Hydraulic power plant intelligent operation and maintenance management system and method based on machine learning

Abstract

The invention discloses a hydraulic power plant intelligent operation and maintenance management system and method based on machine learning, and belongs to the technical field of hydraulic power plant operation and maintenance. Multi-dimensional data of a hydraulic power plant are collected, preprocessed and stored into a data warehouse; defining a unified metadata label, and establishing a data blood relationship tracking mechanism; determining a traceability starting point, backtracking based on a storage position, tracing a data access history, backtracking according to a data flow diagram, and integrating to obtain a data traceability chain; entity identification and relation extraction are carried out, and Neo4j is used to construct a correlation graph; analyzing the relationship between the power grid dispatching data and equipment operation, and adjusting the association graph; identifying dominant anomalies and implicit anomalies, and performing fusion based on the association map to obtain abnormal data of the hydraulic power plant; predicting trends of equipment operation data and sensor monitoring data, analyzing and positioning a deviation value in combination with a residual error, obtaining a data fluctuation mode, analyzing correlation, and obtaining a source of abnormal data of the hydraulic power plant.

Description

Technical Field

[0001] The present invention relates to the technical field of hydropower plant operation and maintenance, and specifically to a hydropower plant intelligent operation and maintenance management system and method based on machine learning. Background Art

[0002] With the rapid development of the power industry, hydropower plants, as a vital source of electricity supply, are crucial for ensuring the reliability of power systems. Hydropower plant operations generate massive amounts of multi-dimensional data. With the trend toward intelligent O&M, leveraging this multi-dimensional data to achieve real-time monitoring, fault warnings, and precise maintenance of hydropower plant equipment has become a critical challenge for the industry. Traditional hydropower plant O&M management methods rely primarily on manual inspections and empirical judgment, which are unable to meet the demands of modern hydropower plants for efficient and precise O&M.

[0003] Existing technologies often lack efficient data collection, preprocessing, and storage mechanisms when processing massive amounts of multidimensional data from hydropower plants. Different types of data may originate from different systems and devices, and data formats and interfaces are inconsistent, making data integration difficult and hindering the full realization of the data's value. Existing technologies lack the ability to deeply mine and analyze the relationships between multidimensional data from hydropower plants, making it difficult to comprehensively understand the plant's operating status. Existing anomaly detection and diagnosis methods primarily rely on single data features or simple threshold judgments, which are prone to misjudgments and missed detections. Summary of the Invention

[0004] The purpose of the present invention is to provide a hydropower plant intelligent operation and maintenance management system and method based on machine learning to solve the problems raised in the prior art.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] In a first aspect, the present invention provides a method for intelligent operation and maintenance management of a hydropower plant based on machine learning, comprising the following steps:

[0007] Collect multi-dimensional data from hydropower plants, including equipment operation data, sensor monitoring data, and grid dispatch data, perform data preprocessing, and store it in a data warehouse. Within the data warehouse, define unified metadata tags and establish a data lineage tracking mechanism. Determine the traceability starting point, trace back based on the storage location, and trace back data access history. This is done based on the data flow diagram and integrated to form a data traceability chain.

[0008] Based on the data warehouse and data traceability chain, we perform entity recognition and relationship extraction, and use Neo4j to build a correlation map. We analyze the relationship between power grid dispatch data and equipment operation, including instruction impact analysis and chain reaction analysis, and adjust the correlation map.

[0009] Based on equipment operation data and sensor monitoring data, anomalies are detected using isolation forests to obtain explicit anomalies. The data distribution is reconstructed through an autoencoder to identify implicit anomalies. Explicit and implicit anomalies are fused based on the correlation graph to obtain abnormal data of the hydropower plant.

[0010] The Prophet model is used to predict the trends of equipment operation data and sensor monitoring data. The deviation values ​​are located by combining residual analysis. The data fluctuation pattern is obtained based on the deviation values. The correlation between the abnormal data of the hydropower plant and the data fluctuation pattern is analyzed to obtain the source of the abnormal data of the hydropower plant.

[0011] In conjunction with the first aspect, in a first implementation of the first aspect of the present application, the multi-dimensional data of the hydropower plant, including equipment operation data, sensor monitoring data, and power grid dispatching data, is collected, pre-processed, and stored in a data warehouse, including:

[0012] Collect equipment operating data, including hydro-generator sets, transformers, and switchgear, at a set collection frequency and transmit it to a data collection server via a fieldbus or wireless network. Sensors monitor equipment and environmental conditions in real time, converting the collected sensor monitoring data into digital signals for transmission. Establish a data connection with the power grid dispatching system and obtain the latest power grid dispatching data, including dispatching instructions and power grid operation data, based on the grid dispatching system's update frequency.

[0013] The data preprocessing includes data cleaning and data conversion. The data warehouse is built using a data lake, wherein the data lake adopts Delta Lake.

[0014] In combination with the first aspect, in a second implementation of the first aspect of the present application, defining unified metadata tags in the data warehouse and establishing a data lineage tracking mechanism includes:

[0015] Based on the source of the hydropower plant's multi-dimensional data, metadata is generated; naming rules are formulated, a hierarchical structure of metadata tags is constructed, and metadata tags are entered; the collection path of each type of metadata from the data source to the data warehouse is obtained, and the collection timestamp is marked;

[0016] Record the storage location of the hydropower plant's multi-dimensional data in the data warehouse, specifically the file path in the data lake; record the access to the hydropower plant's multi-dimensional data, specifically which users or systems operated on which hydropower plant's multi-dimensional data and when; draw a data flow diagram by analyzing the flow of the hydropower plant's multi-dimensional data within the data warehouse and between it and external systems.

[0017] In combination with the first aspect, in a third implementation of the first aspect of the present application, determining the tracing starting point, tracing back based on the storage location, tracing back the data access history, tracing back according to the data flow graph, and integrating to obtain the data tracing chain include:

[0018] The user inputs the data feature to be traced, and based on the data feature to be traced, the data type and data identification information of the data feature specified by the user are determined;

[0019] Based on the determined data type and identification information, a search is performed in the file path records of the data lake. When the data being searched is historical backup data of equipment operation data, the specific file storing the data is located according to the pre-recorded file naming rules and the file path structure in the data lake. In the log table that records the access status of the hydropower plant's multi-dimensional data, all access records of the located data are searched using the storage location of the located data as an index. Based on the queried operation records, the impact of each operation on the data is analyzed.

[0020] Based on the data flow diagram, starting from the located data, determine the path of data flowing from the data source to the data warehouse; follow the data inflow path back to the data collection stage;

[0021] All information obtained during storage location backtracking, data access history tracing, and backtracking based on the data flow graph will be integrated to form a data traceability chain.

[0022] In combination with the first aspect, in a fourth implementation of the first aspect of the present application, entity recognition and relationship extraction are performed based on the data warehouse and data traceability chain, and a relationship graph is constructed using Neo4j, including:

[0023] Extract data related to equipment, sensors, and failure modes from the data warehouse, identify equipment entities based on the unique identifiers of sensor monitoring data, and extract relevant attributes for each equipment entity; identify sensor entities based on sensor monitoring data, and record the attributes of each sensor entity; analyze the fault record data in the equipment operation data and sensor monitoring data to identify different fault mode entities. The attributes of each fault mode entity include fault name, fault phenomenon description, fault severity level, and fault occurrence frequency;

[0024] By analyzing the association between device operation data and sensor monitoring data, as well as the installation configuration information of the device and the sensor, the relationship between the device and the sensor is extracted; combining the device operation data and the fault record data, the device with a fault mode exceeding a set threshold is extracted to obtain the relationship between the device and the fault mode; installing and configuring the Neo4j graph database, and creating the identified device entities, sensor entities, and fault mode entities as nodes in Neo4j; when creating the node, the extracted corresponding attributes are added to the node; based on the extracted relationship between the device and the sensor, the device and the fault mode, and the sensor and the fault mode, corresponding edges are created in Neo4j to connect the nodes, and each edge is labeled with the relationship type and related attributes.

[0025] In conjunction with the first aspect, in a fifth implementation of the first aspect of the present application, analyzing the relationship between power grid dispatch data and equipment operation, including instruction impact analysis and chain reaction analysis, and adjusting the association map, includes:

[0026] The method for analyzing the impact of instructions includes: using regression analysis algorithms to establish a mathematical model between grid dispatch instructions and changes in equipment operating parameters; through model analysis and actual data verification, determining the critical paths and key parameters of the impact of grid dispatch instructions on equipment operation;

[0027] The chain reaction analysis method includes: deriving the relationships between devices based on the physical connections and operating logic of hydropower plant equipment; simulating the chain reaction process by combining the direct impact of grid dispatch instructions on device operation; and analyzing the direct impact of power generation adjustment instructions on generators when the grid issues them. Based on the relationships between devices, the method gradually deduces how the resulting impact is transmitted to other devices.

[0028] Based on the results of instruction impact analysis and chain reaction analysis, new relationship nodes and edges related to power grid scheduling are added to the existing association graph; for indirect relationships between devices involved in chain reactions, corresponding edges are added to the association graph to connect them and annotate relationship attributes. The relationship attributes are information used to describe the characteristics and properties of the relationships between nodes in the association graph.

[0029] In combination with the first aspect, in a sixth implementation of the first aspect of the present application, the method of detecting anomalies based on the device operation data and the sensor monitoring data using an isolation forest to obtain explicit anomalies, reconstructing the data distribution through an autoencoder, and identifying implicit anomalies includes:

[0030] The scikit-learn library in Python was selected as the isolation forest algorithm library to initialize the isolation forest model. The multi-dimensional data of the hydropower plant was input into the isolation forest model for training. Multiple isolation trees were constructed by randomly selecting data subsets. Each tree partitioned the data points until every data point was isolated. After training, the equipment operation data and sensor monitoring data to be tested were input into the isolation forest model. The isolation forest model calculated an anomaly score for each data point. Based on the pre-set anomaly score threshold, the data point was determined to be anomaly. When the anomaly score was higher than the threshold, the data point was marked as an explicit anomaly and the relevant information was recorded.

[0031] Design the model architecture of the autoencoder, including the encoder and decoder; use a multi-layer perceptron as the basic structure of the encoder and decoder, and optimize performance by adjusting the number of hidden layers and neurons; use multi-dimensional data from a hydropower plant to train the autoencoder. During the training process, adjust the parameters by minimizing the reconstruction error so that the data reconstructed by the decoder is close to the original input data; after training, input the equipment operation data and sensor monitoring data to be tested into the autoencoder, reconstruct the data, and calculate the reconstruction error; set a reconstruction error threshold, and when the reconstruction error of the data exceeds the threshold, determine that the data has hidden anomalies.

[0032] In combination with the first aspect, in a seventh implementation of the first aspect of the present application, fusing explicit anomalies and implicit anomalies based on a correlation graph to obtain abnormal data of the hydropower plant includes:

[0033] Map abnormally associated device entities to device nodes in the association graph, map abnormally associated sensor entities to sensor nodes in the association graph, and add historical failure modes as predefined nodes to the graph; create explicit abnormality edges between device nodes and sensor nodes and explicit abnormalities, annotate the abnormality score and type; create implicit abnormality edges between device nodes and sensor nodes and implicit abnormalities, annotate the reconstruction error and trend characteristics;

[0034] Analyze explicit and implicit anomalies of the same device or sensor within the same time period. When the time windows overlap and the parameters are related, mark them as composite anomalies. Calculate the similarity of the explicit and implicit anomalies in the vibration spectrum and temperature gradient, specifically by calculating the cosine similarity. When the similarity exceeds the set threshold, they are determined to be different manifestations of the same fault event.

[0035] Neo4j's Cypher language is used to query abnormal propagation paths, and path weights are assigned based on the strength of the physical connection between devices to obtain high-probability propagation chains. Based on the high-probability propagation chains, when the same device has multiple explicit abnormalities associated, it is determined to be a high-priority fault. When the same device has explicit abnormalities accompanied by implicit abnormalities, it is determined to be a compound fault. When the same device has multiple implicit abnormalities forming a closed loop in the association map, it is determined to be a systemic degradation. Abnormal data integration and dynamic priority sorting are performed to obtain abnormal data of the hydropower plant.

[0036] In conjunction with the first aspect, in an eighth implementation of the first aspect of the present application, predicting trends of equipment operation data and sensor monitoring data using a Prophet model, locating deviation values ​​in combination with residual analysis, obtaining data fluctuation patterns based on the deviation values, analyzing the correlation between abnormal data from the hydropower plant and the data fluctuation patterns, and obtaining the source of the abnormal data from the hydropower plant include:

[0037] Configure a Prophet model and train it using the similarities between explicit and implicit anomalies in vibration spectra and temperature gradients to generate trend, seasonal, and cyclical characteristics of the anomaly data. After training, use the trained Prophet model to predict equipment operation data and sensor monitoring data for a period of time. The prediction results include trend values ​​for the equipment operation data and sensor monitoring data.

[0038] The actual equipment operation data and sensor monitoring data are compared with the trend value predicted by the Prophet model to calculate the residual. The residual is equal to the actual value minus the predicted value, reflecting the difference between the actual data and the model prediction result. Based on the calculated residual, the deviation value is located using the set threshold. When the absolute value of the residual exceeds the threshold, the data point is determined to be a deviation value.

[0039] The located deviation values ​​are analyzed, and based on the variation characteristics and distribution of the deviation values, the data fluctuation patterns are classified into periodic fluctuations, sudden fluctuations, and persistent deviations. A correlation analysis is performed between the abnormal data of the hydropower plant and the data fluctuation patterns, and the degree of association between the abnormal data and different fluctuation patterns is determined by calculating the correlation coefficient. Based on the degree of association and combined with the historical failure records of the hydropower plant equipment, the source of the abnormal data of the hydropower plant is inferred.

[0040] In a second aspect, the present invention provides a hydropower plant intelligent operation and maintenance management system based on machine learning, comprising:

[0041] Data collection and data backtracking module: includes: data collection unit, data lineage tracking unit and data backtracking unit; the data collection unit collects multi-dimensional data of the hydropower plant, including equipment operation data, sensor monitoring data and grid dispatch data, performs data preprocessing, and stores it in the data warehouse; the data lineage tracking unit defines unified metadata tags in the data warehouse and establishes a data lineage tracking mechanism; the data backtracking unit determines the traceability starting point, traces back based on the storage location, traces back the data access history, and traces back according to the data flow diagram to integrate the data traceability chain;

[0042] Association graph generation module: includes an association graph construction unit and an association graph adjustment unit. The association graph construction unit performs entity recognition and relationship extraction based on the data warehouse and data traceability chain, and uses Neo4j to build an association graph. The association graph adjustment unit analyzes the relationship between power grid dispatch data and equipment operation, including instruction impact analysis and chain reaction analysis, and adjusts the association graph.

[0043] Hydropower plant anomaly detection module: This module includes an explicit anomaly analysis unit, a latent anomaly analysis unit, and a hydropower plant anomaly data generation unit. The explicit anomaly analysis unit uses an isolation forest to detect anomalies based on equipment operation data and sensor monitoring data to obtain explicit anomalies. The latent anomaly analysis unit reconstructs data distribution through an autoencoder to identify latent anomalies. The hydropower plant anomaly data generation unit fuses explicit and latent anomalies based on a correlation graph to obtain hydropower plant anomaly data.

[0044] The hydropower plant abnormal data source analysis module includes: a deviation value positioning unit and a hydropower plant abnormal data source analysis unit; wherein, the deviation value positioning unit predicts the trend of equipment operation data and sensor monitoring data through the Prophet model, and locates the deviation value in combination with residual analysis. The hydropower plant abnormal data source analysis unit obtains the data fluctuation pattern based on the deviation value, analyzes the correlation between the hydropower plant abnormal data and the data fluctuation pattern, and obtains the source of the hydropower plant abnormal data.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] 1. The present invention establishes a unified data warehouse by collecting multi-dimensional data of hydropower plants and performing efficient data preprocessing and storage. At the same time, it defines unified metadata tags and establishes a data lineage tracking mechanism, which can achieve comprehensive management and traceability of data.

[0047] 2. Based on the data warehouse and data traceability chain, the present invention performs entity recognition and relationship extraction, uses Neo4j to construct a correlation graph, and deeply analyzes the relationship between power grid dispatching data and equipment operation, which makes it possible to grasp the intrinsic connection between the multi-dimensional data of the hydropower plant from a holistic perspective.

[0048] 3. The present invention uses isolation forests and autoencoders to detect explicit anomalies and implicit anomalies respectively, and fuses the two based on the correlation map to obtain comprehensive and accurate abnormal data of hydropower plants; at the same time, the Prophet model is used to predict the trends of equipment operation data and sensor monitoring data, and the deviation values ​​are located by combining residual analysis. The data fluctuation pattern is obtained based on the deviation value, and the correlation between the abnormal data and the data fluctuation pattern is analyzed, which can accurately determine the source of the abnormal data. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 This is a schematic diagram of the steps of a method for intelligent operation and maintenance management of a hydropower plant based on machine learning according to the present invention;

[0050] Figure 2 This is a system structure diagram of a hydropower plant intelligent operation and maintenance management system based on machine learning in the present invention. DETAILED DESCRIPTION

[0051] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0052] Example: Figure 1-Figure 2 As shown, the present invention provides a technical solution.

[0053] like Figure 1 The present invention provides a method for intelligent operation and maintenance management of a hydropower plant based on machine learning, as shown in a schematic diagram of the steps. The method comprises the following steps:

[0054] Step S100: Collect multi-dimensional data from the hydropower plant, including equipment operation data, sensor monitoring data, and grid dispatch data, perform data preprocessing, and store it in a data warehouse. In the data warehouse, define unified metadata tags and establish a data lineage tracking mechanism. Determine the traceability starting point, trace back based on the storage location, trace the data access history, and trace back based on the data flow diagram to integrate and obtain the data traceability chain.

[0055] Specifically, the system collects equipment operating data, including hydro-generator sets, transformers, and switchgear, at a set collection frequency and transmits it to a data collection server via a fieldbus or wireless network. Sensors monitor the equipment and environmental status in real time, converting the collected sensor monitoring data into digital signals for transmission. The sensors include environmental sensors, electrical sensors, and mechanical sensors. A data connection is established with the power grid dispatching system, and the latest power grid dispatching data, including dispatching instructions and power grid operation data, is obtained according to the update frequency of the power grid dispatching system.

[0056] The data preprocessing includes data cleaning and data conversion. The data warehouse is built using a data lake, wherein the data lake adopts Delta Lake.

[0057] Furthermore, metadata is generated based on the source of the hydropower plant's multi-dimensional data; naming rules are formulated, a hierarchical structure of metadata tags is constructed, and metadata tags are entered; the collection path of each type of metadata from the data source to the data warehouse is obtained, and the collection timestamp is marked;

[0058] Record the storage location of the hydropower plant's multi-dimensional data in the data warehouse, specifically the file path in the data lake; record the access to the hydropower plant's multi-dimensional data, specifically which users or systems operated on which hydropower plant's multi-dimensional data and when; draw a data flow diagram by analyzing the flow of the hydropower plant's multi-dimensional data within the data warehouse and between it and external systems.

[0059] Furthermore, the user inputs the data feature to be traced, and based on the data feature to be traced, the data type and data identification information of the data feature specified by the user are determined;

[0060] Based on the determined data type and identification information, a search is performed in the file path records of the data lake. When the data being searched is historical backup data of equipment operation data, the specific file storing the data is located according to the pre-recorded file naming rules and the file path structure in the data lake. In the log table that records the access status of the hydropower plant's multi-dimensional data, all access records of the located data are searched using the storage location of the located data as an index. Based on the queried operation records, the impact of each operation on the data is analyzed.

[0061] Based on the data flow diagram, starting from the located data, determine the path of data flowing from the data source to the data warehouse; follow the data inflow path back to the data collection stage;

[0062] All information obtained during storage location backtracking, data access history tracing, and backtracking based on the data flow graph will be integrated to form a data traceability chain.

[0063] In one specific embodiment, parameters such as the speed, power, voltage, and current of a hydro-turbine generator set are collected at a 5-minute frequency. For example, during a monitoring session on a particular day, the speed of hydro-turbine generator set A was 150 rpm, the power was 5000 kW, the voltage was 10 kV, and the current was 500 A at 12:00 PM. Similarly, data such as oil temperature, winding temperature, and load factor are collected at 5-minute intervals for transformers. For example, at 1:00 PM for transformer T1, the oil temperature was 45°C, the winding temperature was 50°C, and the load factor was 70%. Data collected for switchgear, including switch status, bus voltage, and branch current, is updated every 5 minutes. For example, at 2:00 PM for switchgear S3, the switch was closed, the bus voltage was 10.5 kV, and the branch current was 300 A. This data is transmitted to the data acquisition server via the fieldbus.

[0064] Environmental sensors collect ambient temperature and humidity data every 10 minutes. On a particular day, at 3:00 PM, the ambient temperature sensor indicated a temperature of 30°C, and the humidity sensor indicated a humidity of 60%. Electrical sensors, such as current transformers and voltage transformers, monitor the electrical parameters of equipment in real time, converting analog signals into digital signals for transmission. For example, an electrical sensor detected a current value of 450A at 4:00 PM, and the converted digital signal value was 0x01C2. Mechanical sensors monitor mechanical conditions such as vibration and displacement. For example, a vibration sensor collects data every two minutes. At 5:00 PM, the vibration sensor of hydro-turbine generator unit B recorded a vibration amplitude of 0.5 mm / s.

[0065] A data connection was established with the power grid dispatch system, and the latest data was obtained every 15 minutes. On one particular day, at 6:00 PM, the dispatch instruction received was to increase the power generation capacity of the hydro-turbine generator to 6,000 kW. At the same time, the grid operating data showed that the current grid frequency was 50.1 Hz and the voltage was 10.2 kV.

[0066] Perform data cleaning and data conversion.

[0067] Metadata is generated based on the multi-dimensional data sources of the hydropower plant. For example, for the speed data of a hydro-turbine generator set, the metadata includes information such as the data source being the speed sensor of hydro-turbine generator set A, the data type being numeric, and the unit being r / min. Naming rules are established, such as using the "Device Name_Parameter Name" format for metadata tags for equipment operation data. A hierarchical structure is constructed, with the top layer representing hydropower plant data and sub-layers such as equipment operation data, sensor monitoring data, and grid dispatch data. Metadata tags are entered, such as "Hydro-turbine Generator Set A_Speed" and "Ambient Temperature Sensor_Temperature." The collection path for each metadata type is obtained. For example, ambient temperature sensor data travels from the sensor through a wireless transmission module, passes through a gateway, enters the data collection server, and is ultimately stored in the data lake. The collection timestamp is marked. For example, data collected by the ambient temperature sensor at 3:00 PM is recorded with the timestamp "2024-08-01 15:00:00." Data access is recorded in a log table. For example, at 7:00 PM, the data analysis system read the power data of hydro-turbine generator set A under this file path. Draw a data flow diagram to show how data flows from sensors and devices to data collection servers, then to the data lake, and between external systems. For example, power grid dispatch data flows from the power grid dispatch system into the data lake through a secure interface. The data analysis system reads and analyzes device operation data from the data lake and transmits the results to the operation and maintenance management system.

[0068] The user inputs the data feature to be traced, "The power data of hydropower generator unit A at 12:00 PM on August 1, 2024, is abnormal." Based on this feature, the data type is determined to be equipment operation data, with the identification information "hydropower generator unit A" and "2024-08-01 12:00 PM." A search within the file path records in the data lake locates the file " / data / hydropower / equipment / 20240801 / water_turbine_generator_A.parquet" storing the data. Using the file path as an index in the log table, a query reveals that the data processing program performed normalization on the data at 1:00 PM. Analysis of this operation record identifies the normalization algorithm and parameters, confirming that this operation did not cause data anomalies. Using the data flow diagram, the data's path from sensor acquisition to entry into the data lake is traced back. It is found that the power data is collected by hydropower generator unit A's power sensor, transmitted via the fieldbus to the data acquisition server, and then stored in the data lake. When looking back to the data collection link, the sensor's working status records were checked and it was found that the power sensor had a brief signal interference between 11:50 and 12:10. However, after data cleaning, the interference data had been processed, and the abnormality was preliminarily ruled out as a result of sensor failure.

[0069] Integrate all the above information to form a data traceability chain: hydro-turbine generator unit A power sensor collects data → field bus transmission → data acquisition server receives → data cleaning (processing interference data) → standardization processing → storage in the data lake (file path: / data / hydropower / equipment / 20240801 / water_turbine_generator_A.parquet).

[0070] Step S200: Based on the data warehouse and data traceability chain, perform entity recognition and relationship extraction, and use Neo4j to build a correlation map; analyze the relationship between power grid dispatch data and equipment operation, including instruction impact analysis and chain reaction analysis, and adjust the correlation map;

[0071] Specifically, data related to equipment, sensors, and failure modes are extracted from the data warehouse. Equipment entities are identified based on the unique identifiers of sensor monitoring data, and relevant attributes are extracted for each equipment entity. Sensor entities are identified based on sensor monitoring data, and their attributes are recorded for each sensor entity. Fault record data in equipment operation data and sensor monitoring data are analyzed to identify different failure mode entities. The attributes of each failure mode entity include the failure name, failure phenomenon description, failure severity level, and failure frequency.

[0072] By analyzing the association between device operation data and sensor monitoring data, as well as the installation configuration information of the device and the sensor, the relationship between the device and the sensor is extracted; combining the device operation data and the fault record data, the device with a fault mode exceeding a set threshold is extracted to obtain the relationship between the device and the fault mode; installing and configuring the Neo4j graph database, and creating the identified device entities, sensor entities, and fault mode entities as nodes in Neo4j; when creating the node, the extracted corresponding attributes are added to the node; based on the extracted relationship between the device and the sensor, the device and the fault mode, and the sensor and the fault mode, corresponding edges are created in Neo4j to connect the nodes, and each edge is labeled with the relationship type and related attributes.

[0073] Furthermore, the method of command impact analysis includes: using regression analysis algorithm to establish a mathematical model between power grid dispatching commands and changes in equipment operating parameters; through model analysis and actual data verification, determining the key paths and key parameters that affect the operation of equipment by power grid dispatching commands;

[0074] The chain reaction analysis method includes: deriving the relationships between devices based on the physical connections and operating logic of hydropower plant equipment; simulating the chain reaction process by combining the direct impact of grid dispatch instructions on device operation; and analyzing the direct impact of power generation adjustment instructions on generators when the grid issues them. Based on the relationships between devices, the method gradually deduces how the resulting impact is transmitted to other devices.

[0075] Based on the results of instruction impact analysis and chain reaction analysis, new relationship nodes and edges related to power grid scheduling are added to the existing association graph; for indirect relationships between devices involved in chain reactions, corresponding edges are added to the association graph to connect them and annotate relationship attributes. The relationship attributes are information used to describe the characteristics and properties of the relationships between nodes in the association graph.

[0076] In one specific embodiment, detailed data for equipment such as hydroelectric generator sets, transformers, and switchgear is retrieved from a data warehouse. For example, the device number "WTG-003" accurately locates a hydroelectric generator set. Its attributes are rich and diverse. The device type is clearly a hydroelectric generator set, belonging to a core component of the power generation system, with a rated power of 12,000 kW. It is manufactured by XY Company and installed at unit 3 in the plant. Furthermore, information such as the equipment's commissioning date (2020-01-01) and the last maintenance date (2024-02-15) are extracted. These attributes are crucial for a comprehensive understanding of the equipment's status.

[0077] Based on the unique numbers in the sensor monitoring data, various sensor entities were successfully identified. Taking the pressure sensor numbered "PS-012" as an example, its attributes cover multiple aspects. The sensor type is a pressure sensor, primarily used to monitor the water pressure at the turbine inlet, with a measurement range of 0-5 MPa and an accuracy of ±0.05 MPa. It is installed near the valve in the turbine inlet pipe to ensure accurate sensing of water pressure changes. The data transmission frequency is set to once every 3 minutes, and the wireless transmission module transmits the data stably to the data acquisition server.

[0078] In-depth analysis of fault records from equipment operating data and sensor monitoring data successfully identified multiple fault mode entities. Taking the "transformer winding short-circuit fault" as an example, the attributes of this fault mode entity provide detailed information about the fault. The fault is named "transformer winding short-circuit fault," and the symptoms are described as a sharp increase in transformer oil temperature, accompanied by abnormal noise, and unstable fluctuations in output voltage and current. The fault severity is rated "extremely severe" because, if not addressed promptly, this type of fault could cause the transformer to burn out, seriously impacting the hydropower plant's normal power generation. Historical data analysis shows that this fault has occurred twice in the past two years.

[0079] By carefully analyzing the inherent correlations between equipment operating data and sensor monitoring data, and combining the equipment and sensor installation configuration drawings with actual recorded information, we successfully extracted the close relationships between the equipment and sensors. For example, there is an "Installed With" relationship between the hydro-turbine generator set WTG-003 and the pressure sensor PS-012. This relationship clearly indicates that the hydro-turbine generator set WTG-003 is equipped with the pressure sensor PS-012 to monitor the water pressure at the turbine inlet in real time, providing critical data support for the stable operation of the equipment.

[0080] By comprehensively considering equipment operating data and fault history data and setting appropriate fault thresholds, we were able to accurately identify the association between equipment and failure modes. In-depth analysis of historical data revealed that turbine generator unit WTG-003 had experienced four "turbine blade cavitation failures" over the past two years, exceeding the established threshold by three. Therefore, a "prone to failure" relationship was established between equipment WTG-003 and the "turbine blade cavitation failure" failure mode.

[0081] When creating the hydro-turbine generator WTG-003 node, add all previously extracted attributes, such as device type, system, rated power, manufacturer, installation location, commissioning time, and last maintenance time, to the node. When creating the pressure sensor PS-012 node, similarly add its sensor type, measurement range, accuracy, installation location, and data transmission frequency.

[0082] Based on the extracted relationships, corresponding edges are created in Neo4j to connect related nodes. An edge with an "Installed With" relationship is created between the hydro-turbine generator set WTG-003 node and the pressure sensor PS-012 node, and the relationship attributes are labeled in detail. For example, the relationship confidence is set to 100%, indicating that this installation relationship is unambiguous; the installation time can also be labeled as "2020-01-05" to record the installation time of the sensor and provide a time clue for subsequent maintenance and troubleshooting. An edge with a "Prone to Failure" relationship is created between the hydro-turbine generator set WTG-003 node and the "Turbine Blade Cavitation Fault" node, and the fault frequency is labeled as "4 times in the past two years." At the same time, relationship attributes such as the time of the first fault occurrence "2023-05-10" are added.

[0083] Grid dispatch command data from the past three months was collected, including power and voltage adjustment commands, as well as corresponding equipment operating parameter changes, such as changes in turbine generator speed, power, voltage, and current. Using a regression analysis algorithm, with the power adjustment command as the independent variable and the turbine generator power change as the dependent variable, a mathematical model was established: P = 0.75I + 1200, where P is the turbine generator power and I is the power adjustment command value. Through repeated training and precise fitting of a large amount of historical data, the model coefficients a = 0.75 and b = 1200 were obtained. Through model analysis and real-world data validation, it was determined that for every 1000kW increase in the power adjustment command, the turbine generator power would increase by approximately 750kW within six minutes. Furthermore, through detailed process analysis and actual monitoring, the critical path was determined to be: the grid dispatch center issues the power adjustment command → the communication network transmits the command to the hydropower plant control system → the turbine governor receives the command and adjusts the guide vane opening → the turbine speed changes accordingly → the generator output power changes accordingly. Key parameters include power generation adjustment instruction value, turbine guide vane opening, turbine speed, generator output power, etc.

[0084] Based on detailed physical connection diagrams of hydropower plant equipment and in-depth analysis of operational logic, the complex relationships between devices were comprehensively identified. For example, the generator and transformer are tightly connected via high-voltage transmission cables, and the transformer and switchgear are connected via busbars, forming a complete power transmission chain. When the power grid issues a power adjustment command, the direct impact of this command on the generator is thoroughly analyzed. When the power grid issues a power increase of 3000kW, the generator's output power rapidly increases by 2250kW within 6 minutes, and the speed rises by 6%.

[0085] Based on the interdependencies between devices, the chain reaction process is deduced step by step: a 2250kW increase in generator output power leads to an increase in input torque and a 6% increase in speed. This change is transmitted to the turbine through the coupling, requiring an increase in water flow to maintain the new speed, which in turn causes a change in water pressure at the turbine inlet. The changes in generator output power and voltage are transmitted to the transformer via the high-voltage transmission cable. The transformer's input power increases by 2250kW, and its input voltage rises by 4%. After the transformer's voltage transformation, the output voltage drops by 3%, while the output current increases by 12%. These changes in the transformer's output voltage and current are transmitted to the switchgear via the busbars. The switchgear's busbar voltage drops by 3%, while the branch current increases by 12%, increasing the load on some switchgear and causing a slight temperature increase.

[0086] Step S300: Based on the equipment operation data and sensor monitoring data, an isolation forest is used to detect anomalies to obtain explicit anomalies. The data distribution is reconstructed through an autoencoder to identify implicit anomalies. The explicit and implicit anomalies are fused based on the correlation graph to obtain abnormal data of the hydropower plant.

[0087] Specifically, the scikit-learn library in Python was selected as the isolation forest algorithm library to initialize the isolation forest model. Multi-dimensional data from the hydropower plant was input into the isolation forest model for training. Multiple isolation trees were constructed by randomly selecting data subsets. Each tree partitioned the data points until each data point was isolated. After training, the equipment operation data and sensor monitoring data to be tested were input into the isolation forest model. The isolation forest model calculated an anomaly score for each data point. Based on a pre-set anomaly score threshold, the data point was determined to be anomaly. When the anomaly score was higher than the threshold, the data point was marked as an explicit anomaly and the relevant information was recorded.

[0088] Design the model architecture of the autoencoder, including the encoder and decoder; use a multi-layer perceptron as the basic structure of the encoder and decoder, and optimize performance by adjusting the number of hidden layers and neurons; use multi-dimensional data from a hydropower plant to train the autoencoder. During the training process, adjust the parameters by minimizing the reconstruction error so that the data reconstructed by the decoder is close to the original input data; after training, input the equipment operation data and sensor monitoring data to be tested into the autoencoder, reconstruct the data, and calculate the reconstruction error; set a reconstruction error threshold, and when the reconstruction error of the data exceeds the threshold, determine that the data has hidden anomalies.

[0089] Furthermore, the device entities associated with the anomaly are mapped to the device nodes in the association graph, the sensor entities associated with the anomaly are mapped to the sensor nodes in the association graph, and the historical failure modes are added to the graph as predefined nodes. Explicit anomaly edges are created between the device nodes and sensor nodes and the explicit anomalies, and the anomaly scores and types are annotated. Implicit anomaly edges are created between the device nodes and sensor nodes and the implicit anomalies, and the reconstruction errors and trend characteristics are annotated.

[0090] Analyze explicit and implicit anomalies of the same device or sensor within the same time period. When the time windows overlap and the parameters are related, mark them as composite anomalies. Calculate the similarity of the explicit and implicit anomalies in the vibration spectrum and temperature gradient, specifically by calculating the cosine similarity. When the similarity exceeds the set threshold, they are determined to be different manifestations of the same fault event.

[0091] Neo4j's Cypher language is used to query abnormal propagation paths, and path weights are assigned based on the strength of the physical connection between devices to obtain high-probability propagation chains. Based on the high-probability propagation chains, when the same device has multiple explicit abnormalities associated, it is determined to be a high-priority fault. When the same device has explicit abnormalities accompanied by implicit abnormalities, it is determined to be a compound fault. When the same device has multiple implicit abnormalities forming a closed loop in the association map, it is determined to be a systemic degradation. Abnormal data integration and dynamic priority sorting are performed to obtain abnormal data of the hydropower plant.

[0092] In one specific embodiment, an isolation forest model was initialized in Python using the scikit-learn library. The number of trees (n_estimators) was set to 200, and the subsample size (max_samples) was set to 0.8 times the total number of samples. The extracted multidimensional data from the hydropower plant was fed into the isolation forest model for training. During training, the model randomly selected a subset of the data to construct 200 isolation trees. Each tree recursively partitioned the data points until every data point was isolated. After training was complete, the data to be tested was fed in. For example, given the operating data of a hydro-turbine generator set at a certain moment, the speed is 145 rpm (normal range: 150 ± 5 rpm), the power is 4800 kW (normal range: 5000 ± 200 kW), the voltage is 9.8 kV (normal range: 10 ± 0.5 kV), and the current is 490 A (normal range: 500 ± 30 A). The corresponding temperature sensor measurement value is 42°C (normal range: 40 ± 3°C), and the vibration sensor measurement value is 0.6 mm / s (normal range: 0.5 ± 0.1 mm / s). The isolation forest model calculates anomaly scores for these data points, such as 0.8 for the speed data point, 0.9 for the power data point, and 0.7 for the temperature sensor data point. The anomaly score threshold is pre-set to 0.75. The anomaly scores of the speed and power data points are higher than the threshold and are marked as obvious anomalies. Relevant information is recorded, including the time of the anomaly occurrence "2024-08-05 10:00:00", the corresponding device is the hydro-turbine generator set "WTG-001", and the sensors are the speed sensor "RS-001" and the power sensor "PS-001".

[0093] An autoencoder model was designed, with both the encoder and decoder using a multilayer perceptron architecture. The encoder consisted of three hidden layers with 128, 64, and 32 neurons, respectively; the decoder also consisted of three hidden layers with 64 and 128 neurons, respectively, matching the input data dimensions. The autoencoder was trained using multidimensional data from a hydropower plant, and model parameters were adjusted by minimizing the reconstruction error (using a mean squared error loss function). After multiple rounds of training, the model was able to reconstruct the input data well. The data to be tested was then fed into the autoencoder for reconstruction. For example, the operating data for the hydropower generator at another time showed a speed of 152 rpm, a power of 5050 kW, a voltage of 10.1 kV, a current of 502 A, a temperature sensor measurement of 41°C, and a vibration sensor measurement of 0.52 mm / s. The reconstruction error was calculated, e.g., 0.04 for the speed data, 0.03 for the power data, and 0.05 for the temperature sensor data. The reconstruction error threshold is set to 0.045. The reconstruction error of the temperature sensor data exceeds the threshold, and the data is determined to have a hidden anomaly. The relevant information is recorded. The anomaly occurred at "2024-08-05 10:30:00", the corresponding device is the hydro-turbine generator set "WTG-001", and the sensor is the temperature sensor "TS-001".

[0094] The device entity "WTG-001" associated with the explicit anomaly is mapped to the hydro-turbine generator device node in the association graph. The associated speed sensor "RS-001" and power sensor "PS-001" are mapped to the sensor node in the graph. The temperature sensor "TS-001" associated with the implicit anomaly is also mapped to the sensor node in the graph. Historical fault modes associated with hydro-turbine generators, such as "hydro-turbine generator blade wear failure," are added to the graph as predefined nodes. Explicit anomaly edges are created between the hydro-turbine generator device node and explicit anomalies in speed and power, annotated with the anomaly score (0.8 for speed and 0.9 for power) and type (device operating parameter anomaly). Implicit anomaly edges are created between the device node and implicit anomalies in the temperature sensor, annotated with the reconstruction error (0.05) and trend characteristics (the temperature data reconstruction error shows a gradual increase).

[0095] The explicit anomalies (speed and power anomalies) and implicit anomalies (temperature anomalies) of the same equipment (hydrogenerator set "WTG-001") in a similar time period (2024-08-05 10:00:00-10:30:00) are analyzed. Since the time windows partially overlap and there is a certain correlation between the speed, power and temperature parameters during equipment operation (for example, power changes may cause the equipment to heat up and thus affect the temperature), they are marked as composite anomalies. The similarity of explicit anomalies and implicit anomalies in the vibration spectrum and temperature gradient is calculated. For example, the vibration spectrum data and temperature gradient data within a period of time are selected and the cosine similarity is calculated. The cosine similarity of the vibration spectrum is 0.85, and the cosine similarity of the temperature gradient is 0.88. The similarity threshold is set to 0.8, and it is determined to be different manifestations of the same fault event.

[0096] Neo4j's Cypher language is used to query anomaly propagation paths. For example, the anomaly propagation path for hydro-turbine generator unit "WTG-001" is queried. Path weights are assigned based on the physical connection strength between devices (e.g., the generator and transformer are connected via high-voltage cables, with a connection strength of 0.9; the transformer and switchgear are connected via busbars, with a connection strength of 0.8). The resulting high-probability propagation chain is: hydro-turbine generator unit "WTG-001" → generator → transformer → switchgear. Based on this high-probability propagation chain, hydro-turbine generator unit "WTG-001" is identified as a high-priority fault because it has multiple explicit anomalies (speed and power anomalies) associated with them. Furthermore, this equipment is identified as a complex fault because the explicit anomalies (speed and power anomalies) are accompanied by implicit anomalies (temperature anomalies). If subsequent monitoring reveals that multiple implicit anomalies for the same equipment form a closed loop in the correlation graph (e.g., implicit anomalies from multiple sensors, such as temperature, pressure, and vibration, forming a closed loop), it is identified as systemic degradation.

[0097] All detected abnormal data is integrated and dynamically prioritized by high-priority faults, complex faults, systemic degradation, and other common faults to generate the hydropower plant's abnormal data. For example, at a certain moment, abnormal data from hydro-turbine generator unit "WTG-001" was ranked at the top of the abnormal data list due to its high-priority and complex fault characteristics, making it easier for operations and maintenance personnel to prioritize processing.

[0098] Step S400: Use the Prophet model to predict the trends of equipment operation data and sensor monitoring data, combine residual analysis to locate deviation values, obtain data fluctuation patterns based on the deviation values, analyze the correlation between the abnormal data of the hydropower plant and the data fluctuation patterns, and obtain the source of the abnormal data of the hydropower plant.

[0099] Specifically, a Prophet model is configured and trained using the similarities between explicit and implicit anomalies in vibration spectra and temperature gradients to generate trend, seasonal, and cyclical characteristics of the anomaly data. After training, the trained Prophet model is used to predict equipment operation data and sensor monitoring data for a period of time. The prediction results include trend values ​​for the equipment operation data and sensor monitoring data.

[0100] The actual equipment operation data and sensor monitoring data are compared with the trend value predicted by the Prophet model to calculate the residual. The residual is equal to the actual value minus the predicted value, reflecting the difference between the actual data and the model prediction result. Based on the calculated residual, the deviation value is located using the set threshold. When the absolute value of the residual exceeds the threshold, the data point is determined to be a deviation value.

[0101] The located deviation values ​​are analyzed, and based on the variation characteristics and distribution of the deviation values, the data fluctuation patterns are classified into periodic fluctuations, sudden fluctuations, and persistent deviations. A correlation analysis is performed between the abnormal data of the hydropower plant and the data fluctuation patterns, and the degree of association between the abnormal data and different fluctuation patterns is determined by calculating the correlation coefficient. Based on the degree of association and combined with the historical failure records of the hydropower plant equipment, the source of the abnormal data of the hydropower plant is inferred.

[0102] In a specific embodiment, the Prophet model is configured in Python using the fbprophet library. For explicit abnormal data, the seasonal cycle is set to 7 days a week and 30 days a month to capture possible weekly and monthly variation patterns; the smoothing parameter of trend change is set to 0.1, so that the model is more sensitive to trend changes. For implicit abnormal data, according to its characteristics, the seasonal cycle is set to 24 hours a day, and changepoint_prior_scale is set to 0.05 to capture potential trend changes in more detail. The similarity feature data of the explicit abnormal data in the vibration spectrum and temperature gradient are input into the Prophet model for training. During the training process, the model learns the trend characteristics of the explicit abnormal data in the time series, such as when the power is abnormal, the change trend of the frequency band where the vibration spectrum energy is concentrated presents a gradually rising trend; seasonal characteristics, such as the amplitude of the abnormal change of the temperature gradient is relatively large on weekends; periodic characteristics, such as the abnormal pattern of the vibration spectrum will be repeated to a certain extent every 10 days or so. Similarly, by training on implicit abnormal data, the model can grasp the unique characteristics of implicit abnormal data in time series. For example, the reconstruction error of temperature sensor data has a more obvious periodic change between 2 and 4 am every day.

[0103] The trained Prophet model is used to predict equipment operation data and sensor monitoring data for the coming week. For example, for the speed of a hydro-turbine generator set, the model predicts that the speed trend will fluctuate within a small range of around 150 r / min over the next week. The speed will be slightly higher during the daytime than at night from Monday to Friday, showing some seasonal variation. For temperature sensor data, the model predicts that the temperature will fluctuate with ambient temperature over the next week. Furthermore, considering the heat generated by equipment operation, the temperature will rise during periods of high equipment load.

[0104] Compare actual equipment operating data and sensor monitoring data with the trend values ​​predicted by the Prophet model to calculate residuals. For example, at 2:00 PM on a certain day, the actual turbine generator speed was 148 rpm, while the model predicted a trend value of 150 rpm. The residual was 148 - 150 = -2 rpm. The actual temperature sensor measurement was 43°C, while the model predicted a trend value of 41°C. The residual was 43 - 41 = 2°C. Set the residual absolute value threshold to 1.5. For speed data, if the residual absolute value |-2| is greater than 1.5, the data point is considered an outlier. For temperature data, if the residual absolute value |2| is greater than 1.5, the data point is also considered an outlier. Record relevant information about these outliers, including the time, corresponding device, and sensor.

[0105] Analyze the located deviations. For example, if a series of deviations show regular changes with a 7-day cycle, it is determined to be a periodic fluctuation pattern. Other deviations suddenly appear at a certain moment and differ greatly from the previous and subsequent data, which is determined to be a sudden fluctuation pattern. Some deviations are continuously above or below the normal range for a period of time, indicating a persistent deviation pattern. For the periodic fluctuation pattern, features such as a 7-day period and an amplitude of ±2 r / min (using the rotational speed as an example) are extracted. For the sudden fluctuation pattern, features such as the sudden time point "2024-08-10 16:00:00" and the degree of deviation exceeding the normal range by 20% (using a certain sensor data as an example) are recorded. For the persistent deviation pattern, features such as the deviation start time "2024-08-05 08:00:00" are determined, the duration has reached 5 days, and the direction of the deviation is continuously above the normal range.

[0106] A correlation analysis was conducted between the hydropower plant's existing abnormal data and data fluctuation patterns. For example, for the power abnormality data of a certain type of hydro-turbine generator set, the correlation coefficient with the periodic fluctuation pattern was calculated to be 0.85, the correlation coefficient with the sudden fluctuation pattern was 0.2, and the correlation coefficient with the persistent deviation pattern was 0.1. This indicates that this type of power abnormality data has a strong correlation with the periodic fluctuation pattern. Combined with the historical fault records of the hydropower plant's equipment, it was found that when the blades of the hydro-turbine generator set show slight wear, its operating data will show a periodic fluctuation pattern, which is similar to the fluctuation pattern of the currently detected power abnormality data. Further inspection of the equipment maintenance records revealed that the hydro-turbine generator set had undergone a blade inspection six months ago, after which power abnormalities gradually emerged. Based on comprehensive judgment, it is inferred that the source of this power abnormality data is the gradual wear of the hydro-turbine generator set blades.

[0107] like Figure 2 As shown in the system structure diagram of a hydropower plant intelligent operation and maintenance management system based on machine learning, the present invention provides a hydropower plant intelligent operation and maintenance management system based on machine learning, including:

[0108] Data collection and data backtracking module: includes: data collection unit, data lineage tracking unit and data backtracking unit; the data collection unit collects multi-dimensional data of the hydropower plant, including equipment operation data, sensor monitoring data and grid dispatch data, performs data preprocessing, and stores it in the data warehouse; the data lineage tracking unit defines unified metadata tags in the data warehouse and establishes a data lineage tracking mechanism; the data backtracking unit determines the traceability starting point, traces back based on the storage location, traces back the data access history, and traces back according to the data flow diagram to integrate the data traceability chain;

[0109] Association graph generation module: includes an association graph construction unit and an association graph adjustment unit. The association graph construction unit performs entity recognition and relationship extraction based on the data warehouse and data traceability chain, and uses Neo4j to build an association graph. The association graph adjustment unit analyzes the relationship between power grid dispatch data and equipment operation, including instruction impact analysis and chain reaction analysis, and adjusts the association graph.

[0110] Hydropower plant anomaly detection module: This module includes an explicit anomaly analysis unit, a latent anomaly analysis unit, and a hydropower plant anomaly data generation unit. The explicit anomaly analysis unit uses an isolation forest to detect anomalies based on equipment operation data and sensor monitoring data to obtain explicit anomalies. The latent anomaly analysis unit reconstructs data distribution through an autoencoder to identify latent anomalies. The hydropower plant anomaly data generation unit fuses explicit and latent anomalies based on a correlation graph to obtain hydropower plant anomaly data.

[0111] The hydropower plant abnormal data source analysis module includes: a deviation value positioning unit and a hydropower plant abnormal data source analysis unit; wherein, the deviation value positioning unit predicts the trend of equipment operation data and sensor monitoring data through the Prophet model, and locates the deviation value in combination with residual analysis. The hydropower plant abnormal data source analysis unit obtains the data fluctuation pattern based on the deviation value, analyzes the correlation between the hydropower plant abnormal data and the data fluctuation pattern, and obtains the source of the hydropower plant abnormal data.

[0112] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the invention can be embodied in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the invention is defined by the appended claims, not the foregoing description, and all variations within the meaning and range of equivalents of the claims are intended to be included therein. Any reference sign in a claim should not be construed as limiting the claim to which it relates.

Claims

1. A method for intelligent operation and maintenance management of a hydropower plant based on machine learning, characterized in that: The following steps are involved: Collect multi-dimensional data from hydropower plants, including equipment operation data, sensor monitoring data, and grid dispatch data, perform data preprocessing, and store it in a data warehouse. Within the data warehouse, define unified metadata tags and establish a data lineage tracking mechanism. Determine the traceability starting point, trace back based on the storage location, and trace back data access history. This is done based on the data flow diagram and integrated to form a data traceability chain. Based on the data warehouse and data traceability chain, we perform entity recognition and relationship extraction, and use Neo4j to build a correlation map. We analyze the relationship between power grid dispatch data and equipment operation, including instruction impact analysis and chain reaction analysis, and adjust the correlation map. Based on equipment operation data and sensor monitoring data, anomalies are detected using isolation forests to obtain explicit anomalies. The data distribution is reconstructed through autoencoders to identify implicit anomalies. The explicit anomalies and implicit anomalies are fused based on the correlation graph to obtain the abnormal data of the hydropower plant; The Prophet model is used to predict the trends of equipment operation data and sensor monitoring data. The deviation values ​​are located by combining residual analysis. The data fluctuation pattern is obtained based on the deviation values. The correlation between the abnormal data of the hydropower plant and the data fluctuation pattern is analyzed to obtain the source of the abnormal data of the hydropower plant.

2. The method for intelligent operation and maintenance of a hydropower plant based on machine learning according to claim 1 is characterized in that: The multi-dimensional data of the hydropower plant, including equipment operation data, sensor monitoring data and grid dispatch data, is collected, pre-processed and stored in the data warehouse, including: Collect equipment operating data, including hydro-generator sets, transformers, and switchgear, at a set collection frequency and transmit it to a data collection server via a fieldbus or wireless network. Sensors monitor equipment and environmental conditions in real time, converting the collected sensor monitoring data into digital signals for transmission. Establish a data connection with the power grid dispatching system and obtain the latest power grid dispatching data, including dispatching instructions and power grid operation data, based on the grid dispatching system's update frequency. The data preprocessing includes data cleaning and data conversion. The data warehouse is built using a data lake, wherein the data lake adopts Delta Lake.

3. The method for intelligent operation and maintenance of a hydropower plant based on machine learning according to claim 1 is characterized in that: In the data warehouse, unified metadata tags are defined and a data lineage tracking mechanism is established, including: Based on the source of the hydropower plant's multi-dimensional data, metadata is generated; naming rules are formulated, a hierarchical structure of metadata tags is constructed, and metadata tags are entered; the collection path of each type of metadata from the data source to the data warehouse is obtained, and the collection timestamp is marked; Record the storage location of the hydropower plant's multi-dimensional data in the data warehouse, specifically the file path in the data lake; record the access to the hydropower plant's multi-dimensional data, specifically which users or systems operated on which hydropower plant's multi-dimensional data and when; draw a data flow diagram by analyzing the flow of the hydropower plant's multi-dimensional data within the data warehouse and between it and external systems.

4. The method for intelligent operation and maintenance of a hydropower plant based on machine learning according to claim 1 is characterized in that: The process of determining the traceability starting point, tracing back based on the storage location, tracing back the data access history, and tracing back based on the data flow graph to integrate the data traceability chain includes: The user inputs the data feature to be traced, and based on the data feature to be traced, the data type and data identification information of the data feature specified by the user are determined; Based on the determined data type and identification information, a search is performed in the file path records of the data lake. When the data being searched is historical backup data of equipment operation data, the specific file storing the data is located according to the pre-recorded file naming rules and the file path structure in the data lake. In the log table that records the access status of the hydropower plant's multi-dimensional data, all access records of the located data are searched using the storage location of the located data as an index. Based on the queried operation records, the impact of each operation on the data is analyzed. Based on the data flow diagram, starting from the located data, determine the path of data flowing from the data source to the data warehouse; follow the data inflow path back to the data collection stage; All information obtained during storage location backtracking, data access history tracing, and backtracking based on the data flow graph will be integrated to form a data traceability chain.

5. The method for intelligent operation and maintenance of a hydropower plant based on machine learning according to claim 1 is characterized in that: The above mentioned process is based on data warehouse and data traceability chain, and entity recognition and relationship extraction are performed. Neo4j is used to build the association graph, including: Extract data related to equipment, sensors, and failure modes from the data warehouse, identify equipment entities based on the unique identifiers of sensor monitoring data, and extract relevant attributes for each equipment entity; identify sensor entities based on sensor monitoring data, and record the attributes of each sensor entity; analyze the fault record data in the equipment operation data and sensor monitoring data to identify different fault mode entities. The attributes of each fault mode entity include fault name, fault phenomenon description, fault severity level, and fault occurrence frequency; By analyzing the association between device operation data and sensor monitoring data, as well as the installation configuration information of the device and the sensor, the relationship between the device and the sensor is extracted; combining the device operation data and the fault record data, the device with a fault mode exceeding a set threshold is extracted to obtain the relationship between the device and the fault mode; installing and configuring the Neo4j graph database, and creating the identified device entities, sensor entities, and fault mode entities as nodes in Neo4j; when creating the node, the extracted corresponding attributes are added to the node; based on the extracted relationship between the device and the sensor, the device and the fault mode, and the sensor and the fault mode, corresponding edges are created in Neo4j to connect the nodes, and each edge is labeled with the relationship type and related attributes.

6. The method for intelligent operation and maintenance of a hydropower plant based on machine learning according to claim 1, characterized in that: The analysis of the relationship between grid dispatch data and equipment operation, including instruction impact analysis and chain reaction analysis, and adjustment of the correlation map includes: The method for analyzing the impact of instructions includes: using regression analysis algorithms to establish a mathematical model between grid dispatch instructions and changes in equipment operating parameters; through model analysis and actual data verification, determining the critical paths and key parameters of the impact of grid dispatch instructions on equipment operation; The chain reaction analysis method includes: deriving the relationships between devices based on the physical connections and operating logic of hydropower plant equipment; simulating the chain reaction process by combining the direct impact of grid dispatch instructions on device operation; and analyzing the direct impact of power generation adjustment instructions on generators when the grid issues them. Based on the relationships between devices, the method gradually deduces how the resulting impact is transmitted to other devices. Based on the results of instruction impact analysis and chain reaction analysis, new relationship nodes and edges related to power grid scheduling are added to the existing association graph; for indirect relationships between devices involved in chain reactions, corresponding edges are added to the association graph to connect them and annotate relationship attributes. The relationship attributes are information used to describe the characteristics and properties of the relationships between nodes in the association graph.

7. The method for intelligent operation and maintenance of a hydropower plant based on machine learning according to claim 1, characterized in that: The method uses isolation forest to detect anomalies based on equipment operation data and sensor monitoring data, obtains explicit anomalies, reconstructs data distribution through autoencoders, and identifies implicit anomalies, including: The scikit-learn library in Python was selected as the isolation forest algorithm library to initialize the isolation forest model. The multi-dimensional data of the hydropower plant was input into the isolation forest model for training. Multiple isolation trees were constructed by randomly selecting data subsets. Each tree partitioned the data points until every data point was isolated. After training, the equipment operation data and sensor monitoring data to be tested were input into the isolation forest model. The isolation forest model calculated an anomaly score for each data point. Based on the pre-set anomaly score threshold, the data point was determined to be anomaly. When the anomaly score was higher than the threshold, the data point was marked as an explicit anomaly and the relevant information was recorded. Design the model architecture of the autoencoder, including the encoder and decoder; use a multi-layer perceptron as the basic structure of the encoder and decoder, and optimize performance by adjusting the number of hidden layers and neurons; use multi-dimensional data from a hydropower plant to train the autoencoder. During the training process, adjust the parameters by minimizing the reconstruction error so that the data reconstructed by the decoder is close to the original input data; after training, input the equipment operation data and sensor monitoring data to be tested into the autoencoder, reconstruct the data, and calculate the reconstruction error; set a reconstruction error threshold, and when the reconstruction error of the data exceeds the threshold, determine that the data has hidden anomalies.

8. The method for intelligent operation and maintenance of a hydropower plant based on machine learning according to claim 1, characterized in that: The above method fuses explicit anomalies and implicit anomalies based on the correlation graph to obtain abnormal data of the hydropower plant, including: Map abnormally associated device entities to device nodes in the association graph, map abnormally associated sensor entities to sensor nodes in the association graph, and add historical failure modes as predefined nodes to the graph; create explicit abnormality edges between device nodes and sensor nodes and explicit abnormalities, annotate the abnormality score and type; create implicit abnormality edges between device nodes and sensor nodes and implicit abnormalities, annotate the reconstruction error and trend characteristics; Analyze explicit and implicit anomalies of the same device or sensor within the same time period. When the time windows overlap and the parameters are related, mark them as composite anomalies. Calculate the similarity of the explicit and implicit anomalies in the vibration spectrum and temperature gradient, specifically by calculating the cosine similarity. When the similarity exceeds the set threshold, they are determined to be different manifestations of the same fault event. Neo4j's Cypher language is used to query abnormal propagation paths, and path weights are assigned based on the strength of the physical connection between devices to obtain high-probability propagation chains. Based on the high-probability propagation chains, when the same device has multiple explicit abnormalities associated, it is determined to be a high-priority fault. When the same device has explicit abnormalities accompanied by implicit abnormalities, it is determined to be a compound fault. When the same device has multiple implicit abnormalities forming a closed loop in the association map, it is determined to be a systemic degradation. Abnormal data integration and dynamic priority sorting are performed to obtain abnormal data of the hydropower plant.

9. The method for intelligent operation and maintenance of a hydropower plant based on machine learning according to claim 1, characterized in that: The Prophet model is used to predict the trends of equipment operation data and sensor monitoring data, and the deviation values ​​are located by combining residual analysis. The data fluctuation pattern is obtained based on the deviation values. The correlation between the abnormal data of the hydropower plant and the data fluctuation pattern is analyzed to obtain the source of the abnormal data of the hydropower plant, including: Configure a Prophet model and train it using the similarities between explicit and implicit anomalies in vibration spectra and temperature gradients to generate trend, seasonal, and cyclical characteristics of the anomaly data. After training, use the trained Prophet model to predict equipment operation data and sensor monitoring data for a period of time. The prediction results include trend values ​​for the equipment operation data and sensor monitoring data. The actual equipment operation data and sensor monitoring data are compared with the trend value predicted by the Prophet model to calculate the residual. The residual is equal to the actual value minus the predicted value, reflecting the difference between the actual data and the model prediction result. Based on the calculated residual, the deviation value is located using the set threshold. When the absolute value of the residual exceeds the threshold, the data point is determined to be a deviation value. The located deviation values ​​are analyzed, and based on the variation characteristics and distribution of the deviation values, the data fluctuation patterns are classified into periodic fluctuations, sudden fluctuations, and persistent deviations. A correlation analysis is performed between the abnormal data of the hydropower plant and the data fluctuation patterns, and the degree of association between the abnormal data and different fluctuation patterns is determined by calculating the correlation coefficient. Based on the degree of association and combined with the historical failure records of the hydropower plant equipment, the source of the abnormal data of the hydropower plant is inferred.

10. A hydropower plant intelligent operation and maintenance management system based on machine learning, using a hydropower plant intelligent operation and maintenance management method based on machine learning according to any one of claims 1 to 9, characterized in that: include: Data collection and data backtracking module: includes: data collection unit, data lineage tracking unit and data backtracking unit; the data collection unit collects multi-dimensional data of the hydropower plant, including equipment operation data, sensor monitoring data and grid dispatch data, performs data preprocessing, and stores it in the data warehouse; the data lineage tracking unit defines unified metadata tags in the data warehouse and establishes a data lineage tracking mechanism; the data backtracking unit determines the traceability starting point, traces back based on the storage location, traces back the data access history, and traces back according to the data flow diagram to integrate the data traceability chain; Association graph generation module: includes an association graph construction unit and an association graph adjustment unit. The association graph construction unit performs entity recognition and relationship extraction based on the data warehouse and data traceability chain, and uses Neo4j to build an association graph. The association graph adjustment unit analyzes the relationship between power grid dispatch data and equipment operation, including instruction impact analysis and chain reaction analysis, and adjusts the association graph. Hydropower plant anomaly detection module: This module includes an explicit anomaly analysis unit, a latent anomaly analysis unit, and a hydropower plant anomaly data generation unit. The explicit anomaly analysis unit uses an isolation forest to detect anomalies based on equipment operation data and sensor monitoring data to obtain explicit anomalies. The latent anomaly analysis unit reconstructs data distribution through an autoencoder to identify latent anomalies. The hydropower plant anomaly data generation unit fuses explicit and latent anomalies based on a correlation graph to obtain hydropower plant anomaly data. The hydropower plant abnormal data source analysis module includes: a deviation value positioning unit and a hydropower plant abnormal data source analysis unit; wherein, the deviation value positioning unit predicts the trend of equipment operation data and sensor monitoring data through the Prophet model, and locates the deviation value in combination with residual analysis. The hydropower plant abnormal data source analysis unit obtains the data fluctuation pattern based on the deviation value, analyzes the correlation between the hydropower plant abnormal data and the data fluctuation pattern, and obtains the source of the hydropower plant abnormal data.

Images