaluation of the risk level of charging stations; S3. Risk modeling and dynamic evaluation to classify the risk level of the causes of operational failures; S4. Respond and automate control based on risk levels; S5. Aggregate multi-site data to update the model.

2. A charging station risk classification method according to claim 1, characterized in that: Collect data from different sensors, equipment and systems, clean, preprocess and store them, and dynamically adjust the sampling frequency and threshold based on historical extreme climate data.

3. A charging station risk classification method according to claim 1 or 2, characterized in that: The voltage safety factor of the electric vehicle charging process and the maximum charging safety factor warning threshold fm are obtained according to the collected data, and the warning voltage safety factor threshold Vfm and the warning current safety factor threshold Afm are set for safety quantitative evaluation.

4. A charging station risk classification method according to claim 3, characterized in that: Different models have independent safety quantitative evaluation indicators. Safety indicators are calculated at each SOC stage of different models. If a SOC stage has multiple different required voltages, required currents, measured voltages and measured currents, the calculation is based on the average value.

5. A charging station risk classification method according to claim 3, characterized in that: The calculated safety factor of electric vehicles follows a normal distribution. First, the standard safety factor is obtained based on the specific model, and then the real-time detected data is substituted to obtain the charging safety factor of the SOC at each stage.

6. A charging station risk classification method according to claim 3, characterized in that: Compare the maximum charging safety factor warning threshold fm and the safety factor of each SOC stage. If the safety factor of a certain SOC stage is greater than the warning threshold, a charging safety warning is issued and the corresponding risk level score is obtained.

7. A charging station risk classification method according to claim 6, characterized in that: If the safety factor of a certain SOC stage is greater than or equal to 0.75fm and less than fm, it is judged as passing; if the safety factor of a certain SOC stage is greater than or equal to 0.5fm and less than 0.75fm, it is judged as medium; if the safety factor of a certain SOC stage is greater than or equal to 0.25fm and less than 0.5fm, it is judged as good; if the safety factor of a certain SOC stage is greater than or equal to 0 and less than 0.25fm, it is judged as excellent.

8. A charging station risk classification method according to claim 1, characterized in that: Risk modeling specifically includes training risk warning models, marking historical fault data as labels, combining isolation forest and autoencoder anomaly detection to identify unknown abnormal patterns, and classifying the causes of operating failures into four levels: normal, low risk, medium risk, and high risk according to risk level scores.

9. A charging station risk classification method according to claim 8, characterized in that: High risk: trigger automatic power outage + push alarm to operation and maintenance personnel and supervision platform + SMS notification of abnormal information to users + pop-up window showing specific abnormal information to users; Medium risk: SMS notification of abnormal information to users + pop-up window showing specific abnormal information to users + generate work order and send it to operation and maintenance APP for processing within a limited time; Low risk: SMS notification of abnormal information to users + page pop-up window showing users specific abnormal information + inclusion in periodic inspection plan.

10. A charging station risk classification method according to claim 1, characterized in that: Through federated learning, multi-site data is aggregated to update the model, protecting privacy while improving generalization capabilities. Operation and maintenance personnel process the results and reversely annotate the system to form a "risk identification-handling-verification-optimization" closed loop, building historical risk events and handling plans into a knowledge graph to assist in decision-making in new scenarios.

Images