A photovoltaic power generation system risk management method

Abstract

The present application relates to the technical field of power system operation management, in particular to a photovoltaic power generation system risk management method, which first collects multi-dimensional operation data to establish a dynamic digital twin health record for equipment; then constructs a time-space causal risk management graph; relies on a Bayesian network to complete causal reasoning, fault root cause diagnosis and risk level determination; can identify extreme working conditions and automatically switch control plans; through cloud-edge collaboration, hierarchical response and risk active blocking are realized; finally, the model and control strategy are iteratively optimized in a closed loop combined with disposal data. The present application can accurately predict the risk evolution trend, greatly reduce false positives and false negatives in extreme environments, and realize the transition from passive alarm to active prevention. According to the actual measurement of multiple types of photovoltaic power stations, the fault diagnosis accuracy of this method is high, effectively shortens the operation and maintenance time, reduces the loss of power generation capacity and reduces the operation and maintenance cost, and has wide applicability.

Description

Technical Field

[0001] This invention relates to the field of power system operation and management technology, specifically a risk management method for photovoltaic power generation systems. Background Technology

[0002] With the deepening of the global energy transition, the proportion of photovoltaic power generation in the energy structure continues to increase. my country's installed photovoltaic power generation capacity has ranked first in the world for many consecutive years, forming a diversified development pattern including large-scale ground-mounted power plants, distributed photovoltaic systems, and residential photovoltaic systems. However, the large-scale deployment of photovoltaic power generation systems has also brought severe challenges to safe operation.

[0003] Photovoltaic power generation systems typically consist of tens or even hundreds of thousands of photovoltaic modules, combiner boxes, inverters, transformers, and other equipment, distributed over vast geographical areas and exposed to complex and ever-changing natural environments for extended periods. During operation, these systems face various risks, including microcracks in modules, hot spots, loose wiring, insulation aging, inverter failures, and transformer malfunctions. These risks not only lead to power generation losses and reduced economic efficiency but can also trigger safety accidents such as fires and electric shocks, causing serious personal injury and property damage. In recent years, in particular, with the increasing age of photovoltaic power plants, equipment aging has become a more prominent issue, and safety accidents are on the rise.

[0004] While some progress has been made in risk management technology for photovoltaic power generation systems, the following prominent issues still exist: First, most existing technologies remain at the level of anomaly detection, lacking in-depth causal reasoning capabilities. Most methods can only determine whether a device is malfunctioning, but cannot accurately explain the root cause of the anomaly. After receiving an alarm, maintenance personnel still need to spend a significant amount of time conducting on-site investigations to determine the root cause of the fault, which severely impacts fault handling efficiency. Furthermore, due to the inability to understand the causal mechanisms of risk generation, existing methods also struggle to accurately predict the future development trend and potential impact of risks.

[0005] Second, existing technologies have poor adaptability to extreme operating conditions. Most risk assessment models are trained on data under normal operating conditions. When encountering extreme weather events such as heavy rain, hail, strong winds, extreme heat, and sandstorms, the system's operating status changes drastically, leading to a large number of false alarms and missed alarms. Especially under extreme weather conditions, the probability of photovoltaic system failure increases significantly, and traditional risk management methods often fail, making it impossible to detect and handle faults in a timely manner, which can easily lead to the escalation of accidents.

[0006] Third, existing technologies lack proactive risk transmission and prevention mechanisms. Most methods can only issue alarms after a fault occurs, waiting for maintenance personnel to handle it, which is a passive risk management approach. However, photovoltaic power generation systems are highly coupled and complex systems. If a fault in one device is not isolated in time, it may quickly propagate to other related devices, triggering a chain reaction and causing wider power outages and equipment damage. Especially in unattended, remote photovoltaic power plants, passive risk management models pose serious safety hazards.

[0007] Fourth, existing technologies do not adequately consider individual differences in equipment. Photovoltaic equipment from different brands, batches, and installation conditions exhibit significant differences in performance characteristics and aging patterns. However, most existing methods use uniform evaluation models and threshold standards, failing to conduct personalized risk assessments for the individual characteristics of each device. This significantly reduces the accuracy of risk assessments.

[0008] Fifth, existing technologies lack sufficient closed-loop management capabilities. While some methods have proposed the concept of closed-loop management, most simply record the results of actions taken without effectively transforming experience into model knowledge, thus failing to achieve continuous self-optimization of the system. As the photovoltaic system operates for longer periods, the accuracy of the model gradually declines, requiring frequent manual maintenance and updates.

[0009] In recent years, although some AI-based photovoltaic fault diagnosis methods have been proposed, most of these methods focus on identifying single types of faults and have not formed a complete risk management system. Furthermore, these methods still have significant shortcomings in causal reasoning, adaptability to extreme operating conditions, and proactive intervention. Therefore, there is an urgent need to develop a new risk management method for photovoltaic power generation systems that can overcome the aforementioned deficiencies of existing technologies and achieve a shift from passive alarm to proactive prevention, and from single detection to system-wide control. Summary of the Invention

[0010] To address the problems in existing technologies, this invention provides a risk management method for photovoltaic power generation systems, which has causal reasoning decision-making capabilities, extreme condition adaptive management capabilities, and proactive risk prevention capabilities, enabling intelligent risk management throughout the entire process.

[0011] The technical solution adopted by this invention to solve its technical problem is: a risk management method for photovoltaic power generation systems, comprising the following steps: S1: Full lifecycle equipment health management and data acquisition, real-time acquisition of electrical operation data, environmental meteorological data and equipment status data at all levels of the photovoltaic system, and establishment and dynamic updating of digital twin health records for each device, including factory parameters, operating history and maintenance records; S2: Construct a spatiotemporal causal risk management map, based on the physical topology of the photovoltaic system, integrating the laws of temporal evolution and causal relationships, to form a dynamic risk management map that includes node health attributes, spatiotemporal connectivity, and risk transmission links; S3: Risk decision assessment based on causal reasoning, calculating individual health deviation based on digital twin health records, performing risk causal reasoning and root cause diagnosis through Bayesian networks, predicting risk evolution trends and economic losses, and generating risk level assessment results; S4: Extreme operating condition adaptive management strategy switching, identify extreme weather and abnormal operating conditions, automatically load the corresponding risk control model and contingency plan, and adjust the risk assessment threshold and response level; S5: Hierarchical and layered risk management and proactive intervention, with real-time risk monitoring and local risk blocking operations performed at the edge, and global risk analysis and operation and maintenance decision optimization performed in the cloud, achieving hierarchical response and proactive prevention of risks through cloud-edge collaboration; S6: Closed-loop optimization of risk control effectiveness, recording the entire process and effect data of risk disposal, continuously updating digital twin health records and risk assessment models, and iteratively optimizing risk control strategies.

[0012] Specifically, the construction of the spatiotemporal causal risk management graph in step S2 includes: constructing a basic management graph with photovoltaic equipment as nodes and electrical connections, spatial proximity, and causal influence relationships as edges; adding quantitative weights to each edge to characterize the intensity of risk influence and transmission probability between nodes; and establishing a risk evolution time series model to describe the dynamic development process of risk from generation to diffusion.

[0013] Specifically, the Bayesian network causal inference in step S3 includes: constructing a Bayesian network decision model based on historical fault data and industry expert knowledge; determining the conditional probability table of each node through parameter learning; calculating the posterior probability of each potential fault cause based on observed abnormal evidence, and outputting the fault root cause diagnosis results with priority ranking.

[0014] Specifically, the extreme working condition adaptive management in step S4 includes: establishing a contingency plan library covering extreme working conditions such as rainstorms, hail, strong winds, extreme high temperatures, sandstorms, and lightning strikes; real-time monitoring of environmental parameters and system operating status, identifying the current working condition type through multi-feature fusion; and automatically switching to the matching control mode, adjusting the risk monitoring frequency and handling procedures.

[0015] Specifically, the local risk blocking operation in step S5 includes: dynamically adjusting the output power of the risky equipment; isolating the faulty electrical circuit; optimizing the operating parameters of adjacent equipment; dispatching the energy storage system to supplement the power gap; and reporting the risk information to the power grid dispatching platform and requesting coordinated control.

[0016] Specifically, the digital twin health record in step S1 includes: basic equipment information, installation and commissioning records, full-cycle operation performance curves, maintenance records, historical fault and handling records, and the record is updated in real time with the equipment's operating status.

[0017] Specifically, the spatiotemporal causal risk management map supports the visualization of risk propagation paths, which can intuitively present the process of risk spreading from the source node to related nodes and predict the range of devices that may be affected in the future.

[0018] Specifically, the risk evolution trend prediction adopts a time-series prediction algorithm, which combines historical fault evolution patterns and current system load status to predict the risk development trend and potential power generation loss within different time windows.

[0019] Specifically, the extreme condition contingency plan library supports online iterative updates and can automatically optimize model parameters and handling procedures based on newly occurring extreme condition events and handling experience.

[0020] Specifically, the edge device is equipped with an independent security control module, which can still autonomously execute preset risk blocking strategies when cloud communication is interrupted, ensuring the basic security operation of the system.

[0021] The beneficial effects of this invention are: This invention constructs a spatiotemporal causal risk management map and introduces Bayesian network causal reasoning technology, which can not only identify abnormal states, but also accurately infer the root cause of the fault, predict the evolution trend and impact range of the risk, provide a scientific basis for operation and maintenance decisions, and significantly improve the efficiency of fault handling.

[0022] This invention designs an adaptive management mechanism for extreme working conditions and establishes a contingency plan library for various extreme working conditions. It can automatically switch risk control modes according to actual working conditions, effectively solving the problems of poor adaptability and high false alarm and false alarm rates of traditional methods under extreme working conditions, and improving the reliability of the system in complex environments.

[0023] This invention integrates proactive risk transmission blocking technology, enabling the edge to immediately implement proactive intervention measures when a risk is detected, promptly isolate faults, block the spread of risks, prevent the escalation of accidents, and realize the transformation from passive alarm to proactive prevention.

[0024] This invention establishes a full lifecycle digital twin health record for each device, enabling personalized risk assessment, fully considering individual differences of devices, and significantly improving the accuracy of risk assessment.

[0025] This invention achieves closed-loop management throughout the entire process, automatically learns from handling experience, and continuously optimizes risk assessment models and control strategies, so that the system's risk control capabilities continue to improve as the operating time increases. Attached Figure Description

[0026] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0027] Figure 1 This is a schematic diagram of the overall process of the risk management method for photovoltaic power generation system described in this invention, showing the complete execution steps from data acquisition to closed-loop management.

[0028] Figure 2 This is a schematic diagram of the spatiotemporal causal risk management graph described in this invention, illustrating the node types of the graph and the three different types of connecting edges and their relationships.

[0029] Figure 3 This is a schematic diagram of the cloud-edge collaborative risk management architecture described in this invention, illustrating the functional module division and data interaction relationship between the edge and cloud.

[0030] Figure 4 This is a schematic diagram of the extreme condition adaptive management process described in this invention, illustrating the process of identifying, switching, executing, and recovering from extreme conditions. Detailed Implementation

[0031] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0032] like Figures 1-4 As shown, the risk management method for a photovoltaic power generation system according to the present invention includes the following steps: S1. Full lifecycle equipment health management and data acquisition: Sensors and data acquisition devices deployed at various nodes of the photovoltaic system collect multi-dimensional data in real time at the module, string, inverter, and power plant levels according to preset frequencies. Specifically, module-level voltage, current, and temperature data are collected once per second; string-level electrical data is collected once per second; inverter and transformer operation data is collected twice per second; and environmental meteorological data is collected once per minute. The types of data collected include: electrical operation data, specifically DC voltage, DC current, AC voltage, AC current, active power, reactive power, power factor, frequency, and total harmonic distortion; environmental meteorological data, specifically horizontal surface irradiance, inclined surface irradiance, ambient temperature, relative humidity, wind speed, wind direction, rainfall, air pressure, and dust concentration; and equipment status data, specifically photovoltaic module backsheet temperature, inverter internal temperature, transformer top oil temperature, transformer winding temperature, equipment vibration acceleration, DC side insulation resistance, AC side insulation resistance, switch contact temperature, and circuit breaker status.

[0033] The collected raw data undergoes preprocessing. First, the 3σ criterion is used to remove obvious outliers and noisy data. Then, linear interpolation is used to complete missing data. Finally, the min-max normalization method is used to map data of different dimensions to the [0,1] interval for easier subsequent analysis and processing. Simultaneously, a unique equipment identifier is assigned to each photovoltaic device, and a digital twin health record is established and dynamically updated. The record includes: a basic equipment information database, storing static information such as equipment model, specifications, manufacturing date, supplier, technical parameters, and installation location; an operational history database, storing daily power generation, average conversion efficiency, and key performance indicator statistics since the device was put into operation; a maintenance record database, storing installation and commissioning records, daily inspection records, periodic maintenance records, fault repair records, and component replacement records; and a health status database, storing the device's current health score, risk level, and aging assessment results. The digital twin health record is automatically updated daily, and an immediate update is triggered immediately when equipment malfunctions or undergoes maintenance.

[0034] S2. Constructing a spatiotemporal causal risk management map: The three-dimensional spatiotemporal causal risk management map is constructed using a five-step method: "basic topology construction - node attribute definition - edge relationship quantification - causal correlation mining - temporal evolution modeling". The specific process is as follows: The first step is to construct the basic physical topology graph. Based on the electrical design drawings and actual geographical layout of the photovoltaic power station, a directed acyclic graph is constructed, using photovoltaic modules, photovoltaic strings, combiner boxes, inverters, transformers, and weather stations as nodes and electrical connections as the basic edges. For example, a photovoltaic module node points to its corresponding string node, a string node points to its corresponding combiner box node, a combiner box node points to its corresponding inverter node, and an inverter node points to its corresponding transformer node.

[0035] The second step is to define node attributes. Three types of attributes are added to each node: static attributes, including static information obtained from the digital twin health record such as device type, device identification code, installation location, and rated parameters; dynamic attributes, including dynamically updated information such as real-time operating data, current health score, and current risk level; and historical attributes, including historical information such as historical fault records and historical health status change curves.

[0036] The third step is to quantify edge relationships. The graph contains three types of edges, each with a corresponding weight coefficient: Electrical connection edge: The weighting coefficient is set according to the tightness of the electrical connection. The weight between components within the same string is 0.9, the weight between strings within the same inverter is 0.7, and the weight between inverters within the same transformer is 0.5. Spatial proximity edges: The weights are determined by calculating the geographical distance between devices. The weight between devices less than 10 meters apart is 0.6, the weight between devices 10-50 meters apart is 0.3, and no spatial proximity edges are established between devices greater than 50 meters apart. Causal correlation edges: Weights are determined by mining historical data; see step four for details.

[0037] The fourth step is to uncover causal relationships. Historical operating data and fault records from the power plant over the past three years are collected, and the Granger causality test is used to analyze the causal relationships between devices. For any two devices A and B, if a change in the state of device A can significantly predict a future change in the state of device B, then a causal relationship is considered to exist between A and B. The weight coefficient of the edge is calculated using the Pearson correlation coefficient, with a value ranging from [0,1]. A larger value indicates a stronger causal influence. For example, if analysis reveals that an abnormal DC-side voltage of the inverter causes current fluctuations in adjacent strings, and the Granger causality test p-value is less than 0.05 and the Pearson correlation coefficient is 0.75, then a causal relationship edge is established from the inverter node to the adjacent string node, with a weight of 0.75.

[0038] The fifth step is to establish a risk evolution time series model. Using time series analysis, the development process of historical failures is modeled, extracting typical time series characteristics of risk from its generation, development, and diffusion. A corresponding risk evolution curve is established for each failure type, describing the change in risk level over time and the transmission time of risk between different devices.

[0039] The completed spatiotemporal causal risk management graph updates the dynamic attributes of nodes every 3 minutes, recalculates the weight coefficients of edges every 24 hours, and performs a comprehensive graph structure update every quarter.

[0040] S3. Risk Decision Assessment Based on Causal Reasoning: Based on digital twin health records and spatiotemporal causal risk management maps, multi-dimensional risk assessment and decision support are conducted. The specific process is as follows: First, the individual health deviation of each device is calculated. For each device, data from its digital twin health record, representing the three consecutive months of normal operation prior to commissioning, is extracted as the training set. Principal component analysis is used to extract six core features: voltage, current, power, temperature, vibration, and insulation. The mean and standard deviation of each feature under normal operating conditions are calculated to establish an individual device health baseline model. In actual operation, the real-time collected feature data is compared with the health baseline to calculate the deviation of each feature. Feature deviation = |Real-time value - Baseline mean| / Baseline standard deviation; Then, the entropy weight method is used to determine the weight of each feature, and the overall health deviation of the device is obtained by weighted summation. When the overall deviation exceeds 0.3, an anomaly warning is triggered.

[0041] Secondly, Bayesian network causal inference and root cause diagnosis are performed. The construction process of a Bayesian network consists of two stages: structure learning and parameter learning. Structure Learning Phase: Based on historical fault data and industry expert knowledge, fault type nodes and abnormal symptom nodes are identified. Fault type nodes include over 20 common faults such as component microcracks, component hot spots, loose wiring, insulation aging, inverter overheating, inverter IGBT faults, transformer overheating, and transformer oil deterioration. Abnormal symptom nodes include over 10 abnormal phenomena such as voltage anomalies, current anomalies, power anomalies, temperature anomalies, vibration anomalies, and insulation anomalies. Preliminary causal relationships between nodes are determined using expert knowledge. Then, the K2 algorithm is used to optimize the network structure using historical fault data, resulting in the final Bayesian network structure.

[0042] Parameter learning phase: The maximum likelihood estimation method is used to learn the conditional probability table for each node using historical fault data. The conditional probability table represents the probability that a child node will take different values ​​given that the state of the parent node is known. For example, the "component hotspot" node is the parent node of the "component temperature anomaly" node, and the conditional probability table indicates that when a hotspot occurs in the component, the probability that the component temperature will be abnormal is 0.92.

[0043] When an equipment malfunction is detected, the observed abnormal symptoms are input into a Bayesian network as evidence. A connection tree algorithm is used for inference to calculate the posterior probability of each potential fault cause. The system outputs the top three most likely fault root causes in descending order of posterior probability, along with the confidence level for each root cause. For example, when a low string current and high component temperature are detected, the Bayesian network infers a posterior probability of 0.87 for "component hotspot," 0.62 for "loose wiring," and 0.43 for "component microcrack." The system then outputs the following fault root cause diagnosis results: 1. Component hotspot (confidence level 87%); 2. Loose wiring (confidence level 62%); 3. Component microcrack (confidence level 43%).

[0044] Finally, risk evolution trend prediction and loss assessment are performed. Using a Long Short-Term Memory (LSTM) network algorithm, with the equipment's operating data from the past 7 days and current root cause diagnosis results as input, the risk's development trend is predicted for the next 24 hours, 72 hours, and 7 days. Simultaneously, combining the risk transmission probability in the spatiotemporal causal risk management map, the potential equipment impact range of the risk is predicted. Based on historical failure loss data, a power generation loss assessment model is established to predict the potential power generation loss and equipment maintenance costs caused by the risk without intervention. Based on the severity, urgency, and scope of impact, the risk is classified into four levels: Normal (Level 0), Attention (Level 1), Warning (Level 2), and Hazardous (Level 3).

[0045] S4. Switching between adaptive management strategies for extreme operating conditions: Establish a contingency plan database covering various extreme working conditions to achieve adaptive switching of risk management modes. The specific process is as follows: First, define the types of extreme operating conditions and establish a contingency plan database. Extreme operating conditions are divided into two main categories: meteorological extreme conditions and operational extreme conditions. Meteorological extreme conditions include heavy rain, hail, strong winds, extreme high temperatures, extreme low temperatures, sandstorms, lightning strikes, and blizzards; operational extreme conditions include sudden rises and falls in grid voltage, abnormal grid frequency, sudden increases in charging load, and large-area shading. For each extreme operating condition, a corresponding contingency plan is established. The plan includes: a dedicated risk assessment model, specifically trained using historical data under that condition; adjusted risk threshold parameters, with alarm thresholds for each indicator corrected according to the characteristics of the condition; optimized response strategies, including risk monitoring frequency, response level, and handling procedures; and preventative measures, including preventative measures that can be taken before the extreme condition occurs.

[0046] Secondly, real-time operating condition identification is performed. Environmental meteorological parameters and system operating status are monitored in real time, and feature vectors characterizing different operating conditions are extracted. A Support Vector Machine (SVM) classification algorithm is used to match the extracted feature vectors with operating condition features in the contingency plan database to identify the current operating condition type. The accuracy rate of operating condition identification is required to reach over 95%.

[0047] Finally, the system performs a switch and recovery of the control mode. When an extreme operating condition is detected, the system automatically issues a condition switching warning, loads the corresponding extreme operating condition contingency plan within 10 seconds, and switches to the appropriate risk control mode. The switching process includes: loading a dedicated risk assessment model, updating alarm thresholds for each indicator, adjusting data collection frequency, and modifying the risk response level. During the duration of the extreme operating condition, the system reassesses the operating status every 5 minutes. When the extreme operating condition is detected to have ended, the system automatically issues a recovery warning and recovers to the normal control mode within 30 seconds. Simultaneously, all operational data and handling processes during the extreme operating condition are analyzed to summarize lessons learned and update the corresponding extreme operating condition contingency plan.

[0048] S5. Tiered and layered risk management and proactive intervention: A distributed architecture combining edge computing and cloud computing is adopted, and tiered responses and proactive interventions are implemented based on risk levels. The specific process is as follows: The edge computing infrastructure is deployed at the photovoltaic power plant site and consists of edge gateways and edge servers. Its main responsibilities include real-time data acquisition, preprocessing, localized risk monitoring, and rapid response. The edge computing infrastructure runs a lightweight risk assessment model, a quantized and compressed version of the full model trained in the cloud. This model reduces the number of parameters by more than 70% and increases inference speed by more than 5 times, enabling anomaly detection within milliseconds. Depending on the risk level, the edge computing infrastructure implements different response measures: For normal risk level (Level 0): Continue monitoring, no intervention required; For Level 1 risk: Double the data collection frequency of the device, continuously monitor its status changes, and report monitoring data to the cloud once an hour; For warning level (Level 2) risks: immediately implement preliminary risk blocking operations, including dynamically reducing the output power of the risky equipment to 50% of its rated power, shutting down unnecessary functions of the risky equipment, and increasing the monitoring frequency of adjacent equipment; at the same time, upload detailed abnormal data to the cloud and request global analysis; For risk level 3: Immediately perform emergency shutdown operations, including cutting off the power supply to the faulty circuit, isolating the faulty area, and activating backup equipment; at the same time, notify maintenance personnel through various means such as audible and visual alarms, SMS, and APP push.

[0049] Deployed in a remote data center and equipped with high-performance computing and storage resources, the cloud platform is primarily responsible for global risk analysis, model training, and operational decision optimization. Upon receiving risk information reported from the edge devices, the cloud platform first performs global correlation analysis using a spatiotemporal causal risk management graph to trace the root cause of the risk and analyze its potential propagation paths and impact scope. Then, combining the root cause diagnosis results from Bayesian networks and risk evolution trend predictions, it generates the optimal operational response plan. This plan includes: fault location, fault type, confidence level, recommended inspection items, handling steps, required tools and spare parts, and estimated handling time. The cloud platform automatically generates operational work orders and dispatches them to the most suitable personnel based on their location, skills, and current work status. Simultaneously, the cloud platform fine-tunes the risk assessment model daily using the previous day's global data and performs a comprehensive model update weekly, distributing the updated model parameters to all edge devices.

[0050] S6. Closed-loop optimization of control effectiveness: Establish a closed-loop management mechanism covering the entire process to achieve continuous self-optimization of the system. The specific process is as follows: First, establish a full lifecycle archive for risk events. For each risk event, record complete information from discovery, diagnosis, dispatch, handling, to acceptance, including: the time, location, level, anomaly characteristics, diagnosis results, dispatch time, maintenance personnel information, arrival time, handling measures, handling time, handling results, acceptance status, relevant operational data, and on-site photos. All information is stored in association with the corresponding digital twin health record of the equipment.

[0051] Secondly, the system evaluates the effectiveness of the response and updates the model. After a risk event is resolved, the system automatically compares the actual cause of the failure with the predicted cause, calculating the root cause diagnosis accuracy. It also compares the actual response effect with the predicted impact to assess the accuracy of the risk evolution prediction model. If the diagnosis accuracy is below 80% or the prediction error exceeds 30%, the model is updated immediately. The system retrains the Bayesian network and the risk evolution prediction model using new failure data, updating the model parameters.

[0052] Finally, the knowledge base is updated and strategies are optimized. Effective handling measures and experiences are extracted from successful cases, transformed into structured rules, and added to the system's knowledge base. Regular statistical analysis of all risk events is conducted to identify high-frequency failure types and weaknesses, optimizing risk management strategies. For example, if a certain type of failure is found to be more frequent in a specific season, the corresponding risk threshold and monitoring frequency for that season are adjusted. Through continuous learning and optimization, the system's risk management capabilities are continuously improved.

[0053] Step S2, which involves constructing a spatiotemporal causal risk management graph, also includes: supporting the visualization of risk propagation paths, which can intuitively present the process of risk spreading from the source node to related nodes in the form of animation, and marking nodes with different risk levels on the graph with different colors; and learning the feature representation of the risk graph through graph neural network algorithms to further improve the accuracy of risk identification and prediction.

[0054] In step S3, the individual health baseline model of the equipment is fully updated every quarter. When the equipment undergoes major repairs or component replacements, a new health baseline model is immediately re-established.

[0055] The risk transmission blocking operation in step S5 also includes: automatically starting the backup power supply or energy storage system when a serious fault that may cause systemic risk is detected to ensure the power supply of critical equipment; sending a power adjustment request to the grid dispatch system to cooperate with the grid in frequency and voltage regulation; and maintaining the normal operation of non-faulty areas as much as possible to minimize power generation loss, provided that safety is ensured.

[0056] Example 1 discloses a risk management method for a photovoltaic power generation system, applied to a 150MW ground-mounted photovoltaic power station. The power station is located in a desert region in Northwest China, covering an area of ​​approximately 4 square kilometers. It consists of 600 photovoltaic strings, equipped with 30 centralized inverters and 6 main transformers. The region has a dry climate, large diurnal temperature variations, and abundant sunshine, but frequently experiences extreme weather events such as strong sandstorms, extreme heat, and hail.

[0057] The risk management method for photovoltaic power generation systems in this embodiment specifically includes the following steps: The first step is full lifecycle equipment health management and data acquisition. Miniature temperature sensors and voltage / current acquisition modules are installed on the back of each photovoltaic module. String-level current / voltage acquisition devices, insulation resistance monitoring devices, and switch status monitoring devices are installed at the combiner boxes of each string. High-precision temperature sensors, vibration sensors, electrical parameter acquisition devices, and partial discharge monitoring devices are installed at each inverter and transformer. Five weather stations are evenly distributed throughout the power station to monitor irradiance, ambient temperature, humidity, wind speed, wind direction, rainfall, air pressure, and dust concentration in real time. All acquisition devices are connected to the edge gateway via industrial Ethernet. Simultaneously, a unique digital twin health record is established for each photovoltaic module, combiner box, inverter, and transformer in the power station. The record includes the equipment model, specifications, manufacturing date, supplier information, installation and commissioning records, maintenance records, key performance indicator change curves during operation, and historical fault records and handling results. The system automatically updates the digital twin health record daily, recording changes in the equipment's operating status.

[0058] The second step involves constructing a spatiotemporal causal risk management map. First, a basic topology map is built based on the power plant's electrical design drawings and actual geographical layout. Nodes are defined by 600 photovoltaic strings, 30 inverters, 6 transformers, and 5 weather stations, with edges representing electrical connections and spatial proximity. Then, based on the power plant's historical fault data and maintenance records from the past five years, causal relationships between devices are analyzed. For example, analysis reveals that abnormal DC-side voltage in inverters can cause current fluctuations in adjacent strings, and increased transformer oil temperature increases the probability of inverter overheating. These causal relationships are added to the map, forming causal edges, and the weight coefficient of each edge is calculated based on historical data to quantify the influence intensity and risk transmission probability between nodes. Next, a risk evolution time-series model is established to describe the dynamic process of risk generation, development, and diffusion. Finally, all fault information that has occurred in the power plant's history is marked on the risk map, including the time, location, cause, development process, and handling results of the fault. The resulting spatiotemporal causal risk management map is updated every 3 minutes, providing real-time visualization of the entire power plant's risk distribution, causal relationships, and evolution trends.

[0059] The third step is risk decision assessment based on causal reasoning. For each device, an individual health baseline model is established based on historical data in its digital twin health profile. The individual health baseline model reflects the performance characteristics and fluctuation patterns of the device under normal operating conditions. In actual operation, real-time collected device operating data is compared with the individual health baseline to calculate the deviation of multi-dimensional features. When the deviation exceeds a preset threshold, a risk warning is triggered. Simultaneously, the Bayesian network causal reasoning module is activated. This Bayesian network is built based on the power plant's fault data over the past 5 years and industry expert knowledge, including various fault types such as component faults, combiner box faults, inverter faults, and transformer faults, as well as their related abnormal symptoms. Observed abnormal features are input into the Bayesian network as evidence to calculate the posterior probability of each potential fault cause. The system outputs the top 3 fault causes with the highest posterior probabilities and their confidence levels as the root cause diagnosis results. In addition, the system uses a long short-term memory network algorithm, combined with historical fault evolution patterns and the current system state, to predict the development of risks in the next 24 hours, 72 hours, and 7 days, the range of equipment that may be affected, and the potential power generation loss.

[0060] The fourth step involves adaptive management strategy switching for extreme operating conditions. The system has established a contingency plan database covering various extreme operating conditions, including severe sandstorms, extreme high temperatures, hail, strong winds, and lightning strikes. Each contingency plan database contains a risk assessment model specifically trained for that condition, adjusted threshold parameters, and optimized handling procedures. For example, during a severe sandstorm, because sand cover can cause a decrease in component power, the system will appropriately relax the power deviation threshold while increasing the detection sensitivity for component temperature and current imbalance. The system monitors weather station data and system operating status in real time, accurately determining the current operating condition type through a multi-feature fusion operating condition identification algorithm. When an extreme operating condition is identified, the system automatically switches to the corresponding risk control mode, loading the appropriate model and parameters. During extreme operating conditions, the system increases the data acquisition frequency to improve the accuracy of risk monitoring. After the extreme operating condition ends, the system automatically reverts to normal control mode and conducts in-depth analysis of the operating data during the extreme operating condition period, summarizing lessons learned and updating the extreme operating condition contingency plan database.

[0061] The fifth step involves tiered and layered risk management and proactive intervention. Six edge servers are deployed at the power plant site, each managing 100 strings and 5 inverters. Lightweight risk assessment models and independent safety control modules run on the edge servers, enabling anomaly detection and response within milliseconds. For low-level risks with low overall deviation, the edge servers record data locally and increase the data acquisition frequency of the device to continuously monitor its status changes. For medium-level warning risks, the edge servers immediately execute preliminary risk blocking operations, such as reducing the output power of the relevant inverters and limiting power output in the fault area. For high-level hazardous risks, the edge servers immediately execute emergency blocking operations, such as cutting off power to the faulty circuit, isolating the fault area, and preventing the risk from spreading. Simultaneously, the edge servers upload risk information and detailed data to the cloud. The cloud is deployed in a remote data center, equipped with high-performance computing and storage resources. Upon receiving risk information reported by the edge, the cloud immediately initiates global causal analysis, tracing the root cause of the risk through a spatiotemporal causal risk management graph, analyzing the possible propagation paths and impact range of the risk, generating the optimal handling plan, and distributing it to the edge for execution. Meanwhile, the cloud will automatically generate maintenance work orders and assign them to the nearest maintenance personnel.

[0062] Step 6: Closed-Loop Optimization of Control Effectiveness. The system fully records the entire process of each risk event, including the time, location, level, anomaly characteristics, diagnostic results, handling measures, handling time, handling effects, and related operational data. After handling the fault, maintenance personnel fill in a detailed fault report in the system, including the actual fault cause, handling process, and experience summary. The system compares the actual fault cause with the predicted fault cause, analyzes the reasons for the differences, and updates the parameters of the Bayesian network. At the same time, it compares the handling results with the predicted impact to optimize the risk evolution prediction model. The system also extracts rules from successful handling cases and adds them to the knowledge base to continuously optimize risk control strategies. Through this continuous learning and optimization, the system's risk control capabilities will continuously improve with the increase of operating time.

[0063] This embodiment underwent an 8-month trial operation at the 150MW photovoltaic power station. The trial results showed that the fault root cause diagnosis accuracy of this invention reached over 94%, capable of detecting progressive faults such as component aging and insulation degradation 5-10 days in advance, and potential faults in critical equipment such as inverters and transformers 48 hours in advance. During multiple severe sandstorms and extreme heat events, the system's false alarm rate was reduced by over 90% compared to traditional methods, with no missed alarms. Through proactive risk propagation blocking, three faults that could have led to large-scale power outages were successfully avoided. Compared to traditional methods, the average fault handling time of the power station was reduced by 70%, operation and maintenance costs were reduced by 40%, and power generation loss was reduced by 3.5%.

[0064] Compared to Example 1, this comparative example uses a traditional fixed threshold alarm method for risk management of the same 150MW photovoltaic power plant. Fixed alarm thresholds are set for various operational indicators, and an alarm signal is issued when an indicator exceeds the threshold. A centralized data processing architecture is adopted, with all data uploaded to a central server for processing, and a data scan and alarm determination are performed every 30 minutes.

[0065] During the 8-month controlled trial, the traditional method issued 2156 alarm signals, of which 1723 were false alarms, resulting in a false alarm rate as high as 79.9%. The number of false alarms increased dramatically, particularly during severe sandstorms and extreme heat, reaching a peak of 126 in a single day, severely disrupting normal operation and maintenance. Simultaneously, 47 actual faults went undetected, resulting in a missed alarm rate of 31.5%. Most faults were only discovered after they occurred, with an average fault response time of 5.6 hours. During the trial, two cascading failures due to untimely fault isolation resulted in significant power generation losses and equipment damage. Overall, the power generation loss due to faults during the trial was 8.9%, and the operation and maintenance costs were 1.7 times that of Example 1.

[0066] A comparison of Example 1 and Comparative Example 1 shows that the photovoltaic power generation system risk management method of the present invention is significantly better than the traditional fixed threshold alarm method in terms of fault diagnosis accuracy, adaptability to extreme operating conditions, proactive intervention capability, and operation and maintenance efficiency. It can effectively improve the safe and stable operation level and economic benefits of photovoltaic power plants.

[0067] Example 2 discloses a risk management method for a photovoltaic power generation system, applied to an 8MW residential distributed photovoltaic cluster. This cluster comprises 240 residential photovoltaic systems distributed across three adjacent villages, with a total installed capacity of 8MW. Each residential system is equipped with one string inverter, installed on the roof of a residential building. The region has a subtropical monsoon climate, characterized by hot and rainy summers, and frequent typhoons and torrential rains.

[0068] The basic steps of the photovoltaic power generation system risk management method in this embodiment are the same as those in Embodiment 1, except that: In terms of full lifecycle equipment health management and data collection, considering the large number and wide distribution of residential photovoltaic systems, a low-cost wireless data acquisition solution was adopted. Each inverter communicates directly with the cloud platform via a 4G network to upload operational data. Simultaneously, a small weather station is installed in each village to monitor local environmental meteorological data. An independent digital twin health record is established for each household's photovoltaic system, recording installation information, operational history, power generation data, and maintenance records. The system also connects to the power grid's electricity consumption data, enabling analysis of the matching between photovoltaic power generation and user electricity consumption.

[0069] In constructing a spatiotemporal causal risk management map, a three-tiered risk management map was built: household-level, village-level, and cluster-level. The household-level risk map displays the device relationships within a single household system; the village-level risk map displays the relationships between all household systems within the same village; and the cluster-level risk map displays the overall risk distribution of the entire cluster. Since there are no direct electrical connections between household systems, the main considerations are spatial proximity and the consistency of meteorological influences. For example, household systems within the same village are affected by the same meteorological conditions, and when one household system experiences an anomaly, other household systems may also face similar risks.

[0070] In terms of risk decision assessment based on causal reasoning, the Bayesian network structure has been optimized to suit the characteristics of residential photovoltaic systems, focusing on common fault types in residential systems such as component failure, inverter failure, and wiring failure. Simultaneously, the system has enhanced its ability to analyze anomalies in power generation, distinguishing between power generation declines caused by weather conditions and those caused by equipment failure. The system also regularly evaluates the power generation performance of each household system, identifying systems with low power generation efficiency and providing optimization suggestions.

[0071] Regarding the adaptive management strategy switching for extreme operating conditions, the system has significantly enhanced its ability to respond to typhoons and heavy rainstorms. A dedicated contingency plan database for typhoons and heavy rainstorms has been established, enabling proactive risk prevention preparations based on weather forecasts. Before a typhoon arrives, the system automatically sends warning messages to users, advising them to disconnect the power supply to their photovoltaic systems. During heavy rainstorms, the system increases its sensitivity to detecting insulation resistance and leakage current to prevent electric shock accidents.

[0072] For tiered risk management and proactive intervention, an architecture combining centralized cloud control and edge intelligence is adopted. Since the inverters in the residential system possess certain computing and control capabilities, some lightweight risk detection and control logic is deployed locally on the inverter. When the inverter detects a serious fault, it can automatically cut off the output to protect the equipment and user safety. The cloud platform is responsible for overall risk monitoring, analysis, and scheduling, and uniformly manages the operation and maintenance of the entire cluster.

[0073] This embodiment underwent a 6-month trial operation on an 8MW residential distributed photovoltaic (PV) cluster. The trial results show that the present invention can effectively adapt to the characteristics of residential distributed PV clusters, achieving unified risk management for a large number of dispersed residential systems. The system's fault diagnosis accuracy reached over 91%, and the false alarm rate was reduced by 85% compared to traditional methods. Through early warning and proactive intervention, several potential safety accidents caused by typhoons and rainstorms were successfully avoided. Compared to traditional methods, the average fault handling time of the cluster was reduced by 65%, operation and maintenance costs were reduced by 35%, and overall power generation increased by 2.3%.

[0074] Compared to Example 2, this comparative example uses a mainstream residential photovoltaic monitoring platform to manage the risk of the same 8MW residential distributed photovoltaic cluster. This platform has basic data acquisition and alarm functions, employs a fixed threshold alarm method, and can achieve remote monitoring and mobile app alarm functionality.

[0075] During the six-month controlled trial, the platform issued 872 alarm signals, of which 698 were false alarms, resulting in a false alarm rate of 80.0%. The number of false alarms increased significantly, particularly during rainy weather and typhoons, overwhelming maintenance personnel. Simultaneously, 23 actual faults went undetected, resulting in a missed alarm rate of 28.4%. Most faults were only addressed after users reported a significant drop in power generation, with an average fault response time of 7.2 hours. During the trial, one fire occurred due to a delayed detection of an inverter fault, causing property damage. Overall, the power generation loss due to faults during the trial was 6.8%, and the maintenance cost was 1.5 times that of Example 2.

[0076] The comparison between Example 2 and Comparative Example 2 shows that the risk management method for photovoltaic power generation systems of the present invention is also applicable to household distributed photovoltaic clusters, and can effectively solve the problems of dispersed, numerous, and difficult-to-manage household photovoltaic systems, thus having wide applicability.

[0077] Example 3 discloses a risk management method for a photovoltaic power generation system, applied to a 20MW integrated photovoltaic-storage-charging power station. The power station includes a 15MW photovoltaic power generation system, a 3MW / 6MWh energy storage system, and 10 120kW DC fast charging piles. Located in a highway service area, the power station primarily provides charging services for electric vehicles while also supplying power to the grid.

[0078] The basic steps of the photovoltaic power generation system risk management method in this embodiment are the same as those in Embodiments 1 and 2, except that: In terms of full lifecycle equipment health management and data collection, in addition to collecting data from photovoltaic power generation systems, data such as battery voltage, current, temperature, SOC, and SOH from energy storage systems are also collected, as well as data on charging power, charging time, and fault status of charging piles. Digital twin health records are established for photovoltaic power generation systems, energy storage systems, and charging piles, and the correlation between the three is established.

[0079] In constructing a spatiotemporal causal risk management map, a comprehensive risk management map was built, encompassing three subsystems: photovoltaic power generation, energy storage, and charging. This map includes not only the equipment relationships within each subsystem but also the energy interaction and risk transmission relationships between them. For example, fluctuations in photovoltaic power generation can affect the charging and discharging power of the energy storage system, and failures in the energy storage system can affect the power supply reliability of charging piles. By analyzing the causal relationships between subsystems, risk transmission paths across subsystems can be identified.

[0080] In terms of risk decision-making assessment based on causal reasoning, a joint risk assessment model for integrated photovoltaic-storage-charging systems has been established. This model can comprehensively consider the operating status of the three subsystems—photovoltaic power generation, energy storage, and charging—to assess the overall risk level of the entire power station. For example, when a failure in the photovoltaic power generation system leads to a decrease in power, the system will assess the impact on the charging and discharging plans of the energy storage system and the power supply capacity of the charging piles, and provide corresponding adjustment suggestions.

[0081] Regarding the adaptive management strategy switching under extreme operating conditions, in addition to considering meteorological factors, extreme fluctuations in charging load are also taken into account. A peak charging condition model has been established to cope with surges in charging demand during holidays and other periods. During peak charging periods, the system optimizes the energy dispatching of photovoltaic power generation, energy storage, and charging, prioritizing charging demand while avoiding equipment overload.

[0082] In terms of tiered and layered risk management and proactive intervention, the system achieves coordinated risk management across the photovoltaic, energy storage, and charging systems. When the photovoltaic power generation system fails, the system automatically adjusts the discharge power of the energy storage system to supplement the power shortage and ensure the normal power supply to the charging piles. When the energy storage system fails, the system adjusts the output power of the photovoltaic power generation system and limits the charging power of the charging piles to prevent system overload. Through the coordinated operation of the subsystems, the impact of failures on the normal operation of the power station can be minimized.

[0083] This embodiment underwent a 5-month trial operation at a 20MW integrated photovoltaic-storage-charging power station. The trial results demonstrate that the present invention can effectively achieve comprehensive risk management of the integrated photovoltaic-storage-charging system. The overall risk identification accuracy rate of the system reached over 93%, accurately identifying risk transmission across subsystems. Through coordinated risk management, the continuous and stable operation of the power station was ensured during multiple photovoltaic power generation system failures and peak charging periods. Compared with traditional independent subsystem management methods, the power station's downtime was reduced by 75%, operation and maintenance costs were reduced by 38%, and overall operational efficiency was improved by 15%.

[0084] Compared to Example 3, this comparative example adopts a subsystem independent management and control model, with the photovoltaic power generation system, energy storage system, and charging pile each using their own monitoring systems for risk management. There is no data exchange or coordinated control between the three systems; they operate independently.

[0085] During the five-month control test, the three systems experienced a total of 32 failures, eight of which had cross-subsystem impacts. For example, one inverter failure in the photovoltaic power generation system caused DC bus voltage fluctuations, which in turn affected the normal charging and discharging of the energy storage system, ultimately causing some charging piles to stop working. Due to the lack of coordination among the three systems, the fault handling time was long, with an average fault response time of 6.8 hours. During one peak charging period, a sudden drop in photovoltaic power generation, coupled with the energy storage system's failure to discharge and replenish in time, caused multiple charging piles to stop working due to insufficient power supply, resulting in user complaints. Throughout the test, the combined power generation loss and charging revenue loss due to failures amounted to 7.2% of the power station's total revenue, and the operation and maintenance costs were 1.6 times that of Example 3.

[0086] The comparison between Example 3 and Comparative Example 3 shows that the photovoltaic power generation system risk management method of the present invention can be effectively applied to the photovoltaic-storage-charging integrated system, realize the collaborative risk management across subsystems, and significantly improve the operational reliability and economic benefits of the integrated power station.

[0087] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of protection claimed by the present invention. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A risk management method for a photovoltaic power generation system, characterized in that, Includes the following steps: S1: Full lifecycle equipment health management and data acquisition, real-time acquisition of electrical operation data, environmental meteorological data and equipment status data at all levels of the photovoltaic system, and establishment and dynamic updating of digital twin health records for each device, including factory parameters, operating history and maintenance records; S2: Construct a spatiotemporal causal risk management map, based on the physical topology of the photovoltaic system, integrating the laws of temporal evolution and causal relationships, to form a dynamic risk management map that includes node health attributes, spatiotemporal connectivity, and risk transmission links; S3: Risk decision assessment based on causal reasoning, calculating individual health deviation based on digital twin health records, performing risk causal reasoning and root cause diagnosis through Bayesian networks, predicting risk evolution trends and economic losses, and generating risk level assessment results; S4: Extreme operating condition adaptive management strategy switching, identify extreme weather and abnormal operating conditions, automatically load the corresponding risk control model and contingency plan, and adjust the risk assessment threshold and response level; S5: Hierarchical and layered risk management and proactive intervention, with real-time risk monitoring and local risk blocking operations performed at the edge, and global risk analysis and operation and maintenance decision optimization performed in the cloud, achieving hierarchical response and proactive prevention of risks through cloud-edge collaboration; S6: Closed-loop optimization of risk control effectiveness, recording the entire process and effect data of risk disposal, continuously updating digital twin health records and risk assessment models, and iteratively optimizing risk control strategies.

2. The method for risk management of a photovoltaic power generation system according to claim 1, characterized in that: The construction of the spatiotemporal causal risk management graph in step S2 specifically includes: constructing a basic management graph with photovoltaic equipment as nodes and electrical connections, spatial proximity, and causal influence relationships as edges; adding quantitative weights to each edge to characterize the intensity of risk influence and transmission probability between nodes; and establishing a risk evolution time series model to describe the dynamic development process of risk from generation to diffusion.

3. The method for risk management of a photovoltaic power generation system according to claim 1, characterized in that: The Bayesian network causal inference in step S3 specifically includes: constructing a Bayesian network decision model based on historical fault data and industry expert knowledge; determining the conditional probability table of each node through parameter learning; calculating the posterior probability of each potential fault cause based on observed abnormal evidence, and outputting the fault root cause diagnosis results with priority ranking.

4. The method for risk management of a photovoltaic power generation system according to claim 1, characterized in that: The extreme condition adaptive management in step S4 specifically includes: establishing a contingency plan library covering extreme conditions such as rainstorms, hail, strong winds, extreme high temperatures, sandstorms, and lightning strikes; real-time monitoring of environmental parameters and system operating status, identifying the current condition type through multi-feature fusion; and automatically switching to the matching control mode, adjusting the risk monitoring frequency and handling procedures.

5. The method for risk management of a photovoltaic power generation system according to claim 1, characterized in that: The local risk blocking operation in step S5 specifically includes: dynamically adjusting the output power of the risky equipment; isolating the faulty electrical circuit; optimizing the operating parameters of adjacent equipment; dispatching the energy storage system to supplement the power gap; and reporting the risk information to the power grid dispatching platform and requesting coordinated control.

6. The method for risk management of a photovoltaic power generation system according to claim 1, characterized in that: The digital twin health record in step S1 specifically includes: basic equipment information, installation and commissioning records, full-cycle operation performance curves, maintenance records, historical fault and handling records, and the record is updated in real time with the equipment's operating status.

7. The method for risk management of a photovoltaic power generation system according to claim 2, characterized in that: The spatiotemporal causal risk management map supports the visualization of risk propagation paths, which can intuitively present the process of risk spreading from the source node to related nodes and predict the range of devices that may be affected in the future.

8. The method for risk management of a photovoltaic power generation system according to claim 3, characterized in that: The risk evolution trend prediction adopts a time-series prediction algorithm, which combines historical fault evolution patterns and current system load status to predict the risk development trend and potential power generation loss within different time windows.

9. The method for risk management of a photovoltaic power generation system according to claim 4, characterized in that: The extreme condition contingency plan library supports online iterative updates and can automatically optimize model parameters and handling procedures based on newly occurring extreme condition events and handling experience.

10. A risk management method for a photovoltaic power generation system according to claim 5, characterized in that: The edge device is equipped with an independent security control module, which can still autonomously execute preset risk blocking strategies when cloud communication is interrupted, ensuring the basic security operation of the system.

Images