A new energy station equipment state evaluation and early warning method

Abstract

The present application relates to the technical field of equipment health management, and provides a new energy station equipment state evaluation and early warning method. It comprises the following steps: S1: collecting operation, environment and inspection image data in real time through multi-source sensing equipment, and generating a data set after standardization processing; S2: selecting an AI large model supporting multi-modal input, injecting a fault knowledge graph and industry standards, and generating an adaptive model through general power data pre-training and station-specific data fine-tuning; S3: extracting time domain / frequency domain features and image damage information, and dynamically weighting to construct a unified health index; S4: predicting the HI value in the next 24 hours based on LSTM, dynamically adjusting the threshold combined with the working condition, and executing the early warning in three levels; and S5: updating the model every quarter, verifying the index requirements, and optimizing the rules through false alarm analysis. The present application realizes intelligent evaluation and accurate early warning of multi-source data fusion, improves operation and maintenance efficiency, reduces non-scheduled downtime by more than 100 hours / year, and promotes the unattended transformation of new energy stations.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of equipment health management, in particular to a new energy station equipment state evaluation and early warning method. BACKGROUND

[0002] New energy stations have high operation and maintenance costs and difficult fault handling due to the dispersion of equipment in remote and complex environments, the variety of types, and the complexity of the system. Traditional management systems cannot meet the needs. The power AI large model can more comprehensively and accurately evaluate the equipment state and intelligently warn with the advantages of multi-source data fusion and processing, power industry knowledge sedimentation and iterative evolution, and cross-device dimension comprehensive evaluation. In addition, the energy and power industry is facing transformation challenges, and such new technology fusion applications are urgently needed to promote the intelligent upgrading of new energy stations.

[0003] Invention patent CN119417679A discloses a multi-region monitoring system and method based on an AI vision model, including the following steps: collecting regional information; generating two-dimensional structure data based on regional information; inputting two-dimensional structure data into a density model, and the density model outputs results representing the flow density of each region; inputting two-dimensional structure data into a control model, and the control model outputs results representing flow control measures; the present application reduces the influence of personnel occlusion on the accuracy of flow density calculation by analyzing real-time regional information of each region and combining historical flow data of the region, and obtains flow control measures according to the flow density of each region. Compared with only calculating flow density, this system can provide flow density reduction effect, low execution cost, and improved safety of each region.

[0004] However, the existing technology mainly relies on visual data, and has limited access to internal performance parameters and deep health state evaluation of equipment. In addition, due to the use of specific AI models, it is difficult to adapt to heterogeneous equipment, does not have self-evolution learning ability, and cannot dynamically adjust model parameters according to equipment operation data. Therefore, in actual application, there are still problems such as insufficient prediction accuracy, poor generalization ability, and inability to quantify risks in advance, making it difficult to discover potential equipment failures in time and take preventive measures. SUMMARY

[0005] The technical problem to be solved by the present application is to provide a new energy station equipment state evaluation and early warning method to improve the operation and maintenance management efficiency of new energy stations.

[0006] The technical solution adopted by the present application to solve the technical problem is a new energy station equipment state evaluation and early warning method, including the following steps:

[0007] S1. Data collection and management: Collect operation data, environmental data and inspection image data of new energy station equipment through multi-source sensing devices, standardize the data, and generate a standardized dataset containing structured data tables and unstructured file libraries.

[0008] S2. Large model selection and pre-training: Select an AI large model that supports multi-modal input, inject equipment fault mechanism knowledge graph and industry regulations for pre-training and adjustment, and generate a multi-modal large model adapted to the power industry.

[0009] S3. Feature extraction and health index construction: Use the large model in S2 to extract time domain, frequency domain and image features of the equipment, construct a unified health index system for the equipment, and obtain a unified health index for the equipment.

[0010] S4. Early warning model development and dynamic strategy: Based on the health index in S3, combine time series prediction model and anomaly detection model to generate fault warning, and dynamically adjust warning threshold and grading strategy according to working conditions.

[0011] S5. Model verification and iterative optimization: Verify the warning effect through historical data backtracking and field demonstration, and feedback to optimize the data collection rules in S1, the large model parameters in S2 and the warning strategy in S4.

[0012] Further, step S1 comprises:

[0013] Connect SCADA system through DL / T634.5104 protocol to obtain operation data, and integrate inspection system data through API interface, with data delay ≤10 seconds;

[0014] Linear interpolation is used to fill in missing data, and the unified time granularity is 5-minute average, with device ID, collection time and working condition label marked for the data.

[0015] Further, step S2 comprises: selecting an AI large model that supports time series data and image input; through two-stage training, the first stage uses general power dataset for pre-training, and the second stage injects new energy station exclusive data, and uses LoRA technology to adjust feature weight.

[0016] Further, in step S3, the construction formula of the unified health index is: HI=A×running feature score+B×environmental adaptation score+C×inspection image score, wherein the weight is dynamically adjusted according to the device type, and the cross-device health state ranking is realized through normalization processing.

[0017] Further, step S4 comprises: predicting the HI value in the next 24 hours based on LSTM and the time series inference ability of the large model in S2, and triggering warning when the predicted value is less than the threshold.

[0018] Further, step S4 further comprises a multi-level warning mechanism:

[0019] The first level early warning is pushed to the operation and maintenance terminal and triggers sound and light alarm;

[0020] The second level early warning generates a to-be-processed work order;

[0021] The third level early warning is included in the periodic inspection plan.

[0022] Further, in step S5, the verification index requirement is that the early warning accuracy is greater than or equal to 90%, the recall rate is greater than or equal to 95%, and the advance amount is greater than or equal to 24 hours; the S2 large model is retrained with new data every quarter, and the work condition association rule of S4 is updated based on the false alarm analysis.

[0023] Further, the knowledge graph constructed in step S2 includes a device fault causal relationship chain and an operation and maintenance procedure conforming to GB / T38335, NB / T11346 and NB / T11570 standards.

[0024] Further, the feature extraction of step S3 further includes extracting a vibration signal characteristic frequency through Fourier transform, and combining the S2 large model to identify weak fault characteristics.

[0025] The beneficial effects of the present application are: the present application fuses multi-source heterogeneous data through the power AI large model, realizes the automatic and intelligent unified evaluation and accurate early warning of the new energy station equipment state, can greatly improve the operation and maintenance efficiency, reduce the unplanned downtime and maintenance cost, and reduce the power generation loss; at the same time, the dynamic adjustment of the early warning threshold and the construction of the knowledge graph realize the knowledge precipitation and algorithm iteration, break through the limitations of traditional operation and maintenance, on the basis of enhancing the fault diagnosis accuracy, guaranteeing the data security and system stability, promote the transformation of the station to unattended, promote the efficient consumption of new energy and the collaborative optimization of energy system, and have technical innovation and industry application value. BRIEF DESCRIPTION OF DRAWINGS

[0026] Figure 1 is a flowchart of the present application. DETAILED DESCRIPTION

[0027] Embodiments of the present application will be described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by referring to the accompanying drawings are exemplary and are only used to explain the present application, and cannot be understood as a limitation of the present application.

[0028] As shown in Figure 1 , a new energy station equipment state evaluation and early warning method of the present application includes the following steps:

[0029] S1. Data collection and management: Collect operation data, environmental data and inspection image data of new energy station equipment through multi-source sensing devices, standardize the data, and generate a standardized dataset containing structured data tables and unstructured file libraries.

[0030] S2. Large model selection and pre-training: Select an AI large model that supports multi-modal input, inject equipment fault mechanism knowledge graph and industry regulations for pre-training and adjustment, and generate a multi-modal large model adapted to the power industry.

[0031] S3. Feature extraction and health index construction: Extract time domain, frequency domain and image features of equipment using the large model in S2, which are the equipment data collected in S1, construct a unified health index system for equipment, and obtain a unified health index for equipment.

[0032] S4. Early warning model development and dynamic strategy: Based on the health index in S3, combine time series prediction model and anomaly detection model to generate fault warning, and dynamically adjust warning threshold and grading strategy according to working conditions.

[0033] S5. Model verification and iterative optimization: Verify the warning effect through historical data backtracking and field demonstration, and feedback to optimize the data collection rules in S1, the large model parameters in S2 and the warning strategy in S4.

[0034] Further, step S1 includes:

[0035] Obtain operation data through DL / T634.5104 protocol interface with SCADA system, integrate inspection system data through API interface, and data delay ≤10 seconds;

[0036] Linear interpolation is used to fill in missing data, and the unified time granularity is 5-minute average, and the data is labeled with equipment ID, collection time and working condition label.

[0037] Further, step S2 includes: Select an AI large model that supports time series data and image input; Through two-stage training: the first stage uses general power dataset for pre-training, and the second stage injects new energy station exclusive data, and uses LoRA technology to adjust feature weight.

[0038] Further, the construction formula of the unified health index in step S3 is: HI=A×running feature score+B×environmental adaptation score+C×inspection image score, wherein the weight is dynamically adjusted according to the type of equipment, and the cross-equipment health state ranking is realized through normalization processing.

[0039] Further, step S4 includes: Based on the time series inference ability of LSTM and the large model in S2, predict the HI value in the next 24 hours, and trigger the warning when the predicted value is less than the threshold.

[0040] Further, step S4 also includes a multi-level early warning mechanism:

[0041] The first level early warning is pushed to the operation and maintenance terminal and triggers sound and light alarm;

[0042] The second level early warning generates a to-be-processed work order;

[0043] The third level early warning is included in the periodic inspection plan.

[0044] Further, the verification indicators in step S5 require that the early warning accuracy is ≥ 90%, the recall rate is ≥ 95%, and the advance amount is ≥ 24 hours; the S2 large model is retrained every quarter with new data, and the work condition association rules of S4 are updated based on false alarm analysis.

[0045] Further, the knowledge graph constructed in step S2 includes: device fault causal relationship chain, and operation and maintenance procedures conforming to GB / T38335, NB / T11346, and NB / T11570 standards.

[0046] Further, the feature extraction of step S3 also includes extracting the frequency characteristics of the vibration signal through Fourier transform, and identifying the weak fault characteristics in combination with the S2 large model; it can also locate the device damage area through the target detection algorithm and quantify the crack length and hot spot area.

[0047] Embodiment 1

[0048] S1: Data collection and management

[0049] S11: Data collection implementation

[0050] Vibration sensors are installed at the fan gear box and generator bearing, and infrared temperature measurement sensors are deployed in the photovoltaic array; a meteorological station is set up at the boundary of the field station to collect wind speed, light intensity, temperature and humidity, and is synchronously connected to the SCADA system; a drone is configured to take pictures of the surface of the fan blades every week, and a robot inspects the photovoltaic panel hot spots (infrared image accuracy ±0.5℃) every day.

[0051] S12: Data management operation

[0052] Cleaning: automatically remove out-of-range data, such as photovoltaic panel temperature > 100℃, and linearly interpolate missing wind speed data (≤5 minutes) between adjacent nodes;

[0053] Standardization: aggregate second-level vibration data into 5-minute averages, and unify power units to kW and temperature units to ℃;

[0054] Association labeling: structured data (such as sensor readings) are stored in a MySQL table, which includes device ID, timestamp, and working condition label; unstructured data, such as infrared images, are stored in a MinIO object storage, and are associated with structured data through device ID.

[0055] S2: Model selection and pre-training

[0056] S21: Model selection:

[0057] Opt for Guangming Power Large Model, support time series vibration signal + infrared image multi-modal input, actual measurement inference delay 380ms.

[0058] S22: Knowledge graph construction:

[0059] Integrate fault mechanism: for example, the causal chain of "gearbox oil temperature > 80℃ will cause the probability of bearing wear to rise";

[0060] Embed industry standards: GB / T38335 "Operation Regulations for Photovoltaic Power Stations", NB / T11346 "Maintenance Regulations for Photovoltaic Power Stations", NB / T11570 "Technical Guidelines for Intelligent Operation of Wind Farms", form a structured knowledge graph.

[0061] S23: Two-stage training:

[0062] Pre-training: pre-train with general power dataset such as public IEEE power equipment fault library, so that the model masters basic power knowledge;

[0063] Fine-tuning: inject historical data of the site (including 200 cases of wind turbine yawing failure and 150 cases of photovoltaic hot spot), and fine-tune the feature weights using LoRA technology.

[0064] S3: Feature extraction and health index construction

[0065] S31: Multi-modal feature extraction:

[0066] For example, the operating characteristics of a wind turbine include vibration (amplitude, frequency), bearing temperature, power, speed, oil pressure, etc., environmental characteristics include wind speed, wind direction, air density, ambient temperature, turbulence intensity, etc., and inspection image characteristics include blade cracks / corrosion, bolt loosening, oil leakage, electrical overheating, etc.

[0067] The operating characteristics of a photovoltaic panel include DC current / voltage, AC power, conversion efficiency, insulation resistance, etc., environmental characteristics include light intensity (irradiance), ambient temperature, wind speed, humidity, snow cover, etc., and inspection image characteristics include cell cracks, hot spots, broken cells, junction box cover damage, support corrosion, deformation, dust, bird droppings, snow coverage area and thickness estimation, and vegetation shading area, etc.

[0068] The operating characteristic score is a comprehensive score value of 0-100 calculated by a sub-model or rule set (for example, comparing the current characteristic value with the threshold range of the health state) after preprocessing and extraction of the operating characteristics.

[0069] The higher the environmental adaptation score, the better the device performs in the current environment; the lower the score, the lower the device's ability to adapt to environmental changes or potential problems.

[0070] The inspection image score is obtained by quantifying the severity and area size of the above-mentioned image features. For example, no cracks and hot spots are 100 points, small cracks deduct 10 points, large area hot spots deduct 30 points, etc.

[0071] S32: Unified health index (HI) calculation:

[0072] Select the values of A, B and C in the HI formula according to different device types, for example, fan HI formula: 0.6 x operating feature score + 0.3 x environmental adaptation score + 0.1 x inspection image score;

[0073] Photovoltaic panel HI formula: 0.5 x operating feature score + 0.4 x environmental adaptation score + 0.1 x inspection image score;

[0074] Cross-device calibration: all HI values are normalized by Z-score to generate a site health ranking list, for example, fan HI = 42, then the fan is in a risk state.

[0075] S4: Early warning model and dynamic strategy

[0076] S41: Time series prediction model:

[0077] LSTM input past 72 hours HI sequence, superimposed with environmental features output by S2 large model, such as future wind speed prediction, output 24 hours HI prediction value; When the predicted value drops below the threshold, for example, fan HI < 50; photovoltaic panel HI < 45, trigger early warning.

[0078] S41: Multi-level early warning execution:

[0079] Level 1 early warning: trigger sound and light alarm and send short message to responsible person; For example, fan HI < 50, and the decline rate is greater than or equal to 50%.

[0080] Level 2 early warning: automatically generate maintenance work order; For example, fan HI < 50, and the decline rate is greater than or equal to 25% and less than 50%.

[0081] Three-level early warning: included in the next week's UAV special inspection plan; for example, the fan HI <50, the decline rate is greater than or equal to 10%, and less than 20%; the decline rate here is the decline rate in a warning period, for example, the fan HI at the end of the last period is 80, and the HI at the end of this period is 40, the decline rate is greater than or equal to 50%, which triggers an audible and light alarm and sends a short message to the person in charge; for example, the fan HI at the end of the last period is 80, and the HI at the end of this period is 70, the decline rate is 12.5%, which is included in the next week's UAV special inspection plan.

[0082] S5: model verification and iterative optimization

[0083] S51: model verification method

[0084] Historical data backtracking: select 500+ fault cases in the past year, test the model early warning lead time (target ≥24 hours), calculate the accuracy (TP / (TP+FP) ≥90%) and recall rate (TP / (TP+FN) ≥95%).

[0085] On-site empirical test: deploy the model in 3 pilot stations, compare the early warning results with the actual faults, and collect operation and maintenance feedback.

[0086] S52: continuous iteration measures

[0087] According to the feedback of S51, supplement the missing data, such as adding the SOC data of the energy storage battery, optimize the cleaning rules of S1, such as correcting the compensation algorithm of sensor drift. Re-train the S2 model every quarter with new data, adjust the feature weights of S3, such as increasing the weight of "humidity on the life of photovoltaic panels". Based on the reasons for false alarms, update the working condition association rules of S4, such as refining the threshold adjustment scheme in different seasons.

[0088] S53: closed-loop optimization effect

[0089] Through 3 rounds of iteration, the false alarm rate is reduced by 50%, the fault detection lead time is extended to 48 hours, and the annual non-scheduled downtime of the station is reduced by ≥100 hours.

[0090] The embodiments of the specific implementation are the preferred embodiments of the present application, not limited to the protection scope of the present application, so: any equivalent changes made according to the structure, shape, principle of the present application should be covered within the protection scope of the present application.

Claims

1. A method for assessing and providing early warning of the condition of equipment in a new energy power station, characterized in that, Includes the following steps: S1. Data Acquisition and Governance: Collect operational data, environmental data, and inspection image data of new energy power station equipment through multi-source sensing devices, standardize and govern the data, and generate a standardized dataset containing structured data tables and unstructured file libraries; S2. Large Model Selection and Pre-training: Select a large AI model that supports multimodal input, inject equipment fault mechanism knowledge graph and industry procedures for pre-training and adjustment, and generate a multimodal large model adapted to the power industry; S3. Feature Extraction and Health Indicator Construction: Using the large model in S2, time-domain, frequency-domain, and image features of the equipment are extracted to construct a unified health index system for the equipment, resulting in a unified health index. The formula for constructing the unified health index is: HI = A × Operation Feature Score + B × Environment Adaptation Score + C × Inspection Image Score, where the weights are dynamically adjusted according to the equipment type, and cross-equipment health status ranking is achieved through normalization processing. The operation feature score is a comprehensive score value of 0-100 points calculated by a sub-model or rule set after the operation data has been preprocessed and extracted. The environmental adaptability score is a comprehensive score of 0-100 points based on the equipment's ability to adapt to the environment; the inspection image score is a comprehensive score of 0-100 points obtained by quantifying and scoring the inspection image data. S4. Early Warning Model Development and Dynamic Strategy: Based on the health index of S3, a fault early warning is generated by combining a time series prediction model and anomaly detection model, and the early warning threshold and grading strategy are dynamically adjusted according to the operating conditions. S5. Model Validation and Iterative Optimization: Verify the early warning effect through historical data backtracking and on-site empirical studies, and optimize the data governance rules of S1, the large model parameters of S2, and the early warning strategy of S4.

2. The method for assessing and issuing early warnings of the status of new energy power station equipment according to claim 1, characterized in that, Step S1 includes: obtaining operational data by connecting to the SCADA system via the DL / T634.5104 protocol; integrating inspection system data via the API interface, with a data delay of ≤10 seconds; filling missing data with linear interpolation and labeling the data with device ID, acquisition time, and operating condition label.

3. The method for assessing and issuing early warnings of the status of new energy power station equipment according to claim 1, characterized in that, Step S2 includes: selecting a large AI model that supports time-series data and image input; and training in two stages. In the first stage, the model is pre-trained using a general power dataset, and in the second stage, it is injected with data specific to new energy power plants, and the feature weights are adjusted using LoRA technology.

4. The method for assessing and issuing early warnings of the status of new energy power station equipment according to claim 1, characterized in that, Step S4 includes: predicting the HI value for the next 24 hours based on the temporal inference capabilities of LSTM and the S2 large model, and triggering an early warning when the predicted value is less than the critical value.

5. The method for assessing and issuing early warnings of the status of new energy power station equipment according to claim 4, characterized in that, Step S4 also includes a multi-level early warning mechanism: A Level 1 warning is pushed to the operation and maintenance terminal and triggers an audible and visual alarm. Level 2 warning generates pending work orders; Level 3 early warnings are incorporated into the periodic inspection plan.

6. The method for assessing and issuing early warnings of the status of new energy power station equipment according to claim 1, characterized in that, The verification metrics in step S5 are as follows: early warning accuracy ≥ 90%, recall ≥ 95%, and lead time ≥ 24 hours; the S2 large model is retrained every quarter with new data, and the working condition association rules in step S4 are updated based on false alarm analysis.

7. A method for assessing and providing early warning of the status of new energy power station equipment according to any one of claims 1-6, characterized in that, The knowledge graph constructed in step S2 includes the causal relationship chain of equipment failure, as well as the operation and maintenance procedures that comply with the standards GB / T38335, NB / T11346, and NB / T11570.

8. A method for assessing and providing early warning of the status of new energy power station equipment according to any one of claims 1-6, characterized in that, The feature extraction in step S3 also includes extracting the characteristic frequencies of the vibration signal through Fourier transform and combining them with the large model in S2 to identify weak fault characteristics.

Images