Intelligent power plant monitoring early warning system based on neural network algorithm model and implementation method

Through the intelligent monitoring system of the power plant based on neural network algorithm model, the LSTM model is used to preprocess and early warning of generator set data, which solves the problems of high professionalism and comprehensive capabilities of traditional manual monitoring, and realizes the intelligence of equipment failures Early warning and diagnosis improve equipment operation safety and grid stability.

CN119093581APending Publication Date: 2024-12-06SHANGHAI BAOSIGHT SOFTWARE CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411042074.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-31
Publication Date
2024-12-06

AI Technical Summary

Technical Problem

Traditional power plant monitoring relies on manual real-time monitoring, which has high professional and comprehensive capabilities requirements, and the existing digital twin technology cannot fully solve the equipment fault warning and diagnosis needs.

Method used

The intelligent monitoring and early warning system of the power plant based on neural network algorithm model, including data acquisition, data processing and data application layers, preprocessing, offline training and online early warning of the generator set operation data through the long-term and short-term memory neural network (LSTM) model. , realize fault diagnosis and health analysis.

Benefits of technology

It realizes intelligent early warning and optimized diagnosis of generator set equipment, reduces the probability of failure, improves equipment operation safety and grid stability, reduces manual monitoring personnel and operation and maintenance costs, and extends the service life of the equipment.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119093581A_ABST
    Figure CN119093581A_ABST
Patent Text Reader

Abstract

The invention provides a power plant intelligent monitoring early warning system based on a neural network algorithm model and an implementation method, and the system employs the technical means of big data analysis, artificial intelligence, machine learning, expert experience and the like to carry out the data mining of the massive historical operation data of a power plant, and forms the intelligent monitoring system through the construction of the neural network algorithm model. According to the method, intelligent early warning, optimization diagnosis, performance evaluation and the like of the power plant are achieved, meanwhile, powerful informatization and digitization technologies serve as supports, the fine management level of the power plant is improved, traditional personnel supervision is replaced, efficiency is improved, emission is reduced, safe production is effectively guaranteed, production decision is optimized, and a modern intelligent power plant which is safer, more efficient, more economical and more environmentally friendly is built.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of power plant early warning systems, and in particular to a power plant intelligent monitoring panel early warning system based on a neural network algorithm model and an implementation method thereof. Background Art

[0002] Traditional manual monitoring is performed by power plant operation monitoring personnel to monitor the process parameters, analyze and handle faults in real time in the power plant operation system through monitoring screens. The monitoring personnel must be proficient in operating procedures, safety regulations and dispatching regulations, and understand the real-time operating conditions on site, system process flow, and equipment parameters and operating status in operation. When an abnormality occurs in the equipment, the monitoring personnel must have the ability to quickly determine the cause of the fault and handle the fault in the first time, which puts forward stronger professional and more comprehensive comprehensive ability requirements for the monitoring personnel.

[0003] Patent document CN118312746A discloses a device status evaluation method and system based on digital twins, the system comprising: a digital twin modeling module, used to establish a proton accessory equipment operation and maintenance platform model through digital twins; a data acquisition module, used to collect the operating status data of the electrical system equipment in the proton accessory equipment, and calculate the operating risk value of the electrical system equipment through the data processing module. The equipment status analysis module can combine the operating risk value of the electrical system equipment for real-time analysis, and combine past data to evaluate the future operating status of the electrical system equipment, and judge whether the electrical system equipment needs to be overhauled based on the analysis and evaluation results. However, this patent cannot completely solve the existing technical problems, nor can it meet the needs of the present invention.

[0004] With the rapid development of new technologies such as big data analysis, artificial intelligence AI, 5G technology and their widespread application in the industrial field, the traditional thermal power industry is actively promoting the digital transformation of smart power plants. Smart monitoring panels continue to assist operation monitoring personnel in making key decisions, build a new monitoring method based on predictive maintenance, realize intelligent early warning and diagnosis of equipment system failures, explore new modes of operation optimization of generator sets and important auxiliary machines, improve the economy and flexibility of units, improve equipment reliability, reduce equipment operation and maintenance costs, and continuously promote the high-quality development of the power generation industry. Summary of the invention

[0005] In view of the defects in the prior art, the purpose of the present invention is to provide a power plant intelligent monitoring and early warning system based on a neural network algorithm model and an implementation method.

[0006] The power plant intelligent monitoring and early warning system based on the neural network algorithm model provided by the present invention includes: a data acquisition and storage layer, a data processing layer and a data application layer;

[0007] The data acquisition and storage layer is connected to the data processing layer in a one-way communication manner, and the data processing layer is connected to the data application layer in a two-way communication manner;

[0008] The data acquisition and storage layer adopts the standard network communication protocol OPC to collect millisecond-level generator set operation historical data and real-time data from the DCS / PLC system or SIS system and store them in the real-time database;

[0009] The data processing layer is based on the database and completes data preprocessing, offline data training, neural network algorithm modeling, online training and operation of early warning models, parameter early warning, health evaluation, fault diagnosis, control loop performance evaluation, actuator performance evaluation and fault root cause analysis;

[0010] The data application layer performs real-time monitoring and data alarm through the prediction results of the neural network algorithm model of the data processing layer.

[0011] Preferably, the data preprocessing includes: screening, filtering, deduplication and padding the collected raw data to remove noise and errors in the data;

[0012] The data offline training includes: performing offline training based on the characteristics of the entire data sample, and completing offline data verification at the same time, so as to determine whether the data obtained by offline training meets the standards for entering the model application.

[0013] Preferably, a long short-term memory neural network LSTM model is used to perform fault warning analysis on the system health status, important parameter operation status, control loop and actuator performance of the power plant; the LSTM model memorizes information about the current and previous time steps through its long short-term memory units, and the information in the long short-term memory units is linearly transmitted between each LSTM unit, so that the information is transmitted and maintained across multiple time steps;

[0014] The LSTM model adds three gate structures to the hidden layer nodes, namely the input gate, output gate and forget gate, which respectively t , the output value s at the previous moment t-1 and the unit state value c at the previous moment t-1 To exercise control;

[0015] The forget gate is used to control the unit state value c at the previous moment t-1 How much should be forgotten, and the remaining information should be saved to the current cell state c t The calculation formula is as follows:

[0016] f t =σ(W f ·[s t-1 ,x t ]+bf )

[0017] Among them, s t-1 is the output at the previous moment, x t is the input at the current moment, W f is the weight matrix of the forget gate, b f is the bias of the forget gate, σ is the sigmoid activation function of the forget gate, and f t The calculated value is between [0,1], f t = 0 means completely forgetting the unit state at the previous moment, f t =1 means that the unit state at the previous moment is fully retained, f t The value of between (0, 1) means that only the information that needs to be saved in the unit state at the previous moment is remembered;

[0018] The input gate is used to control the output value s at the previous moment t-1 and the current input value x t How much information has entered the current cell state c t In , the output gate is used to control how much information in the current state is output; first, the sigmoid function is used to obtain a value between 0 and 1 to determine how much new state to add. The calculation formula is (1); then a new candidate vector is created Used to control the addition of c t The calculation formula is (2); then multiply the two parts together to determine the total influence on c t The amount of the cell, plus the influence of the previous forget gate, finally gets the current cell state c t The update formula of the LSTM unit state is shown in formula (3):

[0019] i t =σ(W i ·[s t-1 ,x t ]+b i ) (1)

[0020]

[0021]

[0022] In formula (1) to formula (3), s t-1 is the output at the previous moment, x t is the input at the current moment, W i , W c is the weight matrix of the input gate, b i 、b c is the bias of the input gate, σ, tanh is the activation function of the input gate, i tThe calculated value is between [0,1], i t = 0 means completely forgetting the previous output and current input, i t =1 means that the output of the previous moment and the current input are all retained, i t The value of is between (0, 1) which means that only the information that needs to be saved in the previous output and the current input is remembered; c t-1 is the unit state value at the previous moment, and the current state c is calculated by equations (1) to (3) t , that is, integrating long-term memory and current memory to obtain the current unit state c t ;

[0023] The output gate is used to control the current state c t How much information is output in c? First, the sigmoid activation function of the output gate is used to determine c t Which part of needs to be output? The calculation formula is shown in formula (4); then c t Put in the tanh activation function to determine the final output c t The calculation formula is shown in formula (5):

[0024] o t =σ(W o ·[s t-1 ,x t ]+b o ) (4)

[0025] s t =o t ·tanh(c t ) (5)

[0026] Among them, s t-1 is the output at the previous moment, x t is the input at the current moment, W o is the weight matrix of the output gate, b o is the bias of the output gate, tanh is the activation function of the output gate, o t The calculated value is between [0,1];

[0027] The LSTM neural network changes the weight of the self-loop through the structure calculation of multiple gates, so that when the data is in the process of error propagation, the error can be directly entered into the next layer intact, so that the gradient will not explode or disappear no matter how far it propagates, thus ensuring the convergence of the data; the LSTM neural network trains the network by using a large amount of historical data, and then uses the trained network to predict future values.

[0028] Preferably, the online training of the early warning model includes: determining the system input variables of the offline training through the validity and convergence of the offline training data, connecting the real-time values ​​of the system input variables determined by the offline training to the input data of the neural network model, and using the neural network model to make real-time predictions on the system output values, so as to judge whether the equipment system is abnormal, so as to achieve the effect of early warning;

[0029] The specific process of offline data training of the early warning model is as follows:

[0030] The input and output variables of offline training are exported from the data source, and organized into the specified XLS format in an Excel spreadsheet; each key input and output variable is trained separately with the LSTM neural network model, that is, using the offline data training platform; the historical data in the specified XLS format is read, the data is preprocessed, and the input and output columns are specified according to the collection of the data source; the number of hidden layers of the neural network model, the number of nodes in each layer, and the number of iterative training times are set, where the number of neural network layers is 3, the number of nodes in the first layer of the neural network is 32, the number of nodes in the second layer of the neural network is 64, the number of nodes in the third layer of the neural network is 16, and the number of iterative training times is 100; for online data training, the first 75% of the data is used for training, and the last 25% is used for testing; after the training is completed, a neural network parameter matrix is ​​generated, and the training error and test error are returned; the data training fitting effect is judged based on the error data and the image, and it is decided whether to re-export the data for training; the data whose logarithmic training effect meets the preset standards is formed into a weight matrix file for the real-time prediction model operation.

[0031] Preferably, the specific process of online operation of the real-time prediction model is as follows:

[0032] In the real-time neural network prediction model, input the variable data that has been trained offline, and ensure the correctness and completeness of each variable; input the weight matrix folder name that has been trained offline, and this folder stores the offline training result file group; debug the number of hidden layers, the number of nodes in each layer, and the number of iterative training in the real-time neural network prediction model, where the offline data training parameters are consistent with the parameters of the real-time neural network prediction model; track the values ​​predicted by the model online in real time, compare the output estimate of the neural network model with the real-time status value of the parameter, and when the real-time value of the parameter exceeds the upper and lower limits of the neural network estimate, issue a warning to the parameter in time to judge the prediction effect of the model.

[0033] Preferably, the specific prediction model includes: a parameter early warning model, a health evaluation model, a control loop performance evaluation model, an actuator performance evaluation model, a fault diagnosis model, and a fault root cause analysis model;

[0034] The parameter early warning model uses the LSTM neural network algorithm model for parameter early warning. The LSTM neural network module includes three layers of neurons. The first, second and third layers have several neuron submodules respectively. The number of input variables is X, the number of output variables is Y, and the reasonable value of the output variable is predicted by the X input variables. For the monitored target parameter Y, the N main explanatory variable measurement points that directly affect the change of Y are determined to be X1, X2, ... XN. The synchronous historical data of Y and X in the historical time of the LSTM model are used to perform fitting training on the mapping relationship between X and Y. The training results are assigned to the online LTSM calculation model in the form of a weight matrix file. After the online calculation model is run, the predicted Y value calculated in real time by the explanatory variable X collected online in real time is compared with the actual instantaneous Y value. When the deviation between the actual value and the predicted value exceeds a predetermined threshold, the system starts to issue an early warning or prompt. By comparing the output estimated value of the neural network model with the real-time state value of the parameter, when the real-time value of the parameter exceeds the upper and lower limits of the neural network estimated value, the parameter is warned in time to remind the monitoring personnel to pay attention and adjust.

[0035] The health evaluation model uses an LSTM neural network algorithm model to analyze the health of the equipment. The LSTM neural network module includes three layers of neurons. The first, second and third layers have multiple neuron submodules respectively. The number of input variables is X, and the number of output variables is Y. The reasonable value of the output variable is predicted by the X input variables. To determine whether a device is in a healthy state, it is necessary to simultaneously detect several key indicators Y1, Y2, ..., YM of the device. For the monitored target parameter Y, it is determined that the N main explanatory variable measurement points that directly affect the change of Y are X1, X2, ..., XN. The synchronous historical data of Y and X in the historical time of the LSTM model are used to perform fitting training on the mapping relationship between X and Y. The training result is assigned to the online LTSM calculation model in the form of a weight matrix file. After the online calculation model is run, the predicted Y value calculated in real time by the explanatory variable X collected in real time online is compared with the actual instantaneous Y value. The cosine similarity of the vector is used to evaluate the health. The sub-health state of the device is reflected according to the degree of deviation between the actual vector and the estimated vector, and the system gives a specific health percentage for characterization.

[0036] Preferably, the health evaluation model inputs the real-time value of the variable into the neural network model and makes a real-time prediction of the output variable; tracks the predicted value of the output variable of the neural network model according to the actual variable parameters of the system operation status, and uses correlation analysis to evaluate and score the health of the system, so as to realize abnormal warning of the system and equipment health analysis, including health analysis of the coal mill, primary fan, blower, induced draft fan, feed water pump and high and low pressure systems of the generator set.

[0037] Preferably, the control loop performance evaluation model evaluates the process performance of the control loop based on the set value and feedback value variables of the control system, the relative deviation of the set value and feedback value of the controlled object in the control loop, and the stability performance of the controlled object;

[0038] The definition of controlled performance evaluation of a control loop includes transient evaluation and normal evaluation of the control loop; the evaluation of the ability to return to a new equilibrium state after the equilibrium state of the control loop is broken by interference excitation is called transient evaluation of the control loop, and the transient evaluation fault types of the main control loop are divided into: sharp increase in positive direction, sharp increase in negative direction, increase in vibration amplitude, increase in vibration frequency, increase in vibration amplitude frequency, sharp increase in positive direction, increase in vibration amplitude, sharp increase in positive direction, increase in vibration frequency, sharp increase in negative direction, increase in vibration amplitude, sharp increase in negative direction, and increase in vibration amplitude frequency; when the deviation between the set value and the feedback value within a defined period of time of the control loop is evaluated using the scientific statistical quantity root mean square to evaluate the average performance of the control loop, it is called normal evaluation of the control loop.

[0039] Preferably, the actuator performance evaluation model is an actuator performance detection method based on signal trend analysis, and the actuator fault type is judged according to the input variables and feedback variable data of the actuator, and the fault type is divided into: sticking / slipping fault, actuator dead zone fault, actuator gain fault, actuator deviation fault, and actuator stuck fault;

[0040] The fault diagnosis model constructs a coal mill mechanism model according to the coal mill mechanism and historical data; at the same time, according to the coal mill fault generation mechanism, the model parameters or control methods are adjusted to generate a large amount of typical fault sample data; the fault diagnosis function performs fault diagnosis and analysis on the coal mill, and performs logical judgment on the coal feeder flow, coal mill inlet air pressure, coal mill lower air pressure, coal mill outlet temperature, coal mill cold air damper opening, and coal mill hot air damper opening variables during operation, and diagnoses four fault states of coal mill normal, coal mill coal outage, coal mill coal blockage, and coal mill spontaneous combustion. The coal mill fault diagnosis types are divided into: normal, coal outage, coal blockage, and natural;

[0041] The fault root cause analysis model is based on historical healthy operation data and uses an LSTM model to train the fuel flow, air supply, main steam flow, main water supply, and actual power. That is, one of the indicators is used as the target variable Y, and the remaining indicators are used as explanatory variables X to obtain a mapping relationship in the weight parameter matrix format, and then another four groups of mapping relationships are obtained by analogy. During online operation, five groups of mapping relationship models are run simultaneously to obtain five predicted Y values, which are compared with the actual system values ​​in real time. According to the network theory of conservation of energy and materials, the one with the largest deviation is the location of the fault. When a fault occurs, the five groups of Y value deviation curves are queried, and the one that first produces a deviation is the fault source.

[0042] The method for implementing the power plant intelligent monitoring panel early warning system based on the neural network algorithm model provided by the present invention comprises the following steps:

[0043] Step S1: using the massive historical data stored in the real-time operation process of the unit, selecting a time period that can cover the full operating range of the unit to collect the historical data;

[0044] Step S2: Determine the input / output variables of the prediction model according to the actual situation;

[0045] Step S3: data preprocessing, including outlier removal and missing value filling;

[0046] Step S4: Offline training of data to obtain a training curve diagram, determine the correlation of input / output variables, and determine whether the input / output variables are highly fitted, so as to find input / output variable data that meets the preset conditions;

[0047] Step S5: Neural network model configuration, using the neural network algorithm platform to perform program logic configuration on the system equipment of parameter warning, health evaluation, control loop performance evaluation, actuator performance evaluation, fault diagnosis, and fault root cause analysis of each system that needs warning of the generator set according to the determined input / output variables;

[0048] Step S6: The prediction model is trained online, and the neural network model is trained with the preprocessed data. According to the number of system input / output variables and the system complexity, the number of hidden layers, the number of nodes, the weights, the learning rate, and the activation function of the deep neural network model are adjusted, and the training is repeated until the accuracy of the neural network prediction model meets the preset requirements;

[0049] The normality test analysis method is used to detect outliers, and then the outliers are processed by linear regression. The specific process is: data collection and organization, including cleaning and formatting, for subsequent analysis; using statistical software to test the normality of the data, including drawing histograms, PP diagrams or QQ diagrams to intuitively determine whether the data distribution is close to the normal distribution;

[0050] Outlier detection based on normal distribution, first, calculate the mean and standard deviation, calculate the mean μ and standard deviation σ of the entire data set; second, determine the outlier threshold, according to the characteristics of normal distribution, obtain the probability of a data point falling within the range of the mean plus or minus 3 times the standard deviation μ±3σ, and the data points outside this range are regarded as outliers; third, identify outliers, compare each data point with the threshold, if the value of a data point is less than μ-3σ or greater than μ+3σ, it is marked as an outlier; finally, record all identified outliers, and choose to delete or replace according to the actual situation, specifically: through linear regression processing, replace the samples with more serious data deviations from the overall data through the approximate function fitted by normal data, and use cubic spline interpolation to fill in the missing data.

[0051] Compared with the prior art, the present invention has the following beneficial effects:

[0052] (1) The present invention mines unit operation data, trains data offline, builds a neural network model, and trains an early warning model online. The neural network model predicts abnormal information of the generator set equipment system and makes early warning judgments in advance, thereby realizing intelligent early warning of the generator set system under various working conditions and optimized diagnosis, health analysis, and performance evaluation of the unit equipment, and assisting the monitoring and operation personnel to handle abnormalities, thereby reducing the probability and risk of unit equipment failure, improving the safety of equipment operation, and ensuring the stability of the power grid system. At the same time, it lays the foundation for the construction of digital and intelligent transformation in the power field, prolongs the life cycle of "production-type" data, and has broad application prospects;

[0053] (2) Before the smart monitoring system was put into use, a single unit required 4-5 people to be on duty 24 hours a day. After the smart monitoring system was put into use, only 2-3 people were needed to be on duty, reducing the time and intensity of manual monitoring; the number of abnormal shutdowns of the unit can be reduced by 1-2 times throughout the year, extending the utilization rate of key equipment and reducing the economic losses caused by abnormal shutdowns of the unit.

[0054] (3) Smart monitoring panels have the ability to capture fault warnings in advance, extend the service life and maintenance cycle of equipment, save inspection and repair costs; reduce unplanned shutdowns of units and improve the stability of the power grid system; assist monitoring personnel in making scientific and correct fault warnings and reduce the time spent manually tracing the cause of the fault through past experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings:

[0056] Figure 1A method for implementing a power plant intelligent monitoring and early warning system based on a neural network algorithm model provided by the present invention is a basic principle structure diagram of an LSTM network;

[0057] Figure 2 This is an offline training curve diagram of a method for implementing a power plant smart monitoring and early warning system based on a neural network algorithm model in this implementation case;

[0058] Figure 3 It is a parameter warning model control program diagram of a method for implementing a power plant intelligent monitoring and early warning system based on a neural network algorithm model in this implementation case;

[0059] Figure 4 This is a control program diagram of a health analysis model for an implementation method of a power plant smart monitoring and early warning system based on a neural network algorithm model in this implementation case;

[0060] Figure 5 This is a fault diagnosis model control program diagram of a method for implementing a power plant intelligent monitoring and early warning system based on a neural network algorithm model in this implementation case;

[0061] Figure 6 This is the warning screen of a method for implementing a smart monitoring and early warning system for a power plant based on a neural network algorithm model in this implementation case. DETAILED DESCRIPTION

[0062] The present invention is described in detail below in conjunction with specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those of ordinary skill in the art, several changes and improvements can also be made without departing from the concept of the present invention. These all belong to the protection scope of the present invention.

[0063] Example

[0064] The present invention provides a power plant intelligent monitoring and early warning system based on a neural network algorithm model, including: a data acquisition and storage layer, a data processing layer, and a data application layer. The data acquisition and storage layer is connected to the data processing layer in a one-way communication manner, and the data processing layer is connected to the data application layer in a two-way communication manner. The data acquisition and storage layer uses the standard network communication protocol OPC to collect millisecond-level generator set operation historical data and real-time data from the DCS / PLC system or SIS system, and store them in a real-time database. The data processing layer is based on the database and completes system early warning processing tasks such as data preprocessing, offline data training, neural network algorithm modeling, online training and operation of early warning models, as well as parameter early warning, health evaluation, fault diagnosis, control loop performance evaluation, actuator performance evaluation, and fault root cause analysis. The data application layer is the system configuration management, neural network model configuration and functional application link of the early warning system. The values ​​predicted by the neural network algorithm model of the data processing layer are transmitted to the data application layer to monitor the key parameters of the power plant process system in real time, predict the change trend, equipment health, system fault diagnosis, control loop performance evaluation, actuator performance evaluation, fault root cause analysis, etc. for real-time monitoring and data alarm, so that the monitoring personnel can find the early warning information in time and solve the faults in production operation. The data acquisition and storage layer can only upload to the data processing layer in one direction, and the data processing layer cannot send the results calculated by the neural network model to the data acquisition and storage layer for direct basic control system control.

[0065] like Figure 1 The figure shows the basic principle structure diagram of the LSTM network, which is an implementation method of the intelligent monitoring and early warning system of a power plant based on a neural network algorithm model. The long short-term memory neural network (LSTM) is a special recurrent neural network. LSTM changes the weight of the self-loop by adding input gates, forget gates, and output gates. The self-loop storage unit can block any external interference. The forget gate of the LSTM network can choose whether to discard the information content obtained from the previous step. The output is 1 to retain the information, and the output is 0 to discard the information. The input gate contains the output of the previous step and the input of this step, which are input into the network unit together. Finally, the output gate will output the final result of the network at this layer. Among them, f t ,i t ,o t Represent the local outputs of the forget gate, input gate, and output gate submodules, respectively, and w f ,w i ,w o Represents the weight matrices of the three, b f ,b i ,b oRepresent the bias matrices of the three respectively. σ is the sigmoid function and tanh is the hyperbolic tangent function, both of which are commonly used activation functions in neural networks. In this example, the long short-term memory neural network (LSTM) is used to train, verify, and test parameter warning and health evaluation.

[0066] The forget gate is used to control the unit state value c at the previous moment t-1 How much should be forgotten, and the remaining information should be saved to the current cell state c t The calculation formula is as follows:

[0067] f t =σ(W f ·[s t-1 ,x t ]+b f )

[0068] Among them, s t-1 is the output at the previous moment, x t is the input at the current moment, W f is the weight matrix of the forget gate, b f is the bias of the forget gate, σ is the sigmoid activation function of the forget gate, and f t The calculated value is between [0,1], f t = 0 means completely forgetting the unit state at the previous moment, f t =1 means that the unit state at the previous moment is fully retained, f t The value of between (0, 1) means that only the information that needs to be saved in the unit state at the previous moment is remembered;

[0069] The input gate is used to control the output value s at the previous moment t-1 and the current input value x t How much information has entered the current cell state c t In , the output gate is used to control how much information in the current state is output; first, the sigmoid function is used to obtain a value between 0 and 1 to determine how much new state to add. The calculation formula is (1); then a new candidate vector is created Used to control the addition of c t The calculation formula is (2); then multiply the two parts together to determine the total influence on c t The amount of the cell, plus the influence of the previous forget gate, finally gets the current cell state c t The update formula of the LSTM unit state is shown in formula (3):

[0070] i t =σ(W i ·[s t-1 ,xt ]+b i ) (1)

[0071]

[0072]

[0073] In formula (1) to formula (3), s t-1 is the output at the previous moment, x t is the input at the current moment, W i , W c is the weight matrix of the input gate, b i , b c is the bias of the input gate, σ, tanh is the activation function of the input gate, i t The calculated value is between [0,1], i t = 0 means completely forgetting the previous output and current input, i t =1 means that the output of the previous moment and the current input are all retained, i t The value of is between (0, 1) which means that only the information that needs to be saved in the previous output and the current input is remembered; c t-1 is the unit state value at the previous moment, and the current state c is calculated by equations (1) to (3) t , that is, integrating long-term memory and current memory to obtain the current unit state c t ;

[0074] The output gate is used to control the current state c t How much information is output in c? First, the sigmoid activation function of the output gate is used to determine c t Which part of needs to be output? The calculation formula is shown in formula (4); then c t Put in the tanh activation function to determine the final output c t The calculation formula is shown in formula (5):

[0075] o t =σ(W o ·[s t-1 ,x t ]+b o ) (4)

[0076] s t =o t ·tanh(c t ) (5)

[0077] Among them, s t-1 is the output at the previous moment, x t is the input at the current moment, W ois the weight matrix of the output gate, b o is the bias of the output gate, tanh is the activation function of the output gate, o t The calculated value is between [0,1];

[0078] The LSTM neural network changes the weight of the self-loop through the structure calculation of multiple gates, so that when the data is in the process of error propagation, the error can be directly entered into the next layer intact, so that the gradient will not explode or disappear no matter how far it propagates, thus ensuring the convergence of the data; the LSTM neural network trains the network by using a large amount of historical data, and then uses the trained network to predict future values.

[0079] The online training of the early warning model includes: determining the system input variables of the offline training through the validity and convergence of the offline training data, connecting the real-time values ​​of the system input variables determined by the offline training to the input data of the neural network model, and using the neural network model to make real-time predictions on the system output values, so as to judge whether the equipment system is abnormal, so as to achieve the effect of early warning;

[0080] The specific process of offline data training of the early warning model is as follows:

[0081] The input and output variables of offline training are exported from the data source, and organized into the specified XLS format in an Excel spreadsheet; each key input and output variable is trained separately with the LSTM neural network model, that is, using the offline data training platform; the historical data in the specified XLS format is read, the data is preprocessed, and the input and output columns are specified according to the collection of the data source; the number of hidden layers of the neural network model, the number of nodes in each layer, and the number of iterative training times are set, where the number of neural network layers is 3, the number of nodes in the first layer of the neural network is 32, the number of nodes in the second layer of the neural network is 64, the number of nodes in the third layer of the neural network is 16, and the number of iterative training times is 100; for online data training, the first 75% of the data is used for training, and the last 25% is used for testing; after the training is completed, a neural network parameter matrix is ​​generated, and the training error and test error are returned; the data training fitting effect is judged based on the error data and the image, and it is decided whether to re-export the data for training; the data whose logarithmic training effect meets the preset standards is formed into a weight matrix file for the real-time prediction model operation.

[0082] The specific process of online operation of the real-time prediction model is as follows:

[0083] In the real-time neural network prediction model, input the variable data that has been trained offline, and ensure the correctness and completeness of each variable; input the weight matrix folder name that has been trained offline, and this folder stores the offline training result file group; debug the number of hidden layers, the number of nodes in each layer, and the number of iterative training in the real-time neural network prediction model, where the offline data training parameters are consistent with the parameters of the real-time neural network prediction model; track the values ​​predicted by the model online in real time, compare the output estimate of the neural network model with the real-time status value of the parameter, and when the real-time value of the parameter exceeds the upper and lower limits of the neural network estimate, issue a warning to the parameter in time to judge the prediction effect of the model.

[0084] Specific prediction models include: parameter warning model, health evaluation model, control loop performance evaluation model, actuator performance evaluation model, fault diagnosis model, and fault root cause analysis model;

[0085] The parameter early warning model uses the LSTM neural network algorithm model for parameter early warning. The LSTM neural network module includes three layers of neurons. The first, second and third layers have several neuron submodules respectively. The number of input variables is X, the number of output variables is Y, and the reasonable value of the output variable is predicted by the X input variables. For the monitored target parameter Y, the N main explanatory variable measurement points that directly affect the change of Y are determined to be X1, X2, ... XN. The synchronous historical data of Y and X in the historical time of the LSTM model are used to perform fitting training on the mapping relationship between X and Y. The training results are assigned to the online LTSM calculation model in the form of a weight matrix file. After the online calculation model is run, the predicted Y value calculated in real time by the explanatory variable X collected online in real time is compared with the actual instantaneous Y value. When the deviation between the actual value and the predicted value exceeds a predetermined threshold, the system starts to issue an early warning or prompt. By comparing the output estimated value of the neural network model with the real-time state value of the parameter, when the real-time value of the parameter exceeds the upper and lower limits of the neural network estimated value, the parameter is warned in time to remind the monitoring personnel to pay attention and adjust.

[0086] The health evaluation model uses an LSTM neural network algorithm model to analyze the health of the equipment. The LSTM neural network module includes three layers of neurons. The first, second and third layers have multiple neuron submodules respectively. The number of input variables is X, and the number of output variables is Y. The reasonable value of the output variable is predicted by the X input variables. To determine whether a device is in a healthy state, it is necessary to simultaneously detect several key indicators Y1, Y2, ..., YM of the device. For the monitored target parameter Y, it is determined that the N main explanatory variable measurement points that directly affect the change of Y are X1, X2, ..., XN. The synchronous historical data of Y and X in the historical time of the LSTM model are used to perform fitting training on the mapping relationship between X and Y. The training result is assigned to the online LTSM calculation model in the form of a weight matrix file. After the online calculation model is run, the predicted Y value calculated in real time by the explanatory variable X collected in real time online is compared with the actual instantaneous Y value. The cosine similarity of the vector is used to evaluate the health. The sub-health state of the device is reflected according to the degree of deviation between the actual vector and the estimated vector, and the system gives a specific health percentage for characterization.

[0087] The health evaluation model inputs the real-time value of the variable into the neural network model and makes a real-time prediction of the output variable; the predicted value of the output variable of the neural network model is tracked according to the actual variable parameters of the system operation status, and the health of the system is evaluated and scored using correlation analysis to achieve abnormal warning of the system and equipment health analysis, including health analysis of the coal mill, primary fan, blower, induced draft fan, feed water pump and high and low pressure systems of the generator set.

[0088] The control loop performance evaluation model evaluates the process performance of the control loop based on the set value and feedback value variables of the control system, the relative deviation of the set value and feedback value of the controlled object in the control loop, and the stability performance of the controlled object;

[0089] The definition of controlled performance evaluation of a control loop includes transient evaluation and normal evaluation of the control loop; the evaluation of the ability to return to a new equilibrium state after the equilibrium state of the control loop is broken by interference excitation is called transient evaluation of the control loop, and the transient evaluation fault types of the main control loop are divided into: sharp increase in positive direction, sharp increase in negative direction, increase in vibration amplitude, increase in vibration frequency, increase in vibration amplitude frequency, sharp increase in positive direction, increase in vibration amplitude, sharp increase in positive direction, increase in vibration frequency, sharp increase in negative direction, increase in vibration amplitude, sharp increase in negative direction, and increase in vibration amplitude frequency; when the deviation between the set value and the feedback value within a defined period of time of the control loop is evaluated using the scientific statistical quantity root mean square to evaluate the average performance of the control loop, it is called normal evaluation of the control loop.

[0090] The actuator performance evaluation model, the actuator performance detection method based on signal trend analysis, determines the actuator fault type according to the actuator input variable and feedback variable data, and the fault type is divided into: sticking / slipping fault, actuator dead zone fault, actuator gain fault, actuator deviation fault, and actuator stuck fault;

[0091] The fault diagnosis model constructs a coal mill mechanism model according to the coal mill mechanism and historical data; at the same time, according to the coal mill fault generation mechanism, the model parameters or control methods are adjusted to generate a large amount of typical fault sample data; the fault diagnosis function performs fault diagnosis and analysis on the coal mill, and performs logical judgment on the coal feeder flow, coal mill inlet air pressure, coal mill lower air pressure, coal mill outlet temperature, coal mill cold air damper opening, and coal mill hot air damper opening variables during operation, and diagnoses four fault states of coal mill normal, coal mill coal outage, coal mill coal blockage, and coal mill spontaneous combustion. The coal mill fault diagnosis types are divided into: normal, coal outage, coal blockage, and natural;

[0092] The fault root cause analysis model is based on historical healthy operation data and uses an LSTM model to train the fuel flow, air supply, main steam flow, main water supply, and actual power. That is, one of the indicators is used as the target variable Y, and the remaining indicators are used as explanatory variables X to obtain a mapping relationship in the weight parameter matrix format, and then another four groups of mapping relationships are obtained by analogy. During online operation, five groups of mapping relationship models are run simultaneously to obtain five predicted Y values, which are compared with the actual system values ​​in real time. According to the network theory of conservation of energy and materials, the one with the largest deviation is the location of the fault. When a fault occurs, the five groups of Y value deviation curves are queried, and the one that first produces a deviation is the fault source.

[0093] In order to further illustrate the specific method of this example, taking the coal mill outlet temperature parameter warning model, coal mill health analysis model, and coal mill fault warning model in the power plant pulverizing system as examples, the specific implementation method of this example is described in detail. The specific steps of the system implementation method in this embodiment are as follows:

[0094] Step S1: Data collection: This step mainly uses the OPC communication protocol to collect real data from the SIS system into the database; and selects a frequency of 2ms for the switch quantity and analog quantity to collect about 600,000 to 1 million historical data of each switch quantity and analog quantity.

[0095] Step S2: Determine the input / output variables of the prediction model; for different systems, the input / output variables of the prediction model are determined. Before establishing the neural network model, the input / output variables of the model need to be determined according to the actual situation, and representative typical type data are selected as the input / output variables of the neural network prediction model. The following Table 1 shows the input / output variable table determined by the coal mill outlet temperature parameter warning model, the coal mill health analysis model, and the coal mill fault warning model.

[0096] Table 1 Input / output variables determined for a coal mill in a power plant Step S3: Data preprocessing; During the operation of the system, there are many influencing factors, some of which inevitably include redundancy. A large amount of historical data will not only increase the burden of neural network model training, but also often affect the accuracy of the model. Therefore, the principal component analysis (PCA) dimensionality reduction method is used to reduce the dimensionality of the data, and the data after PCA dimensionality reduction is preprocessed to remove outliers and fill in missing values. The normality test analysis method is used to detect outliers, and then the outliers are processed by linear regression, that is, the samples with serious deviations from the overall data are replaced by the approximate function fitted by the normal data; the missing data are filled by cubic spline interpolation.

[0097] Step S4: Data offline training; Figure 2 As shown, in this example, the preprocessed data is imported into the offline training platform, and the curve graph of data verification is observed to judge the correlation of input / output variables to determine the fitting of input / output variables. The peak and trough trends of the imported data and the verification data are consistent, indicating that the input / output variables of this phase data have a high fitting degree and can be used as training data for the neural network to find ideal input / output variable data.

[0098] Step S5: Neural network model configuration, such as Figure 3 , Figure 4 , Figure 5 As shown, in this example, a neural network algorithm platform is used to perform program logic configuration on the relevant system equipment for parameter warning, health evaluation, control loop performance evaluation, actuator performance evaluation, fault diagnosis, and fault root cause analysis of each system that requires warning of the generator set according to the determined input / output variables.

[0099] Step S6: The prediction model is trained online, and the neural network model is trained with the preprocessed data. According to the number of system input / output variables and system complexity, the number of hidden layers, number of nodes, weights, learning rate, activation function and other parameters of the deep neural network model are adjusted and repeatedly trained until the accuracy of the neural network prediction model meets the requirements. In this example, the number of hidden layers is 3, the number of nodes in the first layer is 32, the number of nodes in the second layer is 64, and the number of nodes in the third layer is 16. The weight parameters in this example are configured according to the weight parameters trained by the offline training platform and the weight matrix of the neural network.

[0100] In this example, Figure 6 As shown, important parameter indicators related to the unit's parameter warning, health evaluation, control loop performance evaluation, actuator performance evaluation, fault diagnosis, and fault root cause analysis are displayed on the monitoring screen.

[0101] Among them, the parameter warning screen displays upper and lower limits, which are calculated based on historical data through a neural network. They will change with changes in working conditions, and an alarm will be quickly triggered if the actual value exceeds the limit.

[0102] The health analysis display shows the status of the system, which is divided into 4 states: healthy, good, qualified, and unqualified. When the health is in the range of 90%-100%, it means healthy; when the health is in the range of 80%-90%, it means good; when the health is in the range of 80%-60%, it means qualified; when the health is in the range of 0%-60%, it means unqualified.

[0103] The control loop performance evaluation monitors the controlled parameters and divides them into 12 states: normal, sharp increase in positive direction, sharp increase in negative direction, increase in vibration amplitude, increase in vibration frequency, increase in vibration amplitude and frequency, sharp increase in positive direction and increase in vibration amplitude, sharp increase in positive direction and increase in vibration frequency, sharp increase in positive direction and increase in vibration frequency, sharp increase in negative direction and increase in vibration amplitude, sharp increase in negative direction and increase in vibration frequency, and sharp increase in negative direction and increase in vibration amplitude and frequency, a total of 12 states.

[0104] The actuator performance evaluation monitoring controlled parameters are divided into 6 states: normal, sticking / slipping, dead zone, gain fault, deviation fault, and stuck.

[0105] The controlled parameters of fault diagnosis are divided into four states: normal coal mill, coal outage, coal blockage, and natural.

[0106] Those skilled in the art know that, in addition to implementing the system, device and its various modules provided by the present invention in a purely computer-readable program code, it is entirely possible to implement the same program in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers and embedded microcontrollers by logically programming the method steps. Therefore, the system, device and its various modules provided by the present invention can be considered as a hardware component, and the modules included therein for implementing various programs can also be considered as structures within the hardware component; the modules for implementing various functions can also be considered as both software programs for implementing the method and structures within the hardware component.

[0107] The above describes the specific embodiments of the present invention. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which does not affect the essence of the present invention. In the absence of conflict, the embodiments of the present application and the features in the embodiments can be combined with each other arbitrarily.

Claims

1. A power plant intelligent monitoring and early warning system based on a neural network algorithm model, characterized in that: include: Data collection and storage layer, data processing layer and data application layer; The data acquisition and storage layer is connected to the data processing layer in a one-way communication manner, and the data processing layer is connected to the data application layer in a two-way communication manner; The data acquisition and storage layer adopts the standard network communication protocol OPC to collect millisecond-level generator set operation historical data and real-time data from the DCS / PLC system or SIS system and store them in the real-time database; The data processing layer is based on the database and completes data preprocessing, offline data training, neural network algorithm modeling, online training and operation of early warning models, parameter early warning, health evaluation, fault diagnosis, control loop performance evaluation, actuator performance evaluation and fault root cause analysis; The data application layer performs real-time monitoring and data alarm through the prediction results of the neural network algorithm model of the data processing layer.

2. The power plant intelligent monitoring and early warning system based on the neural network algorithm model according to claim 1 is characterized in that: The data preprocessing includes: screening, filtering, deduplication and filling the collected raw data to remove noise and errors in the data; The data offline training includes: performing offline training based on the characteristics of the entire data sample, and completing offline data verification at the same time, so as to determine whether the data obtained by offline training meets the standards for entering the model application.

3. The power plant intelligent monitoring and early warning system based on the neural network algorithm model according to claim 1 is characterized in that: The LSTM model of the long short-term memory neural network is used to conduct fault warning analysis on the system health status, operation status of important parameters, control loop and actuator performance of the power plant. The LSTM model memorizes information about the current and previous time steps through its long short-term memory units. The information in the long short-term memory units is transmitted linearly between each LSTM unit, so that the information is transmitted and maintained across multiple time steps. The LSTM model adds three gate structures to the hidden layer nodes, namely the input gate, output gate and forget gate, which respectively t , the output value s at the previous moment t-1 and the unit state value c at the previous moment t-1 To exercise control; The forget gate is used to control the unit state value c at the previous moment t-1 How much should be forgotten, and the remaining information should be saved to the current cell state c t The calculation formula is as follows: f t =σ(W f ·[s t-1 ,x t ]+b f ) Among them, s t-1 is the output at the previous moment, x t is the input at the current moment, W f is the weight matrix of the forget gate, b f is the bias of the forget gate, σ is the sigmoid activation function of the forget gate, and f t The calculated value is between [0,1], f t = 0 means completely forgetting the unit state at the previous moment, f t =1 means that the unit state at the previous moment is fully retained, f t The value of between (0, 1) means that only the information that needs to be saved in the unit state at the previous moment is remembered; The input gate is used to control the output value s at the previous moment t-1 and the current input value x t How much information has entered the current cell state c t In , the output gate is used to control how much information in the current state is output; first, the sigmoid function is used to obtain a value between 0 and 1 to determine how much new state to add. The calculation formula is (1); then a new candidate vector is created Used to control the addition of c t The calculation formula is (2); then multiply the two parts together to determine the total influence on c t The amount of the cell, plus the influence of the previous forget gate, finally gets the current cell state c t The update formula of the LSTM unit state is shown in formula (3): I t =σ(W i ·[s t-1 ,x t ]+b i ) (1) In formula (1) to formula (3), s t-1 is the output at the previous moment, x t is the input at the current moment, W i , W c is the weight matrix of the input gate, b i 、b c is the bias of the input gate, σ, tanh is the activation function of the input gate, i t The calculated value is between [0,1], i t = 0 means completely forgetting the previous output and current input, i t =1 means that the output of the previous moment and the current input are all retained, i t The value of is between (0, 1) which means that only the information that needs to be saved in the previous output and the current input is remembered; c t-1 is the unit state value at the previous moment, and the current state c is calculated by equations (1) to (3) t , that is, integrating long-term memory and current memory to obtain the current unit state c t ; The output gate is used to control the current state c t How much information is output in c? First, the sigmoid activation function of the output gate is used to determine c t Which part of needs to be output? The calculation formula is shown in formula (4); then c t Put in the tanh activation function to determine the final output c t The calculation formula is shown in formula (5): the t =σ(W o ·[s t-1 ,x t ]+b o ) (4) S t =o t ·tanh(c t ) (5) Among them, s t-1 is the output at the previous moment, x t is the input at the current moment, W o is the weight matrix of the output gate, b o is the bias of the output gate, tanh is the activation function of the output gate, o t The calculated value is between [0,1]; The LSTM neural network changes the weight of the self-loop through the structure calculation of multiple gates, so that when the data is in the process of error propagation, the error can be directly entered into the next layer intact, so that the gradient will not explode or disappear no matter how far it propagates, thus ensuring the convergence of the data; the LSTM neural network trains the network by using a large amount of historical data, and then uses the trained network to predict future values.

4. The power plant intelligent monitoring and early warning system based on the neural network algorithm model according to claim 1 is characterized in that: The online training of the early warning model includes: determining the system input variables of the offline training through the validity and convergence of the offline training data, connecting the real-time values ​​of the system input variables determined by the offline training to the input data of the neural network model, and using the neural network model to make real-time predictions on the system output values, so as to judge whether the equipment system is abnormal, so as to achieve the effect of early warning; The specific process of offline data training of the early warning model is as follows: The input and output variables of offline training are exported from the data source, and organized into the specified XLS format in an Excel spreadsheet; each key input and output variable is trained separately with the LSTM neural network model, that is, using the offline data training platform; the historical data in the specified XLS format is read, the data is preprocessed, and the input and output columns are specified according to the collection of the data source; the number of hidden layers of the neural network model, the number of nodes in each layer, and the number of iterative training times are set, where the number of neural network layers is 3, the number of nodes in the first layer of the neural network is 32, the number of nodes in the second layer of the neural network is 64, the number of nodes in the third layer of the neural network is 16, and the number of iterative training times is 100; for online data training, the first 75% of the data is used for training, and the last 25% is used for testing; after the training is completed, a neural network parameter matrix is ​​generated, and the training error and test error are returned; the data training fitting effect is judged based on the error data and the image, and it is decided whether to re-export the data for training; the data whose logarithmic training effect meets the preset standards is formed into a weight matrix file for the real-time prediction model operation.

5. The power plant intelligent monitoring and early warning system based on the neural network algorithm model according to claim 1 is characterized in that: The specific process of online operation of the real-time prediction model is as follows: In the real-time neural network prediction model, input the determined variable data that has been trained offline, and ensure the correctness and completeness of each variable; input the weight matrix folder name that has been trained offline, which contains the offline training result file group; In the real-time neural network prediction model, the number of hidden layers, the number of nodes in each layer, and the number of iterative training are debugged. The offline data training parameters here are consistent with the parameters of the real-time neural network prediction model. The values ​​predicted by the model are tracked online in real time, and the output estimate of the neural network model is compared with the real-time status value of the parameter. When the real-time value of the parameter exceeds the upper and lower limits of the neural network estimate, the parameter is warned in time to judge the prediction effect of the model.

6. The power plant intelligent monitoring and early warning system based on the neural network algorithm model according to claim 1 is characterized in that: Specific prediction models include: parameter warning model, health evaluation model, control loop performance evaluation model, actuator performance evaluation model, fault diagnosis model, and fault root cause analysis model; The parameter early warning model uses the LSTM neural network algorithm model for parameter early warning. The LSTM neural network module includes three layers of neurons. The first, second and third layers have several neuron submodules respectively. The number of input variables is X, the number of output variables is Y, and the reasonable value of the output variable is predicted by the X input variables. For the monitored target parameter Y, the N main explanatory variable measurement points that directly affect the change of Y are determined to be X1, X2, ... XN. The synchronous historical data of Y and X in the historical time of the LSTM model are used to perform fitting training on the mapping relationship between X and Y. The training results are assigned to the online LTSM calculation model in the form of a weight matrix file. After the online calculation model is run, the predicted Y value calculated in real time by the explanatory variable X collected online in real time is compared with the actual instantaneous Y value. When the deviation between the actual value and the predicted value exceeds a predetermined threshold, the system starts to issue an early warning or prompt. By comparing the output estimated value of the neural network model with the real-time state value of the parameter, when the real-time value of the parameter exceeds the upper and lower limits of the neural network estimated value, the parameter is warned in time to remind the monitoring personnel to pay attention and adjust. The health evaluation model uses an LSTM neural network algorithm model to analyze the health of the equipment. The LSTM neural network module includes three layers of neurons. The first, second and third layers have multiple neuron submodules respectively. The number of input variables is X, and the number of output variables is Y. The reasonable value of the output variable is predicted by the X input variables. To determine whether a device is in a healthy state, it is necessary to simultaneously detect several key indicators Y1, Y2, ..., YM of the device. For the monitored target parameter Y, it is determined that the N main explanatory variable measurement points that directly affect the change of Y are X1, X2, ..., XN. The synchronous historical data of Y and X in the historical time of the LSTM model are used to perform fitting training on the mapping relationship between X and Y. The training result is assigned to the online LTSM calculation model in the form of a weight matrix file. After the online calculation model is run, the predicted Y value calculated in real time by the explanatory variable X collected in real time online is compared with the actual instantaneous Y value. The cosine similarity of the vector is used to evaluate the health. The sub-health state of the device is reflected according to the degree of deviation between the actual vector and the estimated vector, and the system gives a specific health percentage for characterization.

7. The power plant intelligent monitoring and early warning system based on the neural network algorithm model according to claim 6 is characterized in that: The health evaluation model inputs the real-time values ​​of variables into the neural network model and makes real-time predictions on the output variables; The predicted values ​​of the output variables of the neural network model are tracked according to the actual variable parameters of the system operation status. At the same time, the health of the system is evaluated and scored using correlation analysis to achieve abnormal warning of the system and equipment health analysis, including health analysis of the coal mill, primary fan, blower, induced draft fan, feed water pump and high and low pressure systems of the generator set.

8. The power plant intelligent monitoring and early warning system based on the neural network algorithm model according to claim 6 is characterized in that: The control loop performance evaluation model evaluates the process performance of the control loop based on the set value and feedback value variables of the control system, the relative deviation of the set value and feedback value of the controlled object in the control loop, and the stability performance of the controlled object; The definition of controlled performance evaluation of a control loop includes transient evaluation and normal evaluation of the control loop; the evaluation of the ability to return to a new equilibrium state after the equilibrium state of the control loop is broken by interference excitation is called transient evaluation of the control loop, and the transient evaluation fault types of the main control loop are divided into: sharp increase in positive direction, sharp increase in negative direction, increase in vibration amplitude, increase in vibration frequency, increase in vibration amplitude frequency, sharp increase in positive direction, increase in vibration amplitude, sharp increase in positive direction, increase in vibration frequency, sharp increase in negative direction, increase in vibration amplitude, sharp increase in negative direction, and increase in vibration amplitude frequency; when the deviation between the set value and the feedback value within a defined period of time of the control loop is evaluated using the scientific statistical quantity root mean square to evaluate the average performance of the control loop, it is called normal evaluation of the control loop.

9. The power plant intelligent monitoring and early warning system based on the neural network algorithm model according to claim 6 is characterized in that: The actuator performance evaluation model, the actuator performance detection method based on signal trend analysis, determines the actuator fault type according to the actuator input variable and feedback variable data, and the fault type is divided into: sticking / slipping fault, actuator dead zone fault, actuator gain fault, actuator deviation fault, and actuator stuck fault; The fault diagnosis model constructs a coal mill mechanism model according to the coal mill mechanism and historical data; at the same time, according to the coal mill fault generation mechanism, the model parameters or control methods are adjusted to generate a large amount of typical fault sample data; the fault diagnosis function performs fault diagnosis and analysis on the coal mill, and performs logical judgment on the coal feeder flow, coal mill inlet air pressure, coal mill lower air pressure, coal mill outlet temperature, coal mill cold air damper opening, and coal mill hot air damper opening variables during operation, and diagnoses four fault states of coal mill normal, coal mill coal outage, coal mill coal blockage, and coal mill spontaneous combustion. The coal mill fault diagnosis types are divided into: normal, coal outage, coal blockage, and natural; The fault root cause analysis model is based on historical healthy operation data and uses an LSTM model to train the fuel flow, air supply, main steam flow, main feed water, and actual power. That is, one of the indicators is used as the target variable Y, and the remaining indicators are used as explanatory variables X to obtain a mapping relationship in the weight parameter matrix format, and then the other four groups of mapping relationships are obtained by analogy. During online operation, five sets of mapping relationship models are run simultaneously to obtain five predicted Y values, which are compared with the actual system values ​​in real time. According to the network theory of energy and material conservation, the one with the largest deviation is where the fault is located. When a fault occurs, the five sets of Y value deviation curves are queried, and the one that first produces deviation is the source of the fault.

10. A method for implementing a power plant intelligent monitoring and early warning system based on a neural network algorithm model, characterized in that: The power plant intelligent monitoring and early warning system based on the neural network algorithm model according to any one of claims 1 to 9 comprises the following steps: Step S1: using the massive historical data stored in the real-time operation process of the unit, selecting a time period that can cover the full operating range of the unit to collect the historical data; Step S2: Determine the input / output variables of the prediction model according to the actual situation; Step S3: data preprocessing, including outlier removal and missing value filling; Step S4: Offline training of data to obtain a training curve diagram, determine the correlation of input / output variables, and determine whether the input / output variables are highly fitted, so as to find input / output variable data that meets the preset conditions; Step S5: Neural network model configuration, using the neural network algorithm platform to perform program logic configuration on the system equipment of parameter warning, health evaluation, control loop performance evaluation, actuator performance evaluation, fault diagnosis, and fault root cause analysis of each system that needs warning of the generator set according to the determined input / output variables; Step S6: The prediction model is trained online, and the neural network model is trained with the preprocessed data. According to the number of system input / output variables and the system complexity, the number of hidden layers, the number of nodes, the weights, the learning rate, and the activation function of the deep neural network model are adjusted, and the training is repeated until the accuracy of the neural network prediction model meets the preset requirements; The normality test analysis method is used to detect outliers, and then the outliers are processed by linear regression. The specific process is: data collection and organization, including cleaning and formatting, for subsequent analysis; using statistical software to test the normality of the data, including drawing histograms, PP diagrams or QQ diagrams to intuitively determine whether the data distribution is close to the normal distribution; Outlier detection based on normal distribution, first, calculate the mean and standard deviation, calculate the mean μ and standard deviation σ of the entire data set; second, determine the outlier threshold, according to the characteristics of normal distribution, obtain the probability of a data point falling within the range of the mean plus or minus 3 times the standard deviation μ±3σ, and the data points outside this range are regarded as outliers; third, identify outliers, compare each data point with the threshold, if the value of a data point is less than μ-3σ or greater than μ+3σ, it is marked as an outlier; finally, record all identified outliers, and choose to delete or replace according to the actual situation, specifically: through linear regression processing, replace the samples with more serious data deviations from the overall data through the approximate function fitted by normal data, and use cubic spline interpolation to fill in the missing data.