CN115238573A - Hydroelectric generating set performance degradation trend prediction method and system considering working condition parameters - Google Patents

Hydroelectric generating set performance degradation trend prediction method and system considering working condition parameters Download PDF

Info

Publication number
CN115238573A
CN115238573A CN202210787450.4A CN202210787450A CN115238573A CN 115238573 A CN115238573 A CN 115238573A CN 202210787450 A CN202210787450 A CN 202210787450A CN 115238573 A CN115238573 A CN 115238573A
Authority
CN
China
Prior art keywords
generating set
hydroelectric generating
working condition
degradation
unit
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202210787450.4A
Other languages
Chinese (zh)
Inventor
李超顺
甘振豪
邓友汉
吴一凡
陈鹏
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Huazhong University of Science and Technology
Original Assignee
Huazhong University of Science and Technology
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Huazhong University of Science and Technology filed Critical Huazhong University of Science and Technology
Priority to CN202210787450.4A priority Critical patent/CN115238573A/en
Publication of CN115238573A publication Critical patent/CN115238573A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/20—Design optimisation, verification or simulation
    • G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/20—Administration of product repair or maintenance
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y04—INFORMATION OR COMMUNICATION TECHNOLOGIES HAVING AN IMPACT ON OTHER TECHNOLOGY AREAS
    • Y04S—SYSTEMS INTEGRATING TECHNOLOGIES RELATED TO POWER NETWORK OPERATION, COMMUNICATION OR INFORMATION TECHNOLOGIES FOR IMPROVING THE ELECTRICAL POWER GENERATION, TRANSMISSION, DISTRIBUTION, MANAGEMENT OR USAGE, i.e. SMART GRIDS
    • Y04S10/00—Systems supporting electrical power generation, transmission or distribution
    • Y04S10/50—Systems or methods supporting the power network operation or management, involving a certain degree of interaction with the load-side end user applications

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Business, Economics & Management (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Human Resources & Organizations (AREA)
  • Software Systems (AREA)
  • Economics (AREA)
  • General Engineering & Computer Science (AREA)
  • Strategic Management (AREA)
  • Artificial Intelligence (AREA)
  • Operations Research (AREA)
  • Quality & Reliability (AREA)
  • General Business, Economics & Management (AREA)
  • Mathematical Physics (AREA)
  • Tourism & Hospitality (AREA)
  • Computing Systems (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Medical Informatics (AREA)
  • Data Mining & Analysis (AREA)
  • Marketing (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Geometry (AREA)
  • Computer Hardware Design (AREA)
  • Biomedical Technology (AREA)
  • General Health & Medical Sciences (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Development Economics (AREA)
  • Game Theory and Decision Science (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention discloses a method and a system for predicting the performance degradation trend of a hydroelectric generating set by considering working condition parameters, and belongs to the field of state evaluation and early warning of the hydroelectric generating set. The invention adopts a probability interval prediction model which consists of a multi-head attention module, a Bi-GRU module and a quantile loss module. The input not only contains the historical degradation degree of the unit performance, but also contains the unit operation condition parameters. Firstly, a multi-head attention module is used for extracting hidden correlation characteristics among unit operation condition parameters; secondly, the Bi-GRU module is used for extracting a working condition hidden feature diagram and a time sequence feature of the unit degradation degree; and finally, obtaining a confidence probability prediction interval result of the unit performance degradation degree through a quantile loss module. The method overcomes the defect of inaccurate prediction of the unit performance degradation caused by abnormal unit vibration and swing due to environmental factors, realizes accurate prediction of the unit performance degradation trend, and improves the reliability of model prediction. And the method has stronger robustness due to the consideration of the parameter information of the unit operation condition.

Description

Hydroelectric generating set performance degradation trend prediction method and system considering working condition parameters
Technical Field
The invention belongs to the field of hydroelectric generating set state evaluation and early warning, and particularly relates to a hydroelectric generating set performance degradation trend prediction method and system considering working condition parameters.
Background
The hydroelectric generating set is a core device for building a hydropower station, ensures stable, safe and reliable operation of the hydropower station, is an important work of management, operation and maintenance personnel of the power plant, and is related to the safety of the hydropower station and whether a power grid can stably provide power load for social production. The hydroelectric generating set is influenced by multiple factors such as electromagnetism, machinery and water power, and along with the increase of the accumulated operation time of the generating set, abnormal vibration is easily caused to the generating set equipment, so that the generating set is fatigued, damaged or even deteriorated, the safe and reliable operation of the generating set is influenced, the safety and the stability of a power grid are greatly influenced, and the great social and economic losses are caused. The safe operation control theory of the hydroelectric generating set under the constraint of developing water-machine-electricity-magnetism multiple coupling is an important basis for guaranteeing the safety of hydraulic engineering, and has important significance for improving the utilization rate of water resources, ensuring the downstream ecological water use of hydropower stations and promoting the development of national economy. The hydroelectric generating set can experience the process of health state, performance degradation, equipment failure and even invalidation in the long-term operation process, if the operation state of the hydroelectric generating set can be monitored in the process of unit degradation, the change of the health state of the hydroelectric generating set is evaluated, the future degradation trend of the hydroelectric generating set is predicted, the abnormal condition of the hydroelectric generating set can be timely detected, a reasonable maintenance plan is made, the state maintenance of the hydroelectric generating set in a practical sense is realized, and the production benefit maximization of a power generating enterprise is guaranteed. As a key link of unit maintenance, the evaluation of the health state of the hydroelectric generating set and the prediction of the performance degradation trend have strong research value and application value for improving the stable operation maintenance level of the hydroelectric generating set and reducing the shutdown maintenance loss caused by predictable faults.
Therefore, in order to ensure safe and stable operation of the unit equipment, research on predicting the performance degradation trend of the hydroelectric generating set needs to be urgently developed. The method is characterized in that massive state monitoring data of the hydroelectric generating set are taken as a basis, a high-dimensional function mapping relation between unit operation parameters and unit operation states is analyzed, the health state of equipment of the hydroelectric generating set is mastered in real time, the health performance of the historical operation process of the hydroelectric generating set is evaluated, the future health state of the unit is predicted and evaluated, the abnormal state of the unit is found in time, the unit fault is judged scientifically and reasonably, corresponding overhaul and active maintenance are arranged, and decision technical support is provided for fault detection and active maintenance of the unit, so that the hydroelectric generating set is guaranteed to operate safely and stably.
The existing hydroelectric generating set degradation trend prediction method based on vibration signals has the problem that the selected health index is too local and single, so that the degradation trend prediction result is difficult to reflect the overall operation state of the hydroelectric generating set; under the complex operation condition, the precision of the degradation trend prediction model is insufficient, and the degradation trend at the future moment cannot be accurately obtained; in addition, the influence of the historical working condition parameters on the degradation trend prediction is ignored by the degradation prediction model, and the prediction effect and the robustness of the model are further limited.
Disclosure of Invention
Aiming at the defects of the prior art, the invention aims to provide a hydroelectric generating set performance degradation trend prediction method and a hydroelectric generating set performance degradation trend prediction system considering working condition parameters, and aims to solve the problems that the hydroelectric generating set degradation trend prediction method is difficult to reflect the whole operation state of the hydroelectric generating set, the prediction precision is insufficient, and the robustness is limited.
In order to achieve the above object, in a first aspect, the present invention provides a method for predicting a degradation trend of performance of a hydroelectric generating set in consideration of operating condition parameters, the method including:
s1, obtaining historical working condition parameters of the hydroelectric generating set, and calculating a degradation trend time sequence capable of reflecting the overall health state level and degradation performance of the hydroelectric generating set;
s2, inputting historical working condition parameters and historical degradation trend time sequences of the hydroelectric generating set into a trained probability interval prediction model to obtain a reliability prediction interval of the future degradation degree of the hydroelectric generating set;
the probability interval prediction model consists of a multi-head attention module, a Bi-GRU module and a quantile loss module; the multi-head attention module is used for extracting a hidden characteristic diagram of the operation working condition of the hydroelectric generating set from the historical working condition parameters of the hydroelectric generating set, and the hidden characteristic diagram is connected with the historical deterioration degree residual error of the hydroelectric generating set and then input into the Bi-GRU module; the Bi-GRU module is used for obtaining the degradation degree depth time sequence characteristics of the hydroelectric generating set through forward GRU and backward GRU calculation and outputting the degradation degree depth time sequence characteristics to the quantile loss module; and the quantile loss module is used for inputting the deterioration degree depth time sequence characteristics of the hydroelectric generating set into the feedforward full-connection layer and calculating to obtain the deterioration degree probability prediction results of the hydroelectric generating set under different quantiles.
It should be noted that, the samples of the training probability interval prediction model are the operating condition parameters and the historical degradation trend time series, and the labels are the trend time series at the future time.
Preferably, step S1 comprises:
s11, working condition parameters in the running process of the hydroelectric generating set are input into a trained single-channel health state model to obtain a vibration and oscillation health state value of each channel of the hydroelectric generating set under each moment running working condition, the vibration and oscillation health state value is compared with the real-time vibration and oscillation state parameters of the hydroelectric generating set to obtain the degradation degree of each channel, the single-channel health state model comprises a random forest and an RFECV, the random forest is used for fitting the relation between the working condition parameters and the vibration and oscillation state parameters, and the RFECV is used for eliminating non-important working condition parameters in the random forest training process;
s12, carrying out time fusion on the health degradation degree of the single component in each steady-state operation process of the hydroelectric generating set to obtain the health degradation degree of each component capable of describing the whole steady-state process of the hydroelectric generating set;
and S13, carrying out spatial fusion on the health degradation degrees of all parts of the hydroelectric generating set to obtain the fusion degradation degree of the hydroelectric generating set, thereby forming a time sequence of the performance degradation trend of the hydroelectric generating set.
It should be noted that the vibration oscillation of the hydroelectric generating set is selected as a research object, the key working condition parameter of each vibration oscillation is selected by using a recursive characteristic elimination method based on a random forest, a single-channel health performance index model is constructed, the functional relation between the running working condition parameter of the hydroelectric generating set and the dependent variables such as the vibration oscillation is established, and the multi-component health performance trend of the hydroelectric generating set is obtained. On the basis, a hydro-power generating unit health performance space-time model based on a steady-state process is constructed, and a performance degradation trend of a certain key component and the whole unit is formed, so that the whole health performance state of the unit is accurately reflected, a unit performance degradation trend time sequence is obtained, and data support is provided for subsequent prediction tasks. The samples for training the single-channel health state model are working condition parameters, and the labels are actual runout state parameters.
Preferably, in step S11, the degree of degradation DC under each channel i，t The calculation formula is as follows:
Figure BDA0003729411180000031
wherein, DC i，t Representing the degradation degree of the hydroelectric generating set at the t moment channel i, reLU representing a linear rectification function, R i，t Representing the actual oscillation state parameter, V, of the channel i at time t of the hydroelectric generating set i，t And (4) representing the oscillation health state value of the hydroelectric generating set channel i under the operation working condition at the moment t.
The degradation degree calculation formula is preferably adopted in the invention, and the ReLU can ensure that the degradation degree of a single sensing channel of the hydro-power generating unit is not less than 0.
Preferably, in step S12, the calculation formula of the health degradation degree of each component, which can describe the whole steady-state process of the hydroelectric generating set, is as follows:
Figure BDA0003729411180000041
Figure BDA0003729411180000042
where time t represents the start of a steady-state operating process, D j (T) represents the health deterioration degree of a component j at the moment T of the hydroelectric generating set, T represents the steady-state running time of the hydroelectric generating set from starting to stopping, and D j，t Representing the euclidean distance.
It should be noted that, the time fusion mode is preferred in the present invention, and since the degradation degree calculation considers each complete steady-state operation process, the obtained degradation degree of each component can more objectively describe the unit operation condition of the whole steady-state process.
Preferably, in step S13, the spatial fusion time series calculation formula of the performance degradation trend of the hydroelectric generating set is as follows:
Figure BDA0003729411180000043
wherein n represents the number of components in the hydroelectric generating set, t represents the starting time of a certain steady-state operation process, D (t) represents the time series of the performance degradation trend of the hydroelectric generating set, D j (t) represents the degree of health deterioration of the component j at time t of the hydroelectric power generating unit.
In the present invention, the spatial fusion method is preferred, and the deterioration degree obtained by comprehensively considering the deterioration states of the components can accurately reflect the overall health status of the unit.
Preferably, the multi-attention module includes a position coding unit, a first residual connection unit, a multi-head self-attention unit with a mask, a global tie pooling layer, a feed-forward layer, and a second residual connection unit;
the position coding unit is used for carrying out position coding on the historical working condition parameters of the hydroelectric generating set;
the first residual error connecting unit is used for inputting the residual error into the multi-head self-attention unit with the mask after connecting the position code and the working condition parameter;
the multi-head self-attention unit with the mask is used for extracting multi-head attention characteristics and inputting the extracted attention vectors into a feedforward layer;
the feedforward layer is used for being connected in series and combined to form a hidden characteristic diagram of the running working condition of the hydroelectric generating set;
and the second residual connecting unit is used for connecting the hidden characteristic diagram and the historical deterioration trend time sequence of the hydroelectric generating set through residual and then inputting the residual to the Bi-GRU module.
It should be noted that, the multi-head attention module with the above structure is preferred in the present invention, and the hidden correlation features between the unit operation condition parameters are extracted while the model training time is reduced by using the position coding and the multi-head attention mechanism, so that the obtained condition hidden feature map can overcome the problem of inaccurate degradation prediction caused by environmental factors, and improve the accuracy of degradation trend prediction.
In order to achieve the above object, in a second aspect, the present invention provides a system for predicting degradation trend of performance of a hydroelectric generating set in consideration of operating condition parameters, including: a processor and a memory;
the memory is for storing a computer program or instructions;
the processor is adapted to execute the computer program or instructions in the memory such that the method of the first aspect is performed.
Generally, compared with the prior art, the technical scheme conceived by the invention has the following beneficial effects:
the invention provides a hydroelectric generating set performance degradation trend prediction method considering working condition parameters. The input of the model not only comprises the historical deterioration degree of the unit performance, but also comprises the unit operation condition parameters. Firstly, a multi-head attention module is used for extracting hidden correlation characteristics among unit operation condition parameters; secondly, the Bi-GRU module is used for extracting a working condition hidden characteristic diagram and time sequence characteristics of unit degradation degree; and finally, obtaining a confidence probability prediction interval result of the unit performance degradation degree through a quantile loss module. The method overcomes the defect of inaccurate prediction of the unit performance degradation caused by abnormal unit vibration and swing due to environmental factors, realizes accurate prediction of the unit performance degradation trend, and improves the reliability of model prediction. Due to the fact that the parameter information of the unit operation working condition is considered, the model for predicting the performance degradation trend of the hydroelectric generating set has stronger robustness.
Drawings
FIG. 1 is a flow chart of a method for predicting a deterioration trend of a hydroelectric generating set according to an embodiment of the invention;
FIG. 2 is a diagram illustrating a time-space fusion degradation trend sequence result provided by an embodiment of the present invention;
FIG. 3 is a diagram of a degradation trend prediction model according to an embodiment of the present invention;
FIG. 4 is a diagram illustrating a prediction result of a degradation trend interval according to an embodiment of the present invention;
fig. 5 is a deterioration trend section prediction Q-Q diagram provided by the embodiment of the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and do not limit the invention.
The invention provides a method for predicting the degradation trend of a hydroelectric generating set, which adopts the following overall scheme: firstly, establishing a plurality of single-channel health index models by utilizing an RFECV (Recursive Feature Elimination Cross-validation, recursive Feature Elimination) + RF (Random forest) model, wherein the models realize input Feature importance scoring in a self-adaption mode in a training process, screening out optimal working condition features, and finally fitting the health models to obtain the relationship between the optimal working condition features and the vibration swing of each channel; calculating deterioration degree indexes of all channels according to the difference between the vibration swing degree health state value and the actual vibration swing value, calculating Euclidean distances of all channels to obtain the deterioration degree of each component, performing time fusion on the deterioration degree of each component by utilizing a steady-state operation process, and finally performing space fusion on each component to obtain a unit fusion deterioration degree; establishing a hydropower unit degradation trend prediction model based on Healthformer, extracting characteristics of historical operating condition information and unit performance historical degradation degree, and finally predicting to obtain a reliability prediction interval of future degradation degree.
Fig. 1 is a flowchart of a method for predicting a deterioration trend of a hydroelectric generating set according to an embodiment of the present invention. As shown in fig. 1, the invention provides a method for predicting the performance degradation trend of a hydroelectric generating set by considering working condition parameters, which comprises the following steps:
step (1): respectively establishing n single-channel health state index models based on RFECV + RF by taking the working condition parameters of the hydroelectric generating set as independent variables of the health state model and the vibration swing degree state parameters of the generating set in the health state as dependent variables of the model, wherein the models can be expressed as f i ：X t →V i，t ，i∈[1，...，n]Wherein X is t Representing the operating condition parameter, V, of the hydroelectric generating set at time t i，t Represents the ith dimension vibration swing measuring point at the time t, f i Representing the RFECV + RF single channel fitting model that the present invention needs to solve.
The step (1) specifically comprises the following steps:
(1-1): firstly, selecting state monitoring data collected in the early operation stage of the hydroelectric generating set (namely, the hydroelectric generating set is still in a complete health state) as a data set Z of a model, and randomly selecting 80% of samples from the data set as a model training set Z T And the rest 20 percent is taken as a model verification set Z V 。
(1-2): selecting a random forest model as a learner for feature screening, setting a minimum feature selection number, a removal feature number of each stage and cross validation times, and selecting a decision coefficient R 2 As an indicator of the importance score of each feature. Determining the coefficient R 2 The calculation formula of (a) is as follows:
Figure BDA0003729411180000071
Figure BDA0003729411180000072
wherein, y i The measured value is shown as an actual value,
Figure BDA0003729411180000073
represents the model predicted value, and n represents the number of samples.
(1-3): the training process of the RFECV + RF health model can be specifically divided into the following steps:
(1-3-1): will train data set Z T Inputting the feature into an RFECV + RF model to model an original feature set, calculating importance scores of all features, removing part of features with low importance scores, updating the feature set and obtaining a feature subset;
(1-3-2): according to the feature importance determined in the stage, different numbers of features are sequentially selected, the training subset is divided into a sub-training set and a sub-verification set, and the number of the features with the highest average score is determined by performing k-fold cross verification based on a random forest model on the selected feature set.
(1-3-3): if the characteristic quantity meets the condition of the optimal quantity of the characteristics, finishing training, finishing the characteristic screening of the working condition parameters, and storing a random forest model; otherwise, returning to (1-3-1) until the characteristic optimal condition is met.
Step (2): respectively inputting working condition parameters of a future operation process into a plurality of single-channel health state models, and comparing the vibration oscillation health state value with the real-time vibration oscillation state value of the hydroelectric generating set to obtain the degradation degree of each channel; solving the Euclidean distance corresponding to the degradation degree according to the spatial position relation of all the sensing channels of each component to realize spatial fusion; performing time fusion on the health degradation degree of a single component in the steady-state operation process of the hydroelectric generating set to obtain the health degradation degree of each component capable of describing the whole steady-state process of the hydroelectric generating set; fusing the health degradation degrees of the components to obtain a unit fusion degradation degree; the method specifically comprises the following steps:
(2-1): working condition parameter X of hydroelectric generating set in future operation process t′ Inputting the calculated state information into the single-channel health state model, and calculating to obtain the oscillation health state of each channel of the hydroelectric generating set under the operation working condition at the moment tValue V i，t′ And real-time runout state parameter R of hydroelectric generating set i，t′ Comparing to obtain the degradation degree DC of the channel i i，t′ The calculation formula is as follows:
Figure BDA0003729411180000081
Figure BDA0003729411180000082
here, reLU represents a linear rectification function, which is an activation function commonly used in a neural network.
(2-2): for each component, the Euclidean distance is obtained according to all the sensing channels (the swing component comprises X and Y directions, and the vibration component comprises X, Y and Z directions) of each component:
Figure BDA0003729411180000083
(2-3): then, carrying out time fusion on the single component health degradation degree in the steady-state operation process of the hydroelectric generating set to obtain each component health degradation degree D capable of describing the whole steady-state process of the hydroelectric generating set j The average value is obtained by the following formula.
Figure BDA0003729411180000091
Wherein T represents the steady-state operation time of the unit in the process from starting to stopping.
(2-4): and finally, performing spatial fusion on the health degradation degrees of the n parts of the unit to obtain a fusion degradation degree D of the hydroelectric generating set, thereby forming a time sequence of the performance degradation trend of the hydroelectric generating set.
Figure BDA0003729411180000092
And (3): a hydroelectric generating set degradation trend prediction model based on Healthformer is constructed, and the model comprises three modules of attention, bi-GRU and Quantum Loss. Firstly, position coding and multi-head attention feature extraction are carried out on unit historical working condition parameters, then residual errors are carried out on the extracted attention vectors and the unit historical deterioration degree to form the input of a Bi-GRU module, a feedforward layer is used as a regressor, the condition quantiles are calculated, finally the quantile loss of the model is evaluated, and the model parameters are updated and optimized according to the quantile loss.
The step (3) specifically comprises the following steps:
(3-1): the method comprises the steps of carrying out position coding and multi-head attention feature extraction on unit historical working condition parameters, and then carrying out residual connection on the extracted attention vectors and unit historical deterioration degrees.
The specific steps of position coding and multi-head attention are as follows:
(3-1-1): for operating condition parameter x wc Coding is carried out according to the time sequence position to obtain time sequence position information p = (p) 1 ，...，p m ) T And connecting p and x in a residual manner wc 。
(3-1-2): the multi-head attention mechanism adopts a zoom dot product attention by a Query-Key-Value calculation method, and a calculation formula is given by the following formula:
Figure BDA0003729411180000101
in the formula: q denotes a query matrix, K denotes a matrix of correlation of the queried information with other information, and V denotes a matrix of the queried information. At the same time, the user can select the desired position,
Figure BDA0003729411180000102
wherein N and M represent the length of the query matrix and the length of the matrix of the queried information, respectively, D k And D v Representing the dimensions of the matrices K (or Q) and V, respectively. The calculation formula of the multi-head attention method is as follows:
MultiHead(Q，K，V)＝Concat(head 1 ，...，head H )W O
wherein the content of the first and second substances,
Figure BDA0003729411180000103
Figure BDA0003729411180000104
and W O All are parameter matrixes to be trained.
And inputting the residual sum into a multi-head self-attention network with a mask for inquiring and matching, calculating the similarity through a Query-Key-Value method to form an attention matrix, and performing flattening operation by using a one-dimensional global average pooling layer to form an attention vector.
x h ＝MultiHead(x wc +x p )
(3-1-3): finally, the flattened attention vector is input to the feedforward fully-connected layer FF 1 In the method, the unit operation condition hidden characteristic diagram is formed by combining in series
Figure BDA0003729411180000105
And providing effective characteristic parameters for final prediction.
Figure BDA0003729411180000106
(3-2): hiding characteristic diagram of operation condition
Figure BDA0003729411180000107
And the historical deterioration degree x of the unit D Performing residual error connection; secondly, calculating by a forward GRU and a backward GRU to obtain a unit deterioration degree depth time sequence characteristic x gru . The calculation formula is as follows:
Figure BDA0003729411180000108
(3-3): will feed forward layer FF 2 As a generation conditionRegressor of quantile prediction results, x gru And inputting the result into the computer, and calculating to obtain the unit degradation degree probability prediction results under different quantiles gamma.
y D，γ ＝FF 2 (x gru )
And calculating the error between the model output and the true value by adopting a quantile loss function, wherein the quantile loss calculation formula is as follows:
Figure BDA0003729411180000111
wherein, y i And
Figure BDA0003729411180000112
respectively representing the true value and the predicted value, and the value range of the quantile gamma is usually [0, 1%]In between. After the model training is finished, the probability predicted values when the quantile gamma is 0.95 and 0.05 are taken as the upper and lower boundaries of the output interval of the prediction model respectively.
Examples
For a clearer explanation of the invention, the advantages of the invention are highlighted, and the invention is further explained by monitoring and analyzing time series database data in a fault diagnosis system of a certain hydropower station state.
Step1: the time span of the data set is from 31/5/2017 to 20/11/2020, the input quantity of the model comprises 13-dimensional characteristic variables such as active power and reactive power, and the output quantity of the model comprises 15 dependent variables such as vibration of each rack and the swing of each part of a large shaft. And selecting data from 31 days in 5 months in 2017 to 31 days in 5 months in 2018 to train and verify the health model, and using the data from 1 day in 6 months in 2018 to 20 days in 11 months in 2020 to generate the time sequence of the performance degradation trend of the hydroelectric generating set. Experimental data were recorded as 8: the scale of 2 is divided into training set and validation set, obtaining 29091 pieces of training data and 7273 pieces of validation data, each piece of data comprises 13-dimensional feature input and 15-dimensional output.
Because a plurality of environment operation working condition parameters influence the vibration and the swing degree of the hydroelectric generating set, useless environment characteristic parameters are eliminated by adopting a recursive characteristic elimination method, and key working condition parameters with stronger vibration and swing correlation with the generating set are obtained so as to improve the accuracy of a single-component health model of the generating set. And respectively calculating and solving relevant working condition parameters of key components of 6 large units, such as an upper guide bearing, a lower guide bearing, a water guide bearing, the vibration of an upper frame, a lower frame, a top cover and the like, under different sensing channels.
The recursive feature elimination algorithm adopts random forest regression as a learner; selecting 5-fold cross validation as a cross validation model; the evaluation index is a determination coefficient R 2 (ii) a The minimum number of selected features and the number of recursive feature deletions are set to 4 and 1, respectively. The random forest model consists of 20 decision trees, sampling is carried out in a replacement mode, mean square error MSE is selected as an evaluation index of the decision trees, and random seeds are 42.
Table 1 gives the results of ranking experiments on the significance signature on a certain hydropower station data set. It can be seen that all the oscillation parameters have strong correlation with the cooling water inlet O temperature, and the cooling water inlet temperature is related to the seasonality, which indicates that the oscillation value of the hydroelectric generating set is seasonally changed; the unit runout parameters and common working condition parameters in researches such as water head, active power, reactive power and guide vane opening degree have strong correlation; the correlation between the unit runout and the working conditions such as exciting current and exciting voltage is not strong.
TABLE 1 ranking results of importance of characteristic of operating condition parameters (smaller value indicates higher importance)
Figure BDA0003729411180000121
Step2: and (3) taking the characteristic input and output obtained in the RFECV stage as data bases, and constructing a random forest fitting model. Firstly, operating a trained health state index model based on random forests by taking working condition parameter data as input to obtain 15 single-channel regression results. According to the sensing channel of each component, the Euclidean distance of the sensing channel is obtained to obtain the performance degradation trends of 6 key components such as an upper, a lower and a water guide bearing, an upper frame, a lower frame, a top cover and the like; performing time mean value fusion processing on each component according to a steady-state operation process; and finally, carrying out spatial fusion on the degradation degrees of all the components to obtain a time sequence of the fusion degradation trend of the performance of the hydroelectric generating set shown in the figure 2. As can be seen from fig. 2, the overall deterioration degree of the unit gradually increases from 0 to 0.4 in two years and half, and shows an increasing trend with time, which indicates that the deterioration degree of the unit gradually increases in the operation process; the accuracy of the health state index model provided by the invention is reflected from the side by the small reduction of the overall deterioration degree of the unit after the unit is overhauled.
Step3: five variables such as temperature, active power, reactive power, guide vane opening degree and water head are used as unit working condition parameter input of a Healthformer model, and the Healthformer structure is shown in figure 3. The trend of the temperature of the upper water guide inlet, the trend of the temperature of the lower water guide inlet and the trend of the temperature of the water guide inlet are consistent, namely the trend of the temperature of the upper water guide inlet and the trend of the temperature of the lower water guide inlet are consistent with the ambient temperature, and therefore the temperature of the upper water guide inlet is selected as a temperature variable representation. For the working condition parameters, due to the lack of data with the same time sequence interval, a near interpolation principle is selected for data interpolation, and final experimental data of the Healthormer model are formed. The first 80% of the data in the resulting dataset were used for model training and the last 20% for model testing. The number and the dimension of the multi-head attention head of the Healthormer model are set to be 8, the number of hidden neurons of the bidirectional GRU network is 8, the neurons of the two feedforward layers are respectively set to be 8 and 32, and the dimension of the final output layer is 9, namely 9 different quantiles are included: [0.005,0.025,0.05,0.1,0.5,0.9,0.95,0.975,0.995]. Fig. 4 shows the actual value of the unit degradation degree, the prediction model interval, the midpoint of the prediction interval, and the relative error between the midpoint of the prediction interval and the true value. It can be seen from fig. 4 that the proposed Healthormer model keeps the interval width small enough, and the difference between the predicted interval center value and the actual unit degradation degree value is small, and the distribution is uniform. The Q-Q graph intuitively shows whether the cumulative distribution function of the predicted values obeys uniform distribution or not by calculating the PIT value, and is used for verifying the reliability of the model, and FIG. 5 is a degradation trend interval prediction Q-Q graph provided by the embodiment of the invention. As shown in fig. 5, the PIT values of the health former model are uniformly distributed around the diagonal (i.e., theoretical distribution), and almost all PIT points are located near the 5% confidence boundary, which indicates that the prediction results of the proposed model are subject to uniform distribution and have strong reliability.
It will be understood by those skilled in the art that the foregoing is only a preferred embodiment of the present invention, and is not intended to limit the invention, and that any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the scope of the present invention.

Claims (7)

1. A hydroelectric generating set performance degradation trend prediction method considering working condition parameters is characterized by comprising the following steps:
s1, obtaining historical working condition parameters of the hydroelectric generating set, and calculating a degradation trend time sequence capable of reflecting the overall health state level and degradation performance of the hydroelectric generating set;
s2, inputting historical working condition parameters and historical degradation trend time sequences of the hydroelectric generating set into a trained probability interval prediction model to obtain a reliability prediction interval of the future degradation degree of the hydroelectric generating set;
the probability interval prediction model consists of a multi-head attention module, a Bi-GRU module and a quantile loss module; the multi-head attention module is used for extracting a hidden characteristic diagram of the running working condition of the hydroelectric generating set from the historical working condition parameters of the hydroelectric generating set, and the hidden characteristic diagram is connected with the historical deterioration degree residual error of the hydroelectric generating set and then is input to the Bi-GRU module; the Bi-GRU module is used for obtaining the degradation degree depth time sequence characteristics of the hydroelectric generating set through forward GRU and backward GRU calculation and outputting the degradation degree depth time sequence characteristics to the quantile loss module; and the quantile loss module is used for inputting the deep time sequence characteristics of the degradation degree of the hydroelectric generating set into the feedforward full connection layer and calculating to obtain the probability prediction results of the degradation degree of the hydroelectric generating set under different quantiles.
2. The method of claim 1, wherein step S1 comprises:
s11, working condition parameters in the running process of the hydroelectric generating set are input into a trained single-channel health state model to obtain a vibration and oscillation health state value of each channel of the hydroelectric generating set under each moment running working condition, the vibration and oscillation health state value is compared with the real-time vibration and oscillation state parameters of the hydroelectric generating set to obtain the degradation degree of each channel, the single-channel health state model comprises a random forest and an RFECV, the random forest is used for fitting the relation between the working condition parameters and the vibration and oscillation state parameters, and the RFECV is used for eliminating non-important working condition parameters in the random forest training process;
s12, time fusion is carried out on the single component health degradation degree of the hydroelectric generating set in each steady state operation process, and the health degradation degree of each component capable of describing the whole steady state process of the hydroelectric generating set is obtained;
and S13, carrying out spatial fusion on the health degradation degrees of all parts of the hydroelectric generating set to obtain the fusion degradation degree of the hydroelectric generating set, thereby forming a time sequence of the performance degradation trend of the hydroelectric generating set.
3. The method according to claim 2, wherein in step S11, the degree of degradation DC under each channel i，t The calculation formula is as follows:
Figure FDA0003729411170000021
wherein, DC i，t Representing the degradation degree of the hydroelectric generating set at the t moment channel i, reLU representing a linear rectification function, R i，t Representing the actual oscillation state parameter, V, of the channel i at time t of the hydroelectric generating set i，t And the oscillation health state value of the hydroelectric generating set channel i under the operation working condition at the time t is represented.
4. The method according to claim 2, wherein in step S12, the health degradation calculation formula of each component that can describe the whole steady-state process of the hydroelectric generating set is as follows:
Figure FDA0003729411170000022
Figure FDA0003729411170000023
where time t represents the start of a steady-state operating process, D j (T) represents the health deterioration degree of a component j at the moment T of the hydroelectric generating set, T represents the steady-state running time of the hydroelectric generating set in the process from starting to stopping, D j，t Representing the euclidean distance.
5. The method according to claim 2, wherein in step S13, the spatial fusion time series calculation formula of the degradation trend of the performance of the hydroelectric generating set is as follows:
Figure FDA0003729411170000024
wherein n represents the number of components in the hydroelectric generating set, time t represents the starting time of a certain steady-state operation process, D (t) represents the time series of the performance degradation trend of the hydroelectric generating set, D j (t) represents the degree of health deterioration of the component j at time t of the hydroelectric power generating unit.
6. The method of any of claims 1 to 5, wherein the multi-attention module comprises a position coding unit, a first residual concatenation unit, a masked multi-headed self-attention unit, a global horizon pooling layer, a feed forward layer, and a second residual concatenation unit;
the position coding unit is used for carrying out position coding on the historical working condition parameters of the hydroelectric generating set;
the first residual error connecting unit is used for inputting the residual error into the multi-head self-attention unit with the mask after connecting the position code and the working condition parameter;
the multi-head self-attention unit with the mask is used for carrying out multi-head attention feature extraction and inputting the extracted attention vector to a feedforward layer;
the feedforward layer is used for being connected in series and combined to form a hidden characteristic diagram of the running working condition of the hydroelectric generating set;
and the second residual connecting unit is used for connecting the hidden characteristic diagram and the historical deterioration trend time sequence of the hydroelectric generating set through residual and then inputting the residual to the Bi-GRU module.
7. A hydroelectric generating set performance degradation trend prediction system considering operating condition parameters is characterized by comprising the following components: a processor and a memory;
the memory is for storing a computer program or instructions;
the processor is for executing the computer program or instructions in the memory, causing the method of any of claims 1-6 to be performed.
CN202210787450.4A 2022-07-04 2022-07-04 Hydroelectric generating set performance degradation trend prediction method and system considering working condition parameters Pending CN115238573A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210787450.4A CN115238573A (en) 2022-07-04 2022-07-04 Hydroelectric generating set performance degradation trend prediction method and system considering working condition parameters

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210787450.4A CN115238573A (en) 2022-07-04 2022-07-04 Hydroelectric generating set performance degradation trend prediction method and system considering working condition parameters

Publications (1)

Publication Number Publication Date
CN115238573A true CN115238573A (en) 2022-10-25

Family

ID=83671452

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210787450.4A Pending CN115238573A (en) 2022-07-04 2022-07-04 Hydroelectric generating set performance degradation trend prediction method and system considering working condition parameters

Country Status (1)

Country Link
CN (1) CN115238573A (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115619287A (en) * 2022-11-14 2023-01-17 湖北工业大学 Multi-source data fusion-based hydroelectric generating set state degradation evaluation method and system
CN117110871A (en) * 2023-10-13 2023-11-24 北京中航科电测控技术股份有限公司 Test bench for high-power density permanent magnet motor

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115619287A (en) * 2022-11-14 2023-01-17 湖北工业大学 Multi-source data fusion-based hydroelectric generating set state degradation evaluation method and system
CN117110871A (en) * 2023-10-13 2023-11-24 北京中航科电测控技术股份有限公司 Test bench for high-power density permanent magnet motor
CN117110871B (en) * 2023-10-13 2024-05-14 北京中航科电测控技术股份有限公司 Test bench for high-power density permanent magnet motor

Similar Documents

Publication Publication Date Title
CN105302096B (en) Intelligent factory scheduling method
CN115238573A (en) Hydroelectric generating set performance degradation trend prediction method and system considering working condition parameters
de la Hermosa González Wind farm monitoring using Mahalanobis distance and fuzzy clustering
CN110163429B (en) Short-term load prediction method based on similarity day optimization screening
US20230213895A1 (en) Method for Predicting Benchmark Value of Unit Equipment Based on XGBoost Algorithm and System thereof
CN102183699A (en) Method for model mismatching detection and positioning of multivariate predictive control system in chemical process
CN115294671A (en) Air compressor outlet pressure prediction method and prediction system
Zhengcai et al. Bottleneck prediction method based on improved adaptive network-based fuzzy inference system (ANFIS) in semiconductor manufacturing system
Li et al. Deep spatio-temporal wind power forecasting
CN112529053A (en) Short-term prediction method and system for time sequence data in server
CN113689072B (en) Marine wind turbine running state evaluation method based on Markov chain
Bao et al. Wind turbine condition monitoring based on improved active learning strategy and KNN algorithm
Karuppusamy Machine learning approach to predictive maintenance in manufacturing industry-a comparative study
CN117744859A (en) Wind turbine generator fault early warning method based on self-adaptive double control strategy
CN116821610B (en) Method for optimizing wind power generation efficiency by utilizing big data
Surjandari et al. Fault detection system using machine learning on geothermal power plant
Bond et al. A hybrid learning approach to prognostics and health management applied to military ground vehicles using time-series and maintenance event data
Liu et al. Residual useful life prognosis of equipment based on modified hidden semi-Markov model with a co-evolutional optimization method
CN116032020A (en) Intelligent monitoring method and system for new energy power station
CN114971024A (en) Fan state prediction method and device
Won et al. Prediction of remaining useful lifetime of membrane using machine learning
Ou et al. Fault Prediction Model of Wind Power Pitch System Based on BP Neural Network
CN117893030B (en) Power system risk early warning method based on big data
Yu et al. An Economic Perspective on Predictive Maintenance of Filtration Units
Chen et al. A Method for Predicting the Remaining Useful Life of Aircraft Engines Based On NLSTM and Feature Optimization Strategy

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination