Intelligent analysis and control system for safety risks of new energy station

Abstract

The application discloses a new energy station safety risk intelligent analysis and management and control system, and belongs to the technical field of intelligent operation and maintenance and safety management and control of new energy power stations.In the application, three-level intelligent management is realized by isolated forest, self-adaptive wavelet denoising and LSTM time sequence filling in the holographic perception and data management layer, and a high-quality data pedestal is formed;in the double-drive intelligent analysis engine layer, millisecond-level state estimation is carried out by fusing a physical mechanism in an inner PINN small model, and combined with CUSUM and extreme value theory dynamic threshold, small drift detection is realized, and combined with a graph attention network, root cause tracing is realized by an outer industry large model;root cause confirmation and operation and maintenance strategy automatic generation are realized by a multi-agent collaborative decision layer;and through reinforcement learning online optimization of threshold and search parameters, closed-loop evolution is realized in the global management and control execution layer.The application can identify early equipment degradation tens of hours in advance, realize interpretable early warning and automatic decision, and improve station safety and operation and maintenance efficiency.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance and safety management technology for new energy power plants, specifically to a system and method that integrates physical mechanisms and artificial intelligence for early risk warning, root cause analysis and intelligent decision-making for new energy power plants. Background Technology

[0002] Currently, the installed capacity of wind power, photovoltaic, and other new energy power plants continues to expand, the integration of internal equipment is constantly increasing, and the operating conditions of the systems are becoming increasingly complex, significantly increasing the pressure on equipment safety management and stable operation. At present, existing technologies in the field of intelligent monitoring and maintenance of new energy power plants still have many prominent shortcomings, making it difficult to adapt to the refined management and control needs of large-scale new energy power plants. Specific problems are as follows: First, fault early warning has a significant lag. Traditional monitoring methods mostly use fixed threshold triggering alarm modes, which are weak in identifying gradual and slowly changing equipment degradation problems such as micro-short circuits in energy storage batteries and early cracks in wind turbine blades. They cannot capture early abnormal characteristics of equipment, resulting in a large number of monitoring and sensing blind spots, making it difficult to achieve early fault prediction.

[0003] Secondly, the false alarm rate in monitoring alarms is too high. At present, mainstream pure data-driven artificial intelligence models lack industry physical mechanism constraints and are uninterpretable black box models. The modeling process does not take into account the underlying physical laws such as the law of conservation of energy, electrical circuit topology, and equipment operating mechanical characteristics. The monitoring and analysis results often violate the actual operating logic, resulting in insufficient alarm accuracy and inability to accurately trace the root cause of abnormal problems.

[0004] Third, fault handling is highly dependent on manual operation and maintenance. The existing monitoring system can only push abnormal alarms. Subsequent fault root cause investigation, risk assessment, and emergency response plan formulation all rely on the professional experience of front-line operation and maintenance personnel. The response efficiency is low, and human judgment errors can easily induce secondary equipment failures and safety accidents.

[0005] Although some studies have introduced physical information neural networks (PINN) into the condition monitoring of power electronic equipment, the applications are limited to the level of estimating the operating status of a single device. They have not yet established a fully integrated intelligent operation and maintenance link that includes early detection of equipment anomalies, accurate fault assessment, intelligent policy issuance, and closed-loop execution on site.

[0006] In summary, there is an urgent need in the industry to develop an integrated intelligent operation and maintenance management system that deeply integrates the underlying physical operation mechanism of equipment with intelligent algorithm models, enabling precise early warning of equipment failures, autonomous judgment of the root causes of anomalies, and automatic output of operation and maintenance strategies. Summary of the Invention

[0007] I. Purpose of the Invention The present invention aims to provide an intelligent analysis and control system for safety risks of new energy power stations, in order to solve the core technical problem that existing technologies cannot achieve interpretable and accurate early warning and automated decision-making in the "sub-health" stage of equipment.

[0008] II. Technical Solution This invention constructs a technical chain of "holographic perception → dual-drive analysis → multi-agent decision-making → closed-loop control", characterized by including the following levels.

[0009] (I) Holographic Perception and Data Governance Layer It is responsible for collecting heterogeneous data from multiple sources across the entire domain and forming a high-quality data foundation through a three-level intelligent governance mechanism.

[0010] Multi-source data acquisition: Real-time aggregation of electrical quantities (current, voltage, power), environmental quantities (irradiance, temperature, wind speed), physical state quantities (soundprint, thermal image), and network security data (traffic, logs).

[0011] Three-level intelligent governance: a) Anomaly detection: The Isolation Forest algorithm is used to remove outliers caused by sensor jumps online. The number of trees in the Isolation Forest is 100, the subsampling size is 256, and the anomaly score threshold is set to 0.6.

[0012] b) Noise filtering: Adaptive wavelet thresholding is used for noise reduction. For noisy signals... After discrete wavelet transform (using db4 wavelet, decomposition level 10 ... ) to obtain wavelet coefficients ,in To decompose the scale, For translation indexing. Processed using a soft threshold function: in , For the first Layer noise standard deviation (estimated by median absolute deviation). The length of the wavelet coefficients in this layer. For symbolic functions. Reconstruction. Obtain the denoised signal .Signal As input to the next-level "missing value repair" module, the data continues to be processed. After three levels of governance (anomaly removal, noise filtering, and missing value imputation), the data forms a unified high-quality data base for use by the subsequent dual-drive analysis engine (PINN and large model).

[0013] c) Missing value repair: The signal after denoising in the previous step... As input to this module, missing values ​​are imputed. An LSTM model based on a time attention mechanism is used. Let the time series... Define missing mask (1 indicates missing). For each position to be filled... The model inputs include: 1. Observations within the historical window (If there are still missing values, use temporary padding values ​​such as 0 or forward padding). 2. Length of consecutive missing items (A positive integer representing the number of consecutive missing sampling points from the previous valid value to the current time).

[0014] 3. LSTM output predicted value The model is trained using supervised learning: missing patterns are randomly simulated from complete, unmissing historical data (randomly set). ), at this point the true value Given that the loss function is: , equivalent to ,in These are the true, complete values ​​known during the training phase. During online inference, the model only outputs... Losses are not calculated.

[0015] The LSTM hidden layer has a dimension of 128, and the number of attention heads is 4. The input feature vector explicitly contains... This allows the model to adaptively adjust the imputation strategy based on the duration of the missing data (for example, short missing data can refer to neighboring points, while long missing data needs to rely on periodic patterns).

[0016] (II) Dual-Driven Intelligent Analysis Engine Layer (1) Inner layer: Real-time anomaly capture of PINN small model Millisecond-level state estimation is achieved using a Physical Information Neural Network (PINN). Taking a photovoltaic string as an example, a neural network following a photovoltaic single-diode model is constructed. Its loss function is: in The weight coefficients for each component are adaptively updated using the GradNorm algorithm in each training cycle (e.g., daily), and do not participate in subsequent reinforcement learning optimization; initial recommendation values. The definitions of each loss term are as follows.

[0017] 1. Data error items: in This refers to the number of samples within the batch. Predict the current (A) for PINN. The measured current is (A).

[0018] Physical residual term: The residual function of the single-diode model of a photovoltaic cell is: 2. The physical residual loss is: in The measured voltage is (V). For the learnable physical parameters of PINN, This is the thermal voltage (calculated from the sensor temperature and not included in gradient updates).

[0019] 3. Regularization term: The trained PINN serves as a digital twin of device health. The system calculates the multidimensional residual vector in real time. An adaptive cumulative sum (CUSUM) algorithm is used to detect small drifts in the residuals. For any residual component (e.g., current residual)... ), and denot its time series as ,in This indicates that the component is at time [time]. The residual value. The following is a simplified representation, using... The first component representing this component For each sample, calculate the cumulative sum of positive and negative samples: in , This is the standard deviation of the residual component. Determination: Before system commissioning, collect data from 72 consecutive hours of fault-free operation of the equipment, and calculate the standard deviation of each residual component as... Initial value; recalculated every 7 days during operation based on a recent normal data window. Use a value between 0.5 and 1.5 (typical value 1.0). When or Exceeding the dynamic threshold Anomalies are detected in real time.

[0020] The dynamic threshold Automatic tuning is based on extreme value theory and decoupled from real-time detection: it is updated offline every 24 hours. For the residual component currently being processed (e.g., current residual), the sequence of normal operating condition observations from the most recent 7 days is used. ,in For this component, the first The values ​​at each sampling time point. For each residual component, perform the following steps, with the update steps as follows: S1 extracts samples exceeding the threshold. Among them, the threshold Take the 95th percentile of the sequence.

[0021] S2 is fitted to the tail using a generalized Pareto distribution: The shape parameters are obtained through maximum likelihood estimation. and scale parameters .

[0022] S3 sets the expected false alarm rate Calculate the threshold: ,in The total number of samples, This represents the number of samples exceeding the threshold.

[0023] S4 will calculate As the reference threshold for this component In actual operation, the reinforcement learning phase will be... Adjusting the multiplier factor within the range That is, the actual threshold of this component. Each residual component maintains its own independent maintenance. .

[0024] (2) Outer layer: Deep tracing and early warning of industry big model Extract the trigger time window after CUSUM triggers an exception. The multidimensional residual statistical characteristics (mean, variance, skewness, maximum drift rate, etc.) within the data form the feature vector. Search the Top- case database Similar cases.

[0025] The current device's real-time graph sub-graph With the retrieved case map Merging. Before merging, a graph node mapping function is used. By projecting the features of historical case nodes onto the feature space of the current device, a fused graph is obtained. .

[0026] Will The data is fed into a graph attention network (GAT) to calculate the strength of the association between nodes: in: The node feature vector is composed of residual components mapped through a fully connected layer, device type encoding, and operating condition concatenation. and Learnable parameters (Glorot initialization); For splicing operations, It is the set of neighboring nodes.

[0027] The node representation is updated using multi-head attention (8 heads), and the probability distribution of each physical root cause is output.

[0028] Outer-layer industry-specific large-scale model architecture: A generative large language model based on a Transformer decoder (7B parameters, 32 layers, 4096 hidden layers, 32 attention heads). Pre-trained on a general corpus, fine-tuned with power industry instructions and adapted to LoRA, and deployed using 4-bit quantization and a vLLM inference engine.

[0029] Large model input processing: This involves processing the root cause probability distribution of the GAT output, the current residual summary statistics, and... The textual descriptions (generated via a graph serializer using depth-first traversal to create linear text of entities, relationships, and attributes) are concatenated into prompt words. The model generates natural language reports (root causes, confidence scores, evolution predictions, and verification measures).

[0030] (3) Multi-agent collaborative decision-making layer Root cause diagnosis agent: If the maximum root cause probability is below a threshold (Recommended value: 0.78) If this is the case, proactively request supplementary data (higher frequency sampling, adjacent device status) for joint consultation until confirmation. Thresholds are calculated monthly using the formula. Adjustment.

[0031] Operation and maintenance strategy generation intelligent agent: adopting the RAG mode, the confirmed root cause is used as the query, relevant procedure fragments are retrieved in the SOP library (dense search, similarity threshold 0.7), and step-level instructions and high-risk labels are generated by combining the real-time status of the equipment.

[0032] (4) Global control execution layer and closed-loop evolution The system aggregates and disambiguates multi-agent policies (prioritizing instructions with higher security levels in case of conflict) and distributes them to execution units such as EMS and security systems through a cloud-edge-device architecture.

[0033] We introduce reinforcement learning loop closure optimization. The reward function is defined as: in Time (minutes) for residual elimination For power generation loss (kWh). The number of false alarms (times / day), weighted (Determined by AHP method).

[0034] The system aims to maximize cumulative rewards and uses the PPO algorithm to fine-tune the following parameters online: CUSUM threshold multiplier factor (Actual threshold) , (Updated every 24 hours based on extreme value theory) Number of search cases ; Reinforcement learning training: offline simulation environment pre-training + online sandbox security exploration (≤2 policy updates per day, with security checks performed before each update). This closed loop enables end-to-end self-evolution.

[0035] III. Beneficial Effects By constructing a digital twin of a device using a Physical Information Neural Network (PINN), physical constraints are used to overcome the high false alarm rate of a purely data-driven model. By combining the Cumulative Summation (CUSUM) algorithm with dynamic thresholds based on extreme value theory, progressive degradation can be identified tens of hours in advance during the "sub-healthy" stage of a device, thus solving the problem of severe lag in warnings based on fixed thresholds.

[0036] By fusing real-time graphs and historical cases through a graph attention network (GAT), the system outputs a root cause probability distribution and generates a natural language report using a large language model, thus achieving a leap from "just an alert" to "explainable and traceable results" and freeing the system from dependence on human experience.

[0037] Based on the Search Enhanced Generation (RAG) mode, the root cause is identified and matched with the Standard Operating Procedure (SOP) library, and operation instructions are automatically generated and issued for execution, improving emergency response efficiency and safety.

[0038] By introducing reinforcement learning and using rewards such as false alarm rate and processing time, thresholds and retrieval parameters are optimized online to achieve full-chain self-evolution and continuously ensure the safety of the site and the revenue from power generation. Attached Figure Description

[0039] Figure 1 This is a flowchart illustrating the overall workflow of the present invention. Detailed Implementation

[0040] This embodiment uses a 50MW photovoltaic power station as the application scenario. The power station includes 200 photovoltaic strings, 5 box-type transformers, and 1 step-up substation. The system deployment adopts a cloud-edge-device collaborative architecture: an edge computing node (configured with a 4-core CPU, 16GB of memory, and 1 NVIDIA Jetson Orin edge computing card) is deployed on each box-type transformer side to run the holographic perception and data governance layer and the PINN small model; a cloud server cluster (configured with 4 GPU servers, each equipped with 2 NVIDIA A100 40GB GPUs) is deployed in the central control center to run the industry large model, the multi-agent collaborative decision-making layer, and the reinforcement learning closed-loop evolution module.

[0041] I. System Deployment and Data Initialization First, intelligent sensing terminals are deployed on key equipment throughout the site: voltage / current transformers (sampling frequency 1kHz) are installed on the DC side of the photovoltaic strings, thermocouple temperature sensors are attached to the back of the modules, power quality monitoring terminals are installed on the AC side of the inverters, and acoustic fingerprint sensors (sampling frequency 25.6kHz) and infrared thermal imaging pan-tilt units are deployed. These devices are connected to edge computing nodes via a fiber optic ring network.

[0042] Before system commissioning, a 72-hour trouble-free trial run was conducted. During this period, multi-source data was collected from all devices, and initialization was completed as follows: 1. Calculate the standard deviation of the residuals: For the current residual sequence output by the PINN model of each device. Voltage residual sequence Calculate their standard deviations respectively. , which serve as the initial parameters for the subsequent CUSUM algorithm.

[0043] 2. Pre-trained governance model: Taking 7 consecutive days of complete, unmissing data, an LSTM missing data imputation model (128 hidden layer dimensions, 4 attention heads) is trained using supervised learning. The loss function converges to... Then fix the model parameters.

[0044] 3. Initialize physical parameters: For each photovoltaic string, use its first day of normal operation data to initialize the physical parameters in the PINN model. Perform pre-training to make its initial values ​​close to the equipment's factory nominal values.

[0045] II. Specific Operation of Holographic Perception and Data Governance Layer The system executes the following governance process in a loop every 200ms: Step 1: Multi-source data acquisition. Edge nodes simultaneously acquire the DC voltage of the string. (V) DC current (A) Backplate temperature (°C), ambient irradiance (W / m) 2 Meanwhile, the acoustic sensor collects runtime spectrum data, and the infrared thermal imager outputs the component temperature distribution matrix every 5 minutes.

[0046] Step 2: Three-level intelligent governance.

[0047] (a) Anomaly Detection: The real-time sampled current and voltage sequences are input into an isolated forest model (number of trees = 100, subsampling size = 256) with a window of 256 points. If the anomaly score of a sample is >0.6, it is identified as an outlier caused by sensor jumps and is directly removed. For example, if the measured current value at a certain moment is 15.3A (theoretical value is about 8.2A), and the isolated forest score is 0.82, it is identified as an outlier and discarded.

[0048] (b) Noise filtering: Filtering the signal after removing outliers. Adaptive wavelet thresholding denoising is performed. A 5-level decomposition using the db4 wavelet is employed, and the wavelet coefficients at each level are analyzed. Calculate the threshold ( (Estimated by the median absolute deviation of the wavelet coefficients at this layer). A soft thresholding function is used for processing. The denoised signal is obtained after reconstruction. Taking a certain series current signal as an example, the signal-to-noise ratio after denoising is improved from the original 12.3dB to 28.6dB.

[0049] (c) Missing value repair: This involves modifying the denoised signal... Input is fed into the pre-trained LSTM model. Assume that at the... The sampled value at time point is empty ( Model input history window ( and consecutive missing length (If the previous point is valid) (Otherwise, accumulate). The model outputs the predicted value. Filling in missing values. For example, if communication jitter causes three consecutive current sampling points to be lost, LSTM fills in the missing values ​​as 8.15A, 8.12A, and 8.09A respectively based on the trend and periodic pattern of the previous 64 points, with an error of less than 0.5% compared to the subsequent actual recovered values.

[0050] The data, after undergoing three levels of governance, forms a unified, high-frequency, and high-quality data foundation for subsequent engine use.

[0051] III. Specific Operation of the Dual-Driven Intelligent Analysis Engine Layer (I) Inner layer: Real-time anomaly capture of PINN small model Edge nodes maintain a PINN model for each photovoltaic string (structure: input layer receiver). The output layer is the predicted current. The middle four layers are fully connected, with 64 neurons in each layer, and the activation function is... The model's loss function is: The weights are adaptively updated every 24 hours using the GradNorm algorithm, with initial values... .

[0052] The specific calculation process is as follows: 1. Measure the voltage Irradiance (Converted to photocurrent) (initial values), temperature (Calculate thermal voltage) Enter PINN.

[0053] 2. The predicted current is obtained through forward propagation of the model. Simultaneously output learnable physical parameters The current estimate.

[0054] 3. Calculate the data error term .

[0055] 4. Calculate the physical residual term: Substitute into the single diode model equation, and we get... Taking the square gives .

[0056] 5. Calculate the regularization term. .

[0057] 6. The total loss is obtained by weighted summation. Backpropagation updates the network weights.

[0058] CUSUM Anomaly Detection and Dynamic Threshold Setting: The system calculates the current residual in real time at 5-second intervals. To detect minute drift in the residuals, an adaptive cumulative sum (CUSUM) algorithm is employed. For any residual component (e.g., current residual)... ), and denot its time series as Calculate the positive and negative cumulative sums: in , The standard deviation of this residual component is calculated as the initial value based on 72 hours of fault-free data collected before the system is put into operation, and recalculated every 7 days after operation based on the recent normal data window. Take 1.0. When or Exceeding the dynamic threshold Anomalies are detected in real time.

[0059] The dynamic threshold Automatic tuning is based on extreme value theory and decoupled from real-time detection: it is updated offline every 24 hours. The threshold is maintained independently for each residual component (e.g., current residual), and the update steps are as follows: 1. Take the sequence of normal operating condition observations from the most recent 7 days. ( For this component, the first (Values ​​at each sampling time).

[0060] 2. Extract samples exceeding the threshold Among them, the threshold Take the sequence Quantiles.

[0061] 3. Fit the tail using a generalized Pareto distribution: , The shape parameters are obtained through maximum likelihood estimation. and scale parameters .

[0062] 4. Set the expected false alarm rate Calculate the threshold: in The total number of samples, This represents the number of samples exceeding the threshold.

[0063] 5. Calculate the As the reference threshold for this component In actual operation, the reinforcement learning phase will be... Adjusting the multiplier factor within the range That is, the actual threshold of this component. Each residual component maintains its own independent maintenance. .

[0064] CUSUM Anomaly Detection Example: Taking a series of strings where the current slowly decays due to the PID effect as an example. Gradually drift from 0A to -0.8A. Initialization parameters: , The offline calculations obtained on that day (Assuming the current situation) Therefore, in reality ). Calculate the positive cumulative sum. negative cumulative sum .because Continuously negative, After 120 minutes, the efficiency reached 2.6 > 2.5, triggering an anomaly warning. At this point, the measured efficiency only decreased by 3%, while traditional threshold alarms typically only trigger when the efficiency decreases by more than 10%, thus achieving proactive warning.

[0065] (II) Outer Layer: In-depth Tracing and Early Warning of Industry-wide Models When CUSUM is triggered, the edge node will be at an abnormal moment. and multidimensional residual data for the preceding and following 10 minutes (including and the residuals from the conversion of voiceprint features Thermal imaging feature residuals Package and upload to the cloud.

[0066] Cloud processing flow: 1. Feature Extraction: Calculate the mean, variance, skewness, and maximum drift rate of each residual component to form a feature vector. .

[0067] 2. Case Retrieval: In the fault case database, cosine similarity is used to retrieve cases... The top 5 most similar historical cases. Suppose we retrieve case A (root cause "PID effect") and case B (root cause "occlusion attenuation"), with similarities of 0.89 and 0.76 respectively.

[0068] 3. Spectrum Fusion: Extract the real-time spectrum sub-graph of the current device (string #115). This includes nodes "string #115", "inverter #3", and "combiner box #7" and their electrical connections. The diagram for Case A is shown below. Through mapping function (Projecting historical equipment model characteristics onto the current equipment characteristic space) yields .

[0069] 4. GAT Reasoning: The input is fed into an 8-head graph attention network. Taking node "string #115" as an example, the attention coefficient between it and its neighbor node "merge box #7" is calculated. : The final GAT output root cause probability distribution is as follows: PID effect (85%), occlusion attenuation (10%), and others (5%).

[0070] 5. Large model report generation: Includes root cause probability, residual summary ("mean current residual -0.8A, continuous negative bias for 120 minutes"), and... The textual description ("Chain #115 → Combiner Box #7 → Inverter #3") is concatenated into a prompt. The industry-wide model with 7B parameters (fine-tuned by power commands, 4-bit quantization) generates the following natural language report: Root cause analysis: The current in string #115 remained consistently lower than the model's predicted value. Based on negative cumulative current, exceeding the threshold, and historical case matching, this was determined to be a PID effect (85% confidence level). Evolution prediction: Without intervention, efficiency is expected to decrease by more than 10% after 12 hours.

[0071] Verification suggestion: Please check the insulation resistance to ground of string #115 and retrieve the nighttime IV curve.

[0072] IV. Specific Operation of the Multi-Agent Collaborative Decision-Making Layer 1. Root cause diagnosis agent: The current maximum root cause probability is 85%, which is higher than the threshold. Without requesting additional data, the root cause is directly identified as the "PID effect." If the probability of a root cause is only 65%, the agent will proactively request: ① to perform 1kHz high-frequency sampling on the target string for 30 seconds; ② to query the real-time status of the three adjacent strings. After data is supplemented, the inference process is repeated until the probability exceeds a threshold or the maximum number of iterations (3) is reached. Threshold Monthly according to formula Adaptive adjustment.

[0073] 2. Operation and Maintenance Strategy Generation Agent: The confirmed root cause, "PID effect," is used as the query, and a dense search is performed in the SOP database (using the BGE-large-v1.5 model). A procedure fragment with a similarity of 0.85 is found: "Nighttime Anti-PID Repair Operation Ticket." Combined with the real-time device status (string #115 current voltage 620V, ambient nighttime temperature 5°C), step-level instructions are generated: [High-Risk Operation] Step 1: Wait until 2:00 AM tomorrow (irradiance <10W / m²) 2 Disconnect the DC switch of string #115.

[0074]

Safety Confirmation

[0075] [Standard Operation] Step 3: Start the PID repair module, apply +800V to ground voltage for 4 hours.

[0076] [Restore Verification] Step 4: Restore the wiring, measure the IV curve of the string, and compare it with the health model.

[0077] V. Specific Operation of the Overall Control Execution Layer and Closed-Loop Evolution The cloud-based system aggregates and disambiguates the aforementioned operation tickets (assuming no other conflicting instructions) before distributing them to the edge nodes via the MQTT protocol. At 2:00 AM, the edge nodes automatically execute the "Disconnect DC Switch" command and, after confirming the switch status via hardwiring, trigger the PID repair module. During execution, the system continuously monitors the repair effect: if the residual error persists after 4 hours... If the value still has not reached zero, the repair time will be automatically extended.

[0078] Example of a closed-loop reinforcement learning system: Various penalties in the system's daily reward function: The average time from triggering the last 5 exceptions to residual elimination was 42 minutes.

[0079] The average power generation loss due to false alarms is 12 kWh / day.

[0080] The number of false alarms is 1.5 times per day.

[0081] Calculate daily rewards The PPO algorithm aims to maximize cumulative rewards and is adjusted online. CUSUM threshold multiplier factor: Due to the high false alarm rate (target 1%, actual 1.8%), it will be... The value was adjusted from 1.0 to 0.92, thereby adjusting the actual threshold. Reduce, increase sensitivity.

[0082] Number of cases retrieved: It was found that increasing the K value can reduce false negatives, but it will increase inference latency. The current optimal K value is stable at 7, so this value is retained.

[0083] Before updating the strategy, a security check is performed in a sandbox environment (simulating whether the adjustment will cause a surge in false alarms). Up to two updates are performed per day to ensure continuous evolution without introducing new risks.

[0084] VI. Effect Verification This embodiment demonstrates continuous operation of the 50MW photovoltaic power plant for 90 days. Actual operating data shows that: 1. Early Warning: A total of 7 instances of equipment degradation caused by PID effects, hot spots, etc. occurred. The system triggered warnings for all of these instances when the efficiency decreased by less than 5%, with the longest warning lead time being 28 hours. In contrast, the original threshold alarm system only captured 2 of these instances, and both of them were triggered after the efficiency decreased by more than 12%.

[0085] 2. Precise positioning: The consistency rate between the root cause analysis output by the dual-drive engine and the verification results by on-site maintenance personnel was 85.7% (6 / 7), and there was no physical logic error such as "high current on sunny days being falsely reported as shading" that is common in pure data-driven models.

[0086] 3. Automated decision-making: The operation and maintenance strategies generated by the multi-agent system were directly adopted and executed by the operation and maintenance personnel in 6 out of 7 anomalies, reducing the average fault handling time from the original 4.5 hours to 0.5 hours (mainly due to waiting for the night window), and significantly reducing the need for manual intervention.

[0087] The above implementation process fully realizes the technical link of "holographic perception → dual-drive analysis → multi-agent decision-making → closed-loop control" proposed in this invention, which fully demonstrates the practicality, advanced nature and significant beneficial effects of this system in the safety risk analysis and control of new energy power stations.

[0088] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A smart analysis and control system for safety risks at new energy power stations, characterized in that, include: The holographic perception and data governance layer is used to collect multi-source heterogeneous data from new energy power stations and forms a data foundation through three levels of intelligent governance: anomaly detection, noise filtering, and missing value repair. The dual-drive intelligent analysis engine layer comprises an inner layer and an outer layer. The inner layer utilizes the Physical Information Neural Network (PINN) to perform millisecond-level state estimation of the device, calculates multidimensional residual vectors in real time, employs adaptive accumulation and CUSUM algorithms to detect minute residual drifts, and automatically adjusts dynamic thresholds based on extreme value theory. When CUSUM exceeds the dynamic threshold, an anomaly is triggered. After an anomaly is triggered, the outer layer extracts multidimensional residual statistical features, retrieves similar cases from the fault case library, merges the current device's real-time graph subgraph with the retrieved case graphs, and feeds them into the Graph Attention Network (GAT) to output the root cause probability distribution. A natural language report is then generated by an industry-wide large model. The multi-agent collaborative decision-making layer includes a root cause diagnosis agent and an operation and maintenance strategy generation agent. When the root cause diagnosis agent requests supplementary data and conducts joint consultation when the root cause probability is lower than a threshold, the operation and maintenance strategy generation agent adopts the retrieval-enhanced generation (RAG) mode, takes the confirmed root cause as the query, retrieves relevant procedure fragments in the standard operating procedure (SOP) library, and generates step-level instructions. The global control execution layer is used to aggregate and disambiguate multi-agent policies and then distribute them to the execution unit. It also introduces a reinforcement learning closed loop to construct a reward function based on residual elimination time, power generation loss, and false alarm count, and to fine-tune the CUSUM threshold multiplier factor and the number of cases retrieved online.

2. The intelligent analysis and control system for safety risks of new energy power stations according to claim 1, characterized in that, In the holographic perception and data governance layer: the anomaly detection uses the isolated forest algorithm, with 100 trees in the isolated forest, a subsampling size of 256, and an anomaly score threshold of 0.6; the noise filtering uses adaptive wavelet threshold denoising, selecting the db4 wavelet, and the decomposition layer number... Wavelet coefficients are processed using a soft thresholding function, and the threshold value is... , For the first Layer noise standard deviation The length of the wavelet coefficients in this layer; the missing value repair uses an LSTM model based on a time attention mechanism, with a hidden layer dimension of 128 and 4 attention heads. The model input includes the observations within the historical window and the length of consecutive missing values. The loss function is: .

3. The intelligent analysis and control system for safety risks of new energy power stations according to claim 1, characterized in that, In the inner layer: The loss function for PINN is: in The GradNorm algorithm is used to adaptively update the initial value in each training cycle. , , ; The residual function of the single-diode model of the photovoltaic cell. 。 4. The intelligent analysis and control system for safety risks of new energy power stations according to claim 1, characterized in that, In the CUSUM algorithm, for the residual component sequence The formula for calculating the cumulative sum of positive and negative numbers is: in , This is the standard deviation of the residual component. Take a value of 0.5~1.5; when or Exceeding the dynamic threshold Anomalies are detected in real time.

5. The intelligent analysis and control system for safety risks of new energy power stations according to claim 1 or 4, characterized in that, The dynamic threshold The tuning method is as follows: update offline every 24 hours, taking the sequence of normal operating condition observations from the most recent 7 days. Extract samples exceeding the threshold , For the 95th percentile of this sequence, fit the tail using a generalized Pareto distribution: Shape parameters are obtained through maximum likelihood estimation. and scale parameters Set the expected false alarm rate ,calculate: in The total number of samples, The number of samples exceeding the threshold; the calculated As a benchmark threshold Actual threshold , Adjusted online via reinforcement learning, each residual component maintains its own independent maintenance. .

6. The intelligent analysis and control system for safety risks of new energy power stations according to claim 1, characterized in that, In the outer layer: extracting abnormal moments. The multidimensional residual statistical characteristics within the preceding and following 10 minutes form a feature vector. Search in the fault case database Similar cases, ; The graph attention network employs an 8-head attention structure and node feature vectors. Attention coefficient: The industry-specific large model is a generative large language model based on the Transformer decoder, with 7B parameters, 32 layers, 4096 hidden layers, and 32 attention heads. It has been fine-tuned with power industry instructions and adapted to LoRA. During deployment, it uses 4-bit quantization and a vLLM inference engine.

7. The intelligent analysis and control system for safety risks of new energy power stations according to claim 1, characterized in that, In the root cause diagnosis agent, the root cause probability threshold The initial value is 0.78, and it is calculated monthly according to the formula: Adjustment.

8. The intelligent analysis and control system for safety risks of new energy power stations according to claim 1, characterized in that, The reward function for reinforcement learning is: in Time (minutes) for residual elimination For power generation loss (kWh). Number of false alarms (times / day) , , The PPO algorithm is used for online optimization. The training method combines offline simulation environment pre-training with online sandbox security exploration. The policy is updated a maximum of 2 times a day, and security verification is performed before each update.

Images