CN-121981706-A - Thermal power plant operation and maintenance decision method and system based on data large model

CN121981706ACN 121981706 ACN121981706 ACN 121981706ACN-121981706-A

Abstract

The invention discloses a thermal power plant operation and maintenance decision method and system based on a data large model, in particular to the technical field of operation and maintenance decision, comprising S1, multi-mode operation and maintenance data global acquisition, S2, data standardization processing and quality control, S3, constructing a knowledge enhancement type multi-mode operation and maintenance large model, S4, accurately evaluating the health state of equipment and early warning faults, S5, generating a multi-target operation and maintenance decision scheme, S6, executing the decision scheme and collecting data feedback, S7, and upgrading a model iteration optimization and decision system. The invention adopts the knowledge enhancement type multi-mode operation and maintenance large model, thereby greatly improving the comprehensiveness and the accuracy of state evaluation and diagnosis and greatly improving the accuracy of operation and maintenance decisions of the thermal power plant, and further solving the problem that the accuracy of the operation and maintenance decisions of the thermal power plant is greatly reduced due to the fact that the multi-mode data depth fusion is not realized.

Inventors

WANG JIAYIN
LIN SHAOXIONG
LI JUNYI
ZHANG YIQIAN
DING JIE
WANG SHUN
XU XIAOSHAN

Assignees

华能国际电力股份有限公司上海石洞口第一电厂
节点互联(北京)科技有限公司

Dates

Publication Date: 20260505
Application Date: 20260204

Claims (10)

1. The thermal power plant operation and maintenance decision method based on the data large model is characterized by comprising the following specific steps: S1, collecting time sequence sensing data, image data, text data and physical model data of a host computer, an auxiliary machine and a public system of a thermal power plant, and collecting and transmitting the time sequence sensing data, the image data, the text data and the physical model data through an edge computing terminal; s2, data standardization processing and quality control, namely cleaning and standardizing the collected data, and establishing a data quality evaluation index system to form a qualified data set; s3, constructing a knowledge enhancement type multi-mode operation and maintenance large model, namely injecting an operation and maintenance knowledge map of the thermal power plant by adopting a special encoder, a multi-mode fusion layer, a knowledge enhancement layer and a decision output head architecture, and completing model training; S4, accurate evaluation and fault early warning of the health state of the equipment, namely outputting the health score, the fault type, the root cause and the residual useful life of the equipment through a model, and carrying out grading early warning; s5, generating a multi-objective operation and maintenance decision scheme, namely outputting a precise maintenance strategy, a resource allocation scheme and working condition adjustment parameters by combining operation and maintenance constraint conditions; s6, executing the decision scheme, and collecting the execution process data and the effect verification data to form a feedback data set; And S7, model iterative optimization and decision system upgrading, namely training a model based on feedback data increment and optimizing the decision system.
2. The thermal power plant operation and maintenance decision method based on the large data model of claim 1 is characterized in that the time sequence sensing data in S1 is collected through a DCS system, a vibration monitor and oil on-line monitoring equipment, the time sequence sensing data comprises analog quantity and switching quantity, the analog quantity covers temperature hearth outlet smoke temperature and steam turbine cylinder temperature, the precision is 0.1 ℃, the sampling frequency is 10-15 Hz, the pressure comprises main steam pressure and hearth negative pressure, the precision is 0.01-0.05MPa, the sampling frequency is 10-25 Hz, the flow is main steam flow and water supply flow, the precision is 0.5-0.8%, the sampling frequency is 10-25 Hz, the vibration adopts bearing vibration acceleration, the sampling frequency is 2560-2800 Hz, the iron spectrum content in oil parameters is 0.1-0.5ppm, the moisture content is 0.4-0.8%, the sampling frequency is 0.1-0.8 Hz, the switching quantity covers the equipment operation or the state, the interlocking protection action signal, the valve switch state or the switching state, the sampling frequency is 0.1-0.9 Hz, the image is represented by two 0 or 1, the image data is 10-25 Hz, the image data is covered by an infrared image data image sensor, the image sensor is used for covering an infrared image sensor, the image sensor is used for measuring the image sensor, the image sensor is used for covering the image sensor, the image sensor is 5-1024, the image sensor is used for the image sensor, the image sensor is used for displaying the image, and has the image and has the visible display, and has the visible through the visual image, and has the visible image.
3. The thermal power plant operation and maintenance decision method based on the large data model is characterized in that S1 text data are collected through an operation and maintenance management system and an overhaul work order system, the operation and maintenance management system comprises unit load and working condition adjustment records in an operation log, a real-time daily input and overhaul work order fault part, a maintenance mode, replacement spare parts and working hour consumption are carried out, after-fault 22-25 h input and accident report pass through a fault evolution process, root cause analysis and disposal measures, after-accident 72-79 h input and equipment model, parameter threshold and maintenance standard in equipment specifications are recorded, the digital archiving is carried out, and physical model data comprise physical model data, namely, based on equipment design drawing and thermodynamic and dynamics principles, the turbine rotor diameter, the boiler heating surface area, the high-temperature resistant strength in material parameters, the wear-resistant coefficient and the rated load and the rated steam temperature pressure in rated working condition parameters are collected, an equipment physical characteristic database is built, the collection mode adopts an edge computing terminal to collect, the data are transmitted through 5G+ optical fiber double links, the data transmission delay is 50-80ms, and the transmission success rate is 98.5% -99.9%.
4. The thermal power plant operation and maintenance decision method based on a data large model according to claim 1, wherein the method comprises the following steps: the S2 data cleaning comprises the steps of eliminating abrupt change data caused by sensor faults in abnormal values by adopting a 3 sigma rule aiming at time sequence data, supplementing the missing data based on a linear interpolation method, when the missing rate is 5-10%, adopting the same-working-condition historical data migration filling when the missing rate is 10-20%, removing noise by Gaussian filtering aiming at image data, correcting shooting angle deviation by adopting an image registration technology, unifying image size and pixel density, eliminating invalid characters by adopting a natural language processing technology aiming at text data, carrying out word segmentation, part-of-speech marking and semantic normalization, and carrying out bearing abrasion, bearing abrasion and unification as bearing abrasion by adopting a real-time verification technology, wherein the data normalization comprises the steps of converting different-dimension parameters into standardized data with the average value of 0-1 by adopting a Z-score normalization process, mapping the pixel values into the interval of 0-1, establishing a data quality evaluation index system, carrying out real-time verification and returning the acquired data to the end for 5-10 times after unqualified data marking, thereby forming a closed-loop control data acquisition system.
5. The thermal power plant operation and maintenance decision method based on the large data model of claim 1 is characterized in that the S3 model architecture is designed by adopting a four-layer architecture of a special encoder, a multi-mode fusion layer, a knowledge enhancement layer and a decision output head, the special encoder is designed for different data types, time sequence data adopts a Transformer Encoder and TCN time sequence convolution network combined structure, long-short-term dependency relationship is captured, image data adopts a Vision Transformer architecture, characteristics such as infrared hot spots and abnormal appearance are extracted, text data adopts a BERT pre-training model, semantic information is mined, multi-mode fusion is realized, the extracted characteristics of each encoder are aligned and fused in a unified high-dimensional semantic space through a cross-mode attention mechanism, and the fusion dimension is 1024.
6. The thermal power plant operation and maintenance decision method based on the large data model of claim 1 is characterized in that in S3, knowledge enhancement is performed, a thermal power plant operation and maintenance knowledge map is built, rotor dynamics, thermodynamic principles and fault correlation knowledge which cover equipment structure knowledge are adopted, fault type-characteristic parameter-treatment measure mapping relations are adopted, expert experience knowledge is adopted, a typical fault diagnosis rule is adopted, the knowledge map is injected into the model in a constraint condition mode, model prediction results are corrected, output against a physical rule is avoided, model training is conducted, a training data set adopts nearly 5-10 years of operation and maintenance data processed in S2, the sample size is 100 ten thousand-200 ten thousand, normal working conditions, early abnormal and fault state samples are covered, the sample proportion is 3:4:3, an Adam optimizer is adopted, the learning rate is 0.001-0.005, the batch size is 64-85, the training iteration number is 500-800, and the model loss convergence loss value is 0.01-0.05.
7. The thermal power plant operation and maintenance decision method based on the data large model is characterized in that the S4 health state assessment comprises the steps of inputting real-time data processed in the S2 into a trained large model, outputting 0-100 points of equipment health score by the model, dividing 85-100 points into health states, dividing 60-84 points into early abnormal states, dividing 0-60 points into fault states, simultaneously outputting various assessment dimensional operation parameters, equipment abrasion and environmental suitability point scores, defining a state short board, performing fault early warning, accurately identifying fault type identification accuracy rate of 98% -98.9% for the model and fault states, predicting fault evolution paths and residual service lives, outputting fault root cause analysis of 10% -15% by combining a knowledge map, and dividing early warning into four stages, namely I stage slight abnormality, no need of shutdown, 72-78 h internal treatment, II stage general abnormality, monitoring operation, 24-26 h internal treatment accuracy rate of 98% -9%, I stage treatment, 3-IV internal emergency load reduction, and immediate fault treatment of 10-5 h internal emergency shutdown.
8. The thermal power plant operation and maintenance decision method based on the large data model of claim 1, wherein the step S5 is characterized in that a multi-objective optimization decision scheme is generated by combining thermal power plant operation and maintenance constraint conditions, through safety constraint, cost constraint and load demand constraint, a maintenance strategy decision is realized, a precise maintenance mode is output, a level I abnormality adopts an online monitoring and periodic inspection strengthening strategy, a level II abnormality adopts a local maintenance strategy, a level III abnormality adopts a load reduction maintenance strategy, a level IV abnormality adopts an emergency shutdown maintenance strategy, a definite maintenance time window is combined with a unit load valley period, a maintenance flow and technical points, a resource configuration decision is realized, a model, a specification and the number of a spare part list required for maintenance are output, a special tool and detection equipment of a personnel configuration scheme are realized based on historical consumption data and current inventory precision matching, personnel number, a division and tool equipment list, a resource scheduling path is optimized, the allocation cost is reduced, a working condition adjustment decision is realized, a main steam pressure adjustment range is 0.5-0.8MPa, a load adjustment rate is 5% -9% -a rated load/small, a rated load/rated load is supported by a rated load fluctuation curve is realized, and a rapid-change of a schematic diagram is realized, and a rapid operation and a graphic diagram is understood by a graphic form and a graphic diagram is realized.
9. The thermal power plant operation and maintenance decision-making method based on the large data model is characterized in that in S6, operation and maintenance personnel execute operation according to a decision-making scheme, synchronous acquisition of execution process data, including operation steps of maintenance operation records, time consumption, effects, parameter changes before and after adjustment of working condition adjustment data, spare part use amount in resource consumption data, personnel man-hour and energy consumption, effect verification, after execution is completed, health scores and fault characteristic parameters of equipment operation data are acquired, effectiveness of the decision-making scheme is verified, effectiveness evaluation indexes, including that the fault elimination rate is 99.5% -99.8% qualified, the equipment health score lifting amplitude is 20-40 are divided into excellent, feedback data acquisition, namely, scheme optimization proposal of the execution process data, the effect verification data and the operation and maintenance personnel feedback opinion is filed, a feedback data set is formed and is supplemented to a model training database, in S7, the model iteration is implemented in an incremental training mode, the feedback data set in S6 is input into the large model, the adjustment attention weight and the update knowledge of the optimization model parameters are completed once, the iteration decision-making system is completed once per time, the iteration is completed by 50-60 h, the iteration system is updated according to a new iteration constraint condition, and the iteration condition is sequentially updated according to the iteration constraint condition of the iteration system, and the iteration condition is completed once after the iteration condition is changed, and the iteration condition is completed in a new time is changed, and the iteration condition is changed.
10. A thermal power plant operation and maintenance decision system based on a data large model, characterized by comprising a processing host and a memory, wherein the memory stores computer program instructions, which when executed by the processing host, implement the thermal power plant operation and maintenance decision method based on a data large model according to any one of claims 1-9.

Description

Thermal power plant operation and maintenance decision method and system based on data large model Technical Field The invention relates to the technical field of operation and maintenance decision making, in particular to a thermal power plant operation and maintenance decision making method and system based on a data large model. Background The large data model has multiple key roles in operation and maintenance decisions of the thermal power plant, various data of the thermal power plant including equipment states, operation parameters and environmental conditions can be acquired in real time through real-time data acquisition and analysis, and rapid analysis is performed through a large data processing technology, so that operation and maintenance personnel can know the operation conditions of the equipment in time, find potential problems, and make accurate decisions. In the prior art, the patent of the patent publication No. CN119398979A discloses a management method for the safety operation and maintenance of a thermal power plant building structure, and the technology changes passive accident treatment into active prevention and control by establishing a thermal power plant building structure safety operation and maintenance management standard, carries out safety operation and maintenance on the whole life cycle of the thermal power plant building structure, improves daily operation and maintenance and building structure classification management efficiency, and realizes scientific decision of the thermal power plant building structure safety operation and maintenance management. This patent has the following drawbacks. In the operation and maintenance decision-making process of the large model, once the operation and maintenance of the thermal power plant is problematic, decision-making is needed to be carried out in time, but a large amount of data is focused on single equipment fault early warning or local data application, the defects of incomplete data coverage, disjoint of the model and a physical mechanism and lack of operability of decision-making advice exist, and multi-mode data deep fusion cannot be realized, so that the accuracy of the operation and maintenance decision of the thermal power plant is greatly reduced. Disclosure of Invention In order to achieve the purpose, the invention provides the following technical scheme that the thermal power plant operation and maintenance decision method based on the data large model comprises the following specific steps: S1, collecting time sequence sensing data, image data, text data and physical model data of a host computer, an auxiliary machine and a public system of a thermal power plant, and collecting and transmitting the time sequence sensing data, the image data, the text data and the physical model data through an edge computing terminal; s2, data standardization processing and quality control, namely cleaning and standardizing the collected data, and establishing a data quality evaluation index system to form a qualified data set; s3, constructing a knowledge enhancement type multi-mode operation and maintenance large model, namely injecting an operation and maintenance knowledge map of the thermal power plant by adopting a special encoder, a multi-mode fusion layer, a knowledge enhancement layer and a decision output head architecture, and completing model training; S4, accurate evaluation and fault early warning of the health state of the equipment, namely outputting the health score, the fault type, the root cause and the residual useful life of the equipment through a model, and carrying out grading early warning; s5, generating a multi-objective operation and maintenance decision scheme, namely outputting a precise maintenance strategy, a resource allocation scheme and working condition adjustment parameters by combining operation and maintenance constraint conditions; s6, executing the decision scheme, and collecting the execution process data and the effect verification data to form a feedback data set; And S7, model iterative optimization and decision system upgrading, namely training a model based on feedback data increment and optimizing the decision system. In a preferred embodiment, the time series sensing data in S1: the method is characterized in that the method comprises the steps of collecting a DCS system, a vibration monitor and oil on-line monitoring equipment, wherein the DCS system comprises an analog quantity and a switching quantity, the analog quantity comprises a temperature hearth outlet smoke temperature and a steam turbine cylinder temperature, the precision is 0.1 ℃ and the sampling frequency is 10-15 Hz, the pressure comprises main steam pressure and hearth negative pressure, the precision is 0.01-0.05MPa, the sampling frequency is 10-25 Hz, the flow is main steam flow and water supply flow, the precision is 0.5-0.8%, the sampling frequency is 10-25 Hz, the vibration adopts bearing vibration acceleration