CN106682781A - Power equipment multi-index prediction method - Google Patents
Power equipment multi-index prediction method Download PDFInfo
- Publication number
- CN106682781A CN106682781A CN201611263834.7A CN201611263834A CN106682781A CN 106682781 A CN106682781 A CN 106682781A CN 201611263834 A CN201611263834 A CN 201611263834A CN 106682781 A CN106682781 A CN 106682781A
- Authority
- CN
- China
- Prior art keywords
- data
- neuron
- index
- equipment
- layer
- 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
Links
Classifications
-
- 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"
Landscapes
- Engineering & Computer Science (AREA)
- Business, Economics & Management (AREA)
- Human Resources & Organizations (AREA)
- Economics (AREA)
- Strategic Management (AREA)
- Marketing (AREA)
- Game Theory and Decision Science (AREA)
- Entrepreneurship & Innovation (AREA)
- Development Economics (AREA)
- Operations Research (AREA)
- Quality & Reliability (AREA)
- Tourism & Hospitality (AREA)
- Physics & Mathematics (AREA)
- General Business, Economics & Management (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
A power equipment multi-index prediction method includes the main steps of establishing a model, operating the model, obtaining a power equipment multi-index prediction result and the like. A generalized regression neural network (GRNN) is adopted to learn historical data, regulations of indexes when equipment is running are simulated, and thus corresponding values are predicated. The power equipment multi-index prediction method has the characteristics of high universality, robustness and self-adaptability of a method based on data driving, is fast in prediction speed, can predicate many values at the same time, and well meets demands of pre-estimation of equipment relevant indexes during actual operations of a power plant.
Description
Technical field
The present invention relates to detection field, specific design equipment condition monitoring field, more particularly to one kind are returned based on broad sense
Return the power equipment multi objective Forecasting Methodology of neutral net (GRNN).
Background technology
It is well known that the running status of equipment is huge for power plant's Influence of production, such as the boiler, generating in Power Plant
The key equipments such as machine.Functional character, external performance and electric characteristic that equipment is embodied in nominal situation running
It is discrepant with being characterized in that under abnormality.According to equipment state early warning, repair schedule is made, the people huge by power plant is saved
Power, material resources, financial resources, with very big economic potential.
At present, mainly there is three major types method in equipment state warning aspect：One is according to manufacture using equipment production firm
The threshold method of technique initialization all parts difference warning level；Two is Knowledge based engineering method, mainly with associated specialist and behaviour
Annexation, failure based on making the enlightening Heuristics of personnel, during qualitative or quantitative description between each unit
Communication mode etc., by inferential capability of the mode simulation process expert such as reasoning, deduction in monitoring, so as to be automatically performed equipment
Fault pre-alarming and monitoring of equipment；Three is that the mathematics method for digging modeled using non-linear multi-objective planning fits equipment various pieces
Between complex redundancy relation, carry out early warning by analyzing the gap between actual value and assessed value.Three kinds of methods have respectively
From the characteristics of and limitation.
In prior art, disclosed Forecasting Methodology mainly has：1.《Association rules method is improved in electrical equipment fault
Application in prediction》Intelligent grid, 10 phases 3 in 2015；2.《A kind of power equipment time-varying stoppage in transit mould suitable for repair based on condition of component
Type》Proceedings of the CSEE, 25 phases in 2013；3.《A kind of equipment fault early-warning and state monitoring method》,
CN104102773A；4.《Power equipment temperature rise model and its application in current-carrying failure predication》Electric Machines and Control, 2013
07 phase of year；5.《Application of the multidimensional time-series association analysis method in electrical equipment fault prediction》Electrical network and clean energy resource,
12 phases in 2014；6. periodical：《Application of the local optimum weighted regression algorithm in power equipment failure prediction》Computer measurement
With control, 01 phase in 2014, but aforesaid way have it is predictive it is poor, according to historical data, limitation is larger, and mode is answered
It is miscellaneous, the shortcomings of computationally intensive.
For traditional fixed threshold method, determined by the manufacturer of equipment completely.Although this conventional method compares
Insurance, but, but ignore a very important problem.Equipment is dynamic change, is not unalterable, with fortune
The increase of the row time limit, its upper side part has different degree of degenerations because of material, usage degree, the degree of wear, respectively refers on equipment
The inherent law of the equipment operation of mark reflection is also in change.So there is error in traditional fixed threshold method, may
Cause false alarm, fail to report alert phenomenon.
Knowledge based engineering method mainly based on the suggestive working experience of associated specialist and operating personnel, to expert's sheet
The dependence of body is stronger, rather than based on tight mathematical logic, it is impossible to accurate Mathematical Modeling is built, alarming result is still inadequate
Ideal, conventional method mainly includes expert system, failure decision tree, digraph, fuzzy logic etc..Then people increasingly note
The use of nonlinear regression model (NLRM) method is re-established, its general principle is fully to excavate to be with existing Mathematical Modeling Methods
The mass historical data of system equipment sets up the model of highly effective and carries out equipment real-time status evaluation.This method is directly by being
The historical data of system sets up fault pre-alarming model, it is not necessary to know the accurate mechanism model of system, therefore its versatility and adaptive
Should be able to power it is all stronger.On monitoring of equipment and warning algorithm it can be divided into again based on signal transacting, rough set, machine learning,
Information fusion and multivariate statistics this five big class algorithm, wherein machine learning algorithm are to develop to enliven the most in theory and practice
Branch, it includes Bayes classifier, and neutral net, SVMs, correlation rule, k nearest neighbor algorithms, clustering algorithm is main
Constituent analysis scheduling algorithm.Although the method achieves certain effect in power plant's real work, it is not very good to remain unchanged,
Still have much room for improvement.
Existing nonlinear regression modeling method is the historical data using index under equipment, sets up model, draws and comments
Whether valuation, obtain residual error to judge whether to break down, report to the police according to instantaneous value and assessed value.Power plant staff is according to reality
Need of work, is often also required to estimate some indexs in advance, to pass judgment on the status information of equipment, arranges the work of next step.
Existing many workers propose the algorithm model in the fields such as machine learning, multivariate statistics answering in terms of power system
With, however, do not have related algorithm model to directly apply on power equipment index prediction yet, and current field of power
Prediction algorithm be generally single attribute forecast, only according to single attribute historical data modeling, limitation is larger.
The content of the invention
It is an object of the invention to overcome the deficiencies in the prior art, there is provided a kind of power equipment multi objective Forecasting Methodology,
Based on non-linear multi-objective planning model method, status monitoring, fault pre-alarming new approaches and new method are creatively proposed, with wide
Based on adopted recurrent neural networks, equipment multi objective historical data is modeled, complete the prediction to equipment index, be given with index
Predicted value is the new approaches of alarm decision foundation, has filled up the blank of equipment multi objective Forecasting Methodology, there is provided one kind is not only suitable for
Complex Nonlinear System, the equipment multi objective Forecasting Methodology for being easy to staff to analyze and process again, based on device data, comprehensive profit
With the information of multiple indexs of correlation, rather than simply returned using broad sense according to the historical information of the single index to be predicted
Return neutral net (GRNN), historical data is learnt, rule of each index when equipment runs is simulated, so as to predict
Corresponding value.The present invention does not only give equipment condition monitoring and early warning new approaches, and with based on data-driven class method
The characteristics of versatility, robustness, strong adaptive ability；Meanwhile, predetermined speed of the present invention is fast, can simultaneously predict multiple values, preferably
Solve the demand for needing to estimate in power plant's real work to equipment index of correlation.
The invention provides a kind of power equipment multi objective Forecasting Methodology, in turn includes the following steps：
(1) model is set up：
Step 1.1：Training data is obtained, in Power Plant DCS System lane database, under finding target device and equipment
All indication informations, are spaced according to certain peek, and selected equipment historical data under normal operating conditions carries out pre- place
Reason；
Step 1.2：According to actual needs, main index of correlation is selected；
Step 1.3：Main index of correlation to selecting is normalized, and obtains training data；
Step 1.4：Training data is processed using validation-cross method；
Step 1.5：Data to dividing after the process of validation-cross method are trained, and find optimal parameter, obtain optimal
Parameter model；
(2) moving model, obtains power equipment multi objective and predicts the outcome：
Step 2.1：Real time data is obtained, from Power Plant DCS System lane database real time data when equipment runs is obtained, and
Do the pretreatment operation consistent with step 1；
Step 2.2：According to actual needs, corresponding achievement data is selected；
Step 2.3：To selecting corresponding achievement data to carry out the normalized consistent with step 1.3, after being processed
Achievement data；
Step 2.4：Achievement data after the process that step 2.3 is obtained, is input to optimal parameter model, obtains processing knot
Really；
Step 2.5：The result obtained to step 2.4 carries out renormalization process, is predicted the outcome.
Further, also include that the data to choosing are pre-processed in step 1.1 or step 2.1, remove achievement data
More than the record that the record and system reason of index bound cause some value disappearances.
Further, step 1.2 is concretely comprised the following steps：
Hypothesis finally gives N number of index under certain equipment, and M bars are recorded, then the m article of record regards the vector of a N-dimensional as,
It is expressed as：Xm=[xm1,xm2,...xmN], then training data is the matrix form of M*N, and concrete form is as follows：
Further, normalized described in step 1.3 or step 2.3 is specifically carried out according to equation below：
xnew=(2x-2xmax)/(xmax-xmin)+1
Wherein, xnewFor the value after x normalization, xmaxFor the maximum of main index of correlation, xminRefer to for main correlation
Target minimum of a value.
Further, step 1.4 is specially：Using validation-cross method, the equal portions of training data 4 take out every time a
Used as inspection, remaining 3 parts are used for training.
Further, step 1.5 is specially：Using generalized regression nerve networks GRNN mode to the process of validation-cross method
The data for dividing afterwards are trained, and obtain the optimal parameter model based on training data error minimum, specially：
Assume map network input X=[x1,x2,...xm]T, wherein being output as Y=[y1,y2,...yT]T；
The neuron number of input layer is equal to the dimension m of input layer in learning sample, and each neuron is that simple distribution is single
Unit, directly passes to hidden layer by input variable；
The neuron number of mode layer is equal to the number n of learning sample, the different sample of each neuron correspondence, in mode layer
The transmission function of neuron i is：
Wherein, X is network inputs amount, and Xi is god
The Jing units corresponding learning samples of i, σ is smoothing parameter, and neuron i is output as between the corresponding sample Xi of input variable X
Euclidean distance square exponential form：
D2=(X-Xi)T(X-Xi)；
Summation layer includes two class neurons, and output of the one of which neuron to all mode layer neurons carries out counting asking
With each neuron of mode layer is 1 with the connection weight of the neuron, and transmission function is：
Output of other neurons to all mode layer neurons is weighted summation, in mode layer i-th neuron with
Connection weight in summation layer between j-th summation neuron is i-th output sample YiIn j-th element yij, summation god
Jing unit transmission function be
Neuron number in output layer is equal to the dimension L of output vector in learning sample, and each neuron is by layer of suing for peace
Output is divided by, i.e.,
yi=SNj/SDJ=1,2 ..., L.
Further, step 2.2 specially selects the achievement data for authorizing out through expert or staff,
Include purpose prediction index, it is remaining correspondence pointer type, number should be consistent with the pointer type of training data, number.
Further, the concrete of step 2.5 carries out renormalization process using equation below：
Wherein, xpFor a predicted value of correspondence index, xmaxThe maximum of the main index of correlation obtained for step 1.3
Value, xminThe minimum of a value of the main index of correlation obtained for step 1.3, xrFor the predicted value of correspondence index after renormalization.
The power equipment multi objective Forecasting Methodology of the present invention, it is possible to achieve：
1. the present invention provides new thinking for equipment condition monitoring and fault pre-alarming.Traditional equipment condition monitoring and therefore
Barrier Early-warning Model, under facility for study the historical data of health assessing current instantaneous value, then according to instantaneous value and assessed value
Obtain residual error to judge whether to report to the police；The invention is proposed, not only can may be used also using historical data and current real-time data
With using " following " predicted value come assessment equipment state, as the new foundation for whether producing alarm signal.
2. the present invention is based on data-driven, is effectively utilized the historical data of equipment index of correlation, big by neutral net
The training of data volume, rule of each index of Step wise approximation when equipment runs, the final different indexs being capable of achieving to distinct device
Prediction, for complicated nonlinear system has very strong versatility.
3. traditional monodrome prediction algorithm is only modeled prediction according to single index (attribute) historical data, does not account for setting
Standby correlation between upper index and index, can inevitably affect the accuracy for predicting the outcome.Present invention comprehensive utilization
The historical information of multiple indexs, rather than the historical information of single index is based only on, take into full account the phase between multi objective
Mutually affect, can more disclose the complicated causality and conditional relationship implied between index, it is to avoid effective information lacks in a large number
The possibility of mistake, as a result relatively more rationally, accurately.
4. the present invention can be with flexible and changeable, and the index number, type under same equipment can change according to actual needs, tool
There is stronger adaptive ability, can solve the problem that power plant staff to distinct device, the forecast demand of different indexs.
5. the present invention can continue prediction downwards according to predicting the outcome, and realize rolling forecast, and completing power plant's real work will
The prediction in time asked, the even long-term forecast in certain error allowed band.
6. traditional monodrome prediction algorithm model, can only predict every time a value, and efficiency is more low, and the present invention is with broad sense
Recurrent neural networks are model, simulate mutual rule of each index when equipment runs, and can simultaneously realize that multi objective is predicted, effect
The monodrome forecast model efficiency such as the relatively existing BP neural network of rate, SVM will height.
Description of the drawings
Fig. 1 is the flow chart for setting up model
Fig. 2 is the flow chart of moving model
Fig. 3 is the structural representation of generalized regression nerve networks (GRNN)
Fig. 4 is pressure fan hot blast air quantity predicted value and actual value comparison diagram
Specific embodiment
The following detailed description of being embodied as the present invention, it is necessary to it is pointed out here that, below implement to be only intended to this
Bright further illustrates, it is impossible to be interpreted as limiting the scope of the invention, and art skilled person is according to above-mentioned
Some nonessential modifications and adaptations that the content of the invention is made to the present invention, still fall within protection scope of the present invention.
The invention provides a kind of power equipment multi objective Forecasting Methodology,
The present invention is a kind of power equipment multi objective Forecasting Methodology based on generalized regression nerve networks, and it is fully excavated
A large amount of historical informations of index of correlation, and a nonlinear multivariable system is established, when analog machine runs between each index
Rule, can be with the multiple finger target values of real-time estimate.The method mainly includes setting up two processes of model and moving model.
Fig. 1 is the flow chart that the present invention sets up model, and whole modeling process is mainly included the following steps that：
Step 1：Obtain training data.
In Power Plant DCS System lane database, target device and all indication informations under equipment are found, according to certain
Peek interval, selected equipment historical data under normal operating conditions.
Necessary pretreatment of doing to the data got, removes the record that achievement data exceedes index bound, removes system
System reason causes the record of some value disappearances, also some other necessary pretreatment operation for removing exception.After screening
Data should meet claimed below：(1) equipment normal operational condition as complete as possible is covered；(2) record all per data
Represent a normal operating condition of equipment；(3) each desired value per data in record should be the sampling of synchronization
Value.
Step 2：According to actual needs, main index of correlation is selected.
Because electric power factory equipment is different, the index classification being related to is different, and index quantity is also uneven, real work
Middle analysis indexes forecasting problem, concrete equipment to make a concrete analysis of, and only in this way could more accurately explore the operation rule of equipment
Rule, and then finding parameter makes equipment reach optimal operational condition, reduce loss increases the service life of equipment.
For certain equipment, because index quantity reaches tens even up to a hundred, wherein existing substantially incoherent, only
Vertical index, is excluded according to an expert view, so as to improve the operational efficiency of model；Other disturbing factors are removed, is increased
The accuracy of model prediction result.Several indexs under certain equipment are finally given, training data is constituted.
Hypothesis finally gives N number of index under certain equipment, M bars record, then the m article of record is considered as N-dimensional
Vector, is represented by：Xm=[xm1,xm2,...xmN], then the training data should be the matrix form of M*N.Concrete form is such as
Under：
Step 3：Normalized.
Data normalization process is a requisite step.Because each index property on equipment is different, if any pressure
Power, temperature, electric current etc., their unit is different, and weights difference may be very big, therefore, for model impact just
Meeting difference is very big, visually says, their influence power " inconsistent, unfairness ".To solve this problem, the present invention is using as follows
Method for normalizing by certain refer to target value x normalize.
xnew=(2x-2xmax)/(xmax-xmin)+1 (2)
Wherein, xnewFor the value after x normalization, xmaxFor the maximum of the index, xminFor the minimum of a value of the index.To institute
There is index according to said method to normalize, the training data after being normalized.
Step 4：Validation-cross method processes training data.
In order to fully excavate the rule of training data, simulated metrics between equipment operation, the present invention adopts validation-cross
Method, by the equal portions of training data 4, takes out every time a as inspection, and remaining 3 parts are used for training.
Step 5：The data divided to validation-cross method using generalized regression nerve networks (GRNN) are trained, and find optimal
Parameter.
Artificial neural network is a kind of imitation animal nerve network behavior feature, carries out the number of distributed parallel information processing
Learn model.Generalized regression nerve networks (GRNN) are one kind therein, are typically used as the regression algorithm model of large-scale data amount.
The present invention simulates inherent law of each index of power equipment when equipment normally runs using the method, next so as to reach prediction
Carve the purpose of correlation.
The structural representation of generalized regression nerve networks (GRNN) can represent the form made in Fig. 3, including input layer, mould
Formula layer, summation layer and four layers of neuron of output layer.Map network is input into X=[x1,x2,...xm]T, wherein being output as Y=[y1,
y2,...yT]T。
The neuron number of input layer is equal to the dimension m of input layer in learning sample, and each neuron is that simple distribution is single
Unit, directly passes to hidden layer by input variable.
The neuron number of mode layer is equal to the number n of learning sample, the different sample of each neuron correspondence, in mode layer
The transmission function of neuron i is
Wherein, X is network inputs amount, and Xi is the corresponding learning samples of neuron i, and σ is smoothing parameter, that is to say, that neuron i
The exponential form of the Euclidean distance square being output as between the corresponding sample Xi of input variable X
D2=(X-Xi)T(X-Xi) (4)
Summation layer includes two class neurons, and output of the one of which neuron to all mode layer neurons carries out counting asking
With each neuron of mode layer is 1 with the connection weight of the neuron, and transmission function is
Output of other neurons to all mode layer neurons is weighted summation, in mode layer i-th neuron with
Connection weight in summation layer between j-th summation neuron is i-th output sample YiIn j-th element yij, summation god
Jing unit transmission function be
Neuron number in output layer is equal to the dimension L of output vector in learning sample, and each neuron is by layer of suing for peace
Output is divided by, i.e.,
yi=SNj/SDJ=1,2 ..., L (7)
Training data after former step process is carried out into generalized regression nerve networks mode process, is obtained based on training number
According to the optimal parameter model that error is minimum.
Fig. 2 is the flow chart of moving model of the present invention, is mainly included the following steps that：
Step 1：Obtain real time data.
Real time data when equipment runs is obtained from Power Plant DCS System lane database, and does identical with training pattern step 1
Pretreatment operation, ensure data accuracy.
Step 2：Select corresponding achievement data.
The achievement data that selection authorizes out through expert or staff, but do not include purpose prediction index,
It is remaining correspondence pointer type, number should be consistent with the pointer type of training data, number.
Step 3：Normalized.
It is identical with training pattern normalization processing method.
Step 4：Using train come optimal parameter model be predicted.
The data that above step process is obtained, are input to optimal parameter model, obtain result.
Step 5：Renormalization, is predicted the outcome.
By the result renormalization of step 4, obtain it is final predict the outcome, for staff analyzing and processing relevant device letter
Breath and the further work of arrangement.
The present invention adopts following renormalization method by the renormalization that predicts the outcome of certain index.
Wherein, xpFor a predicted value of the index, xmaxThe maximum of the index obtained for modeling procedure 3, xminFor
The minimum of a value of the index that modeling procedure 3 is obtained, xrFor the predicted value of the index after renormalization.To all purposes prediction index
According to said method renormalization, is predicted the outcome.
Embodiment
With the pressure fan of northern certain power plant #1 units as monitoring object, pressure fan is the important auxiliary of power plant to the present embodiment
Equipment, its complex structure, influence factor is more, it is difficult to sets up accurate complicated mathematics mechanism model, and easily multiple malfunctions, meets this
The characteristics of inventing targeted nonlinear multivariable system.Elaborating by the present embodiment, further illustrates the reality of the present invention
Apply process.
First, the embodiment of the present invention is as follows to certain power plant's pressure fan index prediction modeling procedure：
Step 1：Blower fan achievement data is fetched and delivered from Power Plant DCS System lane database, with pressure fan normal operating condition
Historical data is used as training data.The index related to the operation of the pressure fan of the power plant has 45, including power (MW), electric current
(A), air blower inlet temperature (DEG C), axle 1X is to vibration (mm/s) etc..This example training data is from Power Plant DCS System data
The historical data from April, 2016 No. 4 power plant 1# machine pressure fan of No. 25 in June, 2016 that Ku Li gets, peek at intervals of
10min。
Step 2：Index of correlation data are selected according to the practical problem of pressure fan, and does certain pretreatment work.
According to being actually needed, index hot blast air quantity, outlet temperature are particularly important, need prediction in time, grasp related data
The operation conditions of assessment equipment.Through evaluation, NF is excluded, select 15 leading indicators altogether, including power, electric current, entered
Mouth movable vane position feedback etc., as basis for forecasting index.Exclude because power plant system reason cause comprising dirty data, vacancy value etc.
Data record；The foundation achievement data of all current data records and next common group of the purpose prediction index data for recording
Record into final training data, such as the anabolic process of the data of training data first record, (current value 1, entrance movable vane position is anti-
Feedback value 1 ... hot blast airflow value 2, Outlet Temperature value 2), therein 1,2 represent from DCS system database take out when data note
The original line number of record.Partial data is presented below：
Step 3：Normalized.
To exclude impact of the index weight value to forecast model, the training data of pressure fan is normalized.Partial data shows
It is as follows：
| Electric current | Feed back entrance movable vane position | …… | Hot blast air quantity | Outlet temperature |
| -0.54349433 | -0.23648701 | …… | -0.78145672 | -0.42048523 |
| -0.56531589 | -0.23721738 | …… | -0.78236392 | -0.42503240 |
| -0.54358623 | -0.22880003 | …… | -0.69105943 | -0.42334336 |
| -0.41899617 | -0.21598148 | …… | -0.72052971 | -0.42396483 |
| -0.45204450 | -0.20839068 | …… | -0.73119393 | -0.42032569 |
Step 4：Validation-cross method processes the training data of pressure fan.
For the inherent law for fully excavating pressure fan training data, the training number of pressure fan is processed using validation-cross method
According to.
Step 5：With generalized regression nerve networks (GRNN) as forecast model, the minimum parameter of error identifying is optimal ginseng
Number.
The training data training general regression neural network that step 4 is obtained, obtains optimum prediction model, Jing Guofen
Analysis, error is in pressure fan actual motion allowed band.
2nd, the present invention is as follows the step of pressure fan example moving model：
Step 1：All indication informations of pressure fan are taken out from Power Plant DCS System lane database.
The achievement data of pressure fan is taken out from Power Plant DCS System lane database, the time is from June, 2016 No. 5 to 2016
June No. 25 in year, as prediction data.Peek is at intervals of 10min.
Step 2：Select corresponding pressure fan achievement data.
Training data record most next two columns do not include purpose prediction index, should be (current value, entrance movable vane position feedback
Value ...), it is not (current value, entrance movable vane position value of feedback ... hot blast airflow value, Outlet Temperature value)；Process comprising vacancy value etc.
Data record, impact of the removal system noise to predicting the outcome.Partial results are presented below：
| Electric current | Feed back entrance movable vane position | …… |
| 66.27044678 | 59.92971802 | …… |
| 69.51446533 | 63.63220215 | …… |
| 66.25900269 | 61.45934677 | …… |
| 66.26027679 | 60.3863678 | …… |
| 68.98269653 | 60.91820145 | …… |
Step 3：Normalized.
The normalization operation consistent with modeling process is done to pressure fan data.Partial data is presented below：
Step 4：Pressure fan data are predicted with the optimal parameter model for training.
The pressure fan data input optimal parameter model that step 3 is obtained, obtains output result.
Step 5：The output result renormalization that step 4 is obtained.
The output result of renormalization step 4, obtains final predicting the outcome.This result be electric current (A), entrance movable vane position
Feedback (%), power (MW), air blower inlet temperature (DEG C), the hot blast wind that axle 1X is predicted to achievement datas such as vibrations (mm/s)
Amount (t/h), Outlet Temperatures (DEG C).Partial data is presented below：
| Hot blast air quantity | Outlet temperature |
| 137.25958853 | 25.87092028 |
| 140.41691601 | 26.06566395 |
| 138.91559737 | 26.82168667 |
| 140.26331837 | 26.03597391 |
| 141.59082622 | 25.01163275 |
Corresponding True Data is as follows：
| Hot blast air quantity | Outlet temperature |
| 136.25951234 | 25.83432553 |
| 141.41696754 | 26.03434325 |
| 138.91454333 | 27.65767675 |
| 140.56334555 | 26.23565775 |
| 142.00234322 | 24.73225443 |
The mean square error of pressure fan hot blast air quantity prediction is that 3.1986, i.e. AME are interior in [- 1.788,1.788],
In the area requirement that power plant allows, pressure fan outlet temperature error is also in allowed band, it was demonstrated that method is feasible for error.
Again by taking hot blast air quantity as an example, first 400 predict the outcome and real data comparison diagram such as Fig. 4, and rhombus represents true
Value, circle represents predicted value；Actual value and predicted value are basically identical, when illustrating that the model meets pressure fan operation between each index
Rule, method is feasible.
Power plant staff with the predicted value of hot blast air quantity and outlet temperature as foundation, with reference to other professional knowledges, method,
Running status, the health degree of assessment pressure fan, adjusts working condition, makes repair schedule.
Although for illustrative purposes, it has been described that the illustrative embodiments of the present invention, those skilled in the art
Member will be understood that, in the case of the scope and spirit without departing from the invention disclosed in claims, can be in form and details
On carry out various modifications, addition and replace etc. change, and it is all these change should all belong to claims of the present invention
Protection domain, and each department of claimed product and each step in method, can be in any combination
Form is combined.Pair therefore, disclosed in this invention the description of embodiment be not intended to limit the scope of the present invention,
But for describing the present invention.Correspondingly, the scope of the present invention is not limited by embodiment of above, but by claim or
Its equivalent is defined.
Claims (8)
1. a kind of power equipment multi objective Forecasting Methodology, it is characterised in that in turn include the following steps：
(1) model is set up：
Step 1.1：Training data is obtained, in Power Plant DCS System lane database, is found all under target device and equipment
Indication information, is spaced according to certain peek, and selected equipment historical data under normal operating conditions is pre-processed；
Step 1.2：According to actual needs, main index of correlation is selected；
Step 1.3：Main index of correlation to selecting is normalized, and obtains training data；
Step 1.4：Training data is processed using validation-cross method；
Step 1.5：Data to dividing after the process of validation-cross method are trained, and find optimal parameter, obtain optimal parameter
Model；
(2) moving model, obtains power equipment multi objective and predicts the outcome：
Step 2.1：Obtain real time data, from Power Plant DCS System lane database obtain equipment run when real time data, and do with
The consistent pretreatment operation of step 1；
Step 2.2：According to actual needs, corresponding achievement data is selected；
Step 2.3：To selecting corresponding achievement data to carry out the normalized consistent with step 1.3, the finger after being processed
Mark data；
Step 2.4：Achievement data after the process that step 2.3 is obtained, is input to optimal parameter model, obtains result；
Step 2.5：The result obtained to step 2.4 carries out renormalization process, is predicted the outcome.
2. the method for claim 1, it is characterised in that：Also include that the data to choosing are entered in step 1.1 or step 2.1
Row pretreatment, removing achievement data causes the record of some values disappearances more than the record and system reason of index bound.
3. method as claimed in claim 1 or 2, it is characterised in that step 1.2 is concretely comprised the following steps：
Hypothesis finally gives N number of index under certain equipment, and M bars are recorded, then the m article of record regards the vector of a N-dimensional as, are represented
For：Xm=[xm1,xm2,...xmN], then training data is the matrix form of M*N, and concrete form is as follows：
4. the method for claim 1, it is characterised in that normalized is specifically pressed described in step 1.3 or step 2.3
Carry out according to equation below：
xnew=(2x-2xmax)/(xmax-xmin)+1
Wherein, xnewFor the value after x normalization, xmaxFor the maximum of main index of correlation, xminFor main index of correlation
Minimum of a value.
5. the method for claim 1, it is characterised in that step 1.4 is specially：Using validation-cross method, will train
The equal portions of data 4, take out every time a as inspection, and remaining 3 parts are used for training.
6. the method for claim 1, it is characterised in that step 1.5 is specially：Using generalized regression nerve networks GRNN
Mode is trained to the data divided after the process of validation-cross method, obtains based on the minimum optimal parameter of training data error
Model, specially：
Assume map network input X=[x1,x2,...xm]T, wherein being output as Y=[y1,y2,...yT]T；
The neuron number of input layer is equal to the dimension m of input layer in learning sample, and each neuron is simple distribution unit, directly
Connect and input variable is passed to into hidden layer；
The neuron number of mode layer is equal to the number n of learning sample, the different sample of each neuron correspondence, nerve in mode layer
The transmission function of first i is：
Pi=exp is [(X-Xi)T(X-Xi)/2σ2] i=1,2 ..., n；Wherein, X is network inputs amount, and Xi is neuron i correspondences
Learning sample, σ is smoothing parameter, and neuron i is output as the Euclidean distance between the corresponding sample Xi of input variable X
Square exponential form：
D2=(X-Xi)T(X-Xi)；
Summation layer includes two class neurons, and output of the one of which neuron to all mode layer neurons carries out the summation that counts,
The each neuron of mode layer is 1 with the connection weight of the neuron, and transmission function is：
Output of other neurons to all mode layer neurons is weighted summation, i-th neuron and summation in mode layer
Connection weight in layer between j-th summation neuron is i-th output sample YiIn j-th element yij, neuron of suing for peace
Transmission function be
Neuron number in output layer be equal to learning sample in output vector dimension L, each neuron by sue for peace layer output
It is divided by, i.e.,
yi=SNj/SDJ=1,2 ..., L.
7. the method for claim 1, it is characterised in that：Step 2.2 is specially selected through expert or staff
The achievement data for authorizing out, not including purpose prediction index, remaining correspondence pointer type, number should be with training numbers
According to pointer type, number it is consistent.
8. the method as described in claim 1-7, it is characterised in that the concrete of step 2.5 carries out anti-normalizing using equation below
At change
Reason：
Wherein, xpFor a predicted value of correspondence index, xmaxThe maximum of the main index of correlation obtained for step 1.3, xmin
The minimum of a value of the main index of correlation obtained for step 1.3, xrFor the predicted value of correspondence index after renormalization.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201611263834.7A CN106682781A (en) | 2016-12-30 | 2016-12-30 | Power equipment multi-index prediction method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201611263834.7A CN106682781A (en) | 2016-12-30 | 2016-12-30 | Power equipment multi-index prediction method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN106682781A true CN106682781A (en) | 2017-05-17 |
Family
ID=58849509
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201611263834.7A Pending CN106682781A (en) | 2016-12-30 | 2016-12-30 | Power equipment multi-index prediction method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106682781A (en) |
Cited By (16)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108846489A (en) * | 2018-07-02 | 2018-11-20 | 天泽信息产业股份有限公司 | A kind of large scale equipment health Evaluation method based on SVM |
| CN109344976A (en) * | 2018-08-24 | 2019-02-15 | 华能国际电力股份有限公司海门电厂 | A kind of electrical system operating status intelligent analysis method based on Keras |
| CN109492826A (en) * | 2018-12-06 | 2019-03-19 | 远光软件股份有限公司 | A kind of information system operating status Risk Forecast Method based on machine learning |
| CN109582657A (en) * | 2018-11-28 | 2019-04-05 | 国家电网有限公司 | Capacitor faults prediction technique based on machine learning and big data analysis |
| CN109726463A (en) * | 2018-12-25 | 2019-05-07 | 中铁隧道局集团有限公司 | A kind of shield TBM fault early warning method based on SVM algorithm |
| CN110009384A (en) * | 2019-01-07 | 2019-07-12 | 阿里巴巴集团控股有限公司 | Predict the method and device of operational indicator |
| CN108509645B (en) * | 2018-04-13 | 2020-03-24 | 华润电力风能(威海)有限公司 | Equipment early warning method |
| CN111738536A (en) * | 2020-08-28 | 2020-10-02 | 北京每日优鲜电子商务有限公司 | Device operation method, device, electronic device and computer readable medium |
| CN111784133A (en) * | 2020-06-22 | 2020-10-16 | 上海化学工业区中法水务发展有限公司 | Method and equipment for making sewage plant energy consumption performance index |
| CN112328847A (en) * | 2019-12-26 | 2021-02-05 | 国家电网有限公司 | Transformer overload visualization method and system based on big data |
| CN112651512A (en) * | 2020-12-10 | 2021-04-13 | 北京北大千方科技有限公司 | Service reasonableness calculation method and device |
| CN112862180A (en) * | 2021-02-04 | 2021-05-28 | 华电国际电力股份有限公司技术服务分公司 | Denitration system inlet NOx concentration prediction method |
| CN113656238A (en) * | 2021-08-17 | 2021-11-16 | 北京京航计算通讯研究所 | AI industrial application capability test method and system of intelligent terminal |
| CN113837227A (en) * | 2021-08-26 | 2021-12-24 | 北京智芯微电子科技有限公司 | Load prediction method, device, chip, electronic device and storage medium |
| CN114757380A (en) * | 2022-04-29 | 2022-07-15 | 西安热工研究院有限公司 | Thermal power plant fault early warning system and method, electronic equipment and storage medium |
| CN115169650A (en) * | 2022-06-20 | 2022-10-11 | 四川观想科技股份有限公司 | Equipment health prediction method for big data analysis |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104915747A (en) * | 2015-02-03 | 2015-09-16 | 远景能源（江苏）有限公司 | Electricity generation performance evaluation method of generator set and equipment thereof |
| CN105302848A (en) * | 2014-10-11 | 2016-02-03 | 山东鲁能软件技术有限公司 | Evaluation value calibration method of equipment intelligent early warning system |
| CN105550943A (en) * | 2016-01-18 | 2016-05-04 | 重庆大学 | Method for identifying abnormity of state parameters of wind turbine generator based on fuzzy comprehensive evaluation |
| CN105930901A (en) * | 2016-07-18 | 2016-09-07 | 河海大学 | RBPNN-based transformer fault diagnosis method |
-
2016
- 2016-12-30 CN CN201611263834.7A patent/CN106682781A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105302848A (en) * | 2014-10-11 | 2016-02-03 | 山东鲁能软件技术有限公司 | Evaluation value calibration method of equipment intelligent early warning system |
| CN104915747A (en) * | 2015-02-03 | 2015-09-16 | 远景能源（江苏）有限公司 | Electricity generation performance evaluation method of generator set and equipment thereof |
| CN105550943A (en) * | 2016-01-18 | 2016-05-04 | 重庆大学 | Method for identifying abnormity of state parameters of wind turbine generator based on fuzzy comprehensive evaluation |
| CN105930901A (en) * | 2016-07-18 | 2016-09-07 | 河海大学 | RBPNN-based transformer fault diagnosis method |
Non-Patent Citations (1)
| Title |
|---|
| 张宇航 等: ""基于广义神经网络与模糊聚类的变压器故障诊断"", 《高压电器》 * |
Cited By (22)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108509645B (en) * | 2018-04-13 | 2020-03-24 | 华润电力风能(威海)有限公司 | Equipment early warning method |
| CN108846489A (en) * | 2018-07-02 | 2018-11-20 | 天泽信息产业股份有限公司 | A kind of large scale equipment health Evaluation method based on SVM |
| CN109344976A (en) * | 2018-08-24 | 2019-02-15 | 华能国际电力股份有限公司海门电厂 | A kind of electrical system operating status intelligent analysis method based on Keras |
| CN109582657A (en) * | 2018-11-28 | 2019-04-05 | 国家电网有限公司 | Capacitor faults prediction technique based on machine learning and big data analysis |
| CN109492826A (en) * | 2018-12-06 | 2019-03-19 | 远光软件股份有限公司 | A kind of information system operating status Risk Forecast Method based on machine learning |
| CN109726463A (en) * | 2018-12-25 | 2019-05-07 | 中铁隧道局集团有限公司 | A kind of shield TBM fault early warning method based on SVM algorithm |
| CN110009384A (en) * | 2019-01-07 | 2019-07-12 | 阿里巴巴集团控股有限公司 | Predict the method and device of operational indicator |
| CN112328847A (en) * | 2019-12-26 | 2021-02-05 | 国家电网有限公司 | Transformer overload visualization method and system based on big data |
| CN111784133B (en) * | 2020-06-22 | 2023-01-24 | 上海化学工业区中法水务发展有限公司 | Method and equipment for making sewage plant energy consumption performance index |
| CN111784133A (en) * | 2020-06-22 | 2020-10-16 | 上海化学工业区中法水务发展有限公司 | Method and equipment for making sewage plant energy consumption performance index |
| CN111738536A (en) * | 2020-08-28 | 2020-10-02 | 北京每日优鲜电子商务有限公司 | Device operation method, device, electronic device and computer readable medium |
| CN112651512A (en) * | 2020-12-10 | 2021-04-13 | 北京北大千方科技有限公司 | Service reasonableness calculation method and device |
| CN112862180A (en) * | 2021-02-04 | 2021-05-28 | 华电国际电力股份有限公司技术服务分公司 | Denitration system inlet NOx concentration prediction method |
| CN113656238A (en) * | 2021-08-17 | 2021-11-16 | 北京京航计算通讯研究所 | AI industrial application capability test method and system of intelligent terminal |
| CN113656238B (en) * | 2021-08-17 | 2023-07-04 | 北京京航计算通讯研究所 | AI industrial application capability test method and system of intelligent terminal |
| CN113837227A (en) * | 2021-08-26 | 2021-12-24 | 北京智芯微电子科技有限公司 | Load prediction method, device, chip, electronic device and storage medium |
| CN113837227B (en) * | 2021-08-26 | 2024-02-02 | 北京智芯微电子科技有限公司 | Load prediction method, load prediction device, chip, electronic equipment and storage medium |
| CN114757380A (en) * | 2022-04-29 | 2022-07-15 | 西安热工研究院有限公司 | Thermal power plant fault early warning system and method, electronic equipment and storage medium |
| WO2023207190A1 (en) * | 2022-04-29 | 2023-11-02 | 西安热工研究院有限公司 | Fault early warning system and method for fossil fuel power plant, electronic device, and storage medium |
| CN114757380B (en) * | 2022-04-29 | 2024-02-20 | 西安热工研究院有限公司 | Fault early warning system and method for thermal power plant, electronic equipment and storage medium |
| CN115169650A (en) * | 2022-06-20 | 2022-10-11 | 四川观想科技股份有限公司 | Equipment health prediction method for big data analysis |
| CN115169650B (en) * | 2022-06-20 | 2022-11-22 | 四川观想科技股份有限公司 | Equipment health prediction method for big data analysis |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106682781A (en) | Power equipment multi-index prediction method | |
| CN105096053B (en) | A kind of health control decision-making technique suitable for complicated technology system | |
| CN106407589B (en) | Fan state evaluation and prediction method and system | |
| CN105095963A (en) | Method for accurately diagnosing and predicting fault of wind tunnel equipment | |
| CN107430715A (en) | Cascade identification in building automation | |
| CN106199174A (en) | Extruder energy consumption predicting abnormality method based on transfer learning | |
| Jana et al. | Novel type-2 fuzzy logic approach for inference of corrosion failure likelihood of oil and gas pipeline industry | |
| CN114118673A (en) | Workshop intelligent fault diagnosis early warning method based on digital twin technology | |
| CN108921230A (en) | Method for diagnosing faults based on class mean value core pivot element analysis and BP neural network | |
| Gadhavi et al. | Student final grade prediction based on linear regression | |
| Garg et al. | A two-phase approach for reliability and maintainability analysis of an industrial system | |
| CN107291830A (en) | A kind of creation method of equipment knowledge base | |
| CN117389236B (en) | Propylene oxide production process optimization method and system | |
| CN117196159A (en) | Intelligent water service partition metering system based on Internet big data analysis | |
| CN117393076B (en) | Intelligent monitoring method and system for heat-resistant epoxy resin production process | |
| CN115438726A (en) | Device life and fault type prediction method and system based on digital twin technology | |
| Eltyshev et al. | Intelligent decision support in the electrical equipment diagnostics | |
| Tripathy et al. | Explaining Anomalies in Industrial Multivariate Time-series Data with the help of eXplainable AI | |
| Dong et al. | A combining condition prediction model and its application in power plant | |
| CN115983714A (en) | Static security assessment method and system for edge graph neural network power system | |
| Elbazi et al. | Digital twin-enabled monitoring of mining haul trucks with expert system integration: A case study in an experimental open-pit mine | |
| CN115378136A (en) | Intelligent early warning method and early warning device for safety risk of secondary equipment of power grid | |
| Zhang et al. | Decision-supporting models for human-reliability based safety promotion in offshore Liquid Natural Gas terminal | |
| Alzeraif et al. | Predicting Maintenance Labor Productivity in Electricity Industry using Machine Learning: A Case Study and Evaluation | |
| Castro-Puma et al. | Automatic learning algorithm for troubleshooting in hydraulic machinery |
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 | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20170517 |
|
| RJ01 | Rejection of invention patent application after publication |