CN108051035A - The pipe network model recognition methods of neural network model based on gating cycle unit - Google Patents
The pipe network model recognition methods of neural network model based on gating cycle unit Download PDFInfo
- Publication number
- CN108051035A CN108051035A CN201710998436.8A CN201710998436A CN108051035A CN 108051035 A CN108051035 A CN 108051035A CN 201710998436 A CN201710998436 A CN 201710998436A CN 108051035 A CN108051035 A CN 108051035A
- Authority
- CN
- China
- Prior art keywords
- data
- network model
- training
- neural network
- water supply
- 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.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01F—MEASURING VOLUME, VOLUME FLOW, MASS FLOW OR LIQUID LEVEL; METERING BY VOLUME
- G01F1/00—Measuring the volume flow or mass flow of fluid or fluent solid material wherein the fluid passes through a meter in a continuous flow
-
- 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
- G06N3/04—Architecture, e.g. interconnection topology
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Fluid Mechanics (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The present invention provides a kind of neural network model based on gating cycle unit and its training method and applications.Wherein, the neural network model based on gating cycle unit, including：Multiple gating cycle elementary layers are configured to receive water supply network in traffic characteristic in different time periods；First full articulamentum, composition input layer in parallel with the multiple gating cycle elementary layer, the first full articulamentum are configured to receive the Meteorological Characteristics of water supply network；The merging layer being connected with the input layer with the pattern that tensor is connected；With the merging layer with the second full articulamentum that the pattern that tensor is connected is connected to the full articulamentums of M, M is the integer more than or equal to 2；And the output layer being connected with the full articulamentums of M with the pattern that tensor is connected, it is configured to predicted flow rate of the output water supply network in subsequent time.
Description
Technical field
The present invention relates to field of neural networks, and in particular to a kind of neural network model based on gating cycle unit and its
Training method and application.
Background technology
Under the influence of the factors such as pipeline aging, limited, the Supervision backwardness of Technical investment, the leakage of China's public supply mains
It is more universal to damage phenomenon.Water supply network leakage loss not only has brought tremendous economic losses to Running-water Company, also results in the energy
With the waste of resource.Major part Running-water Company of China lacks scientific and effective management method, the base of pipe network for water supply network
Plinth data are also not perfect, these problems restrict the efficiency of management and service level of Running-water Company always.Therefore, test tube is treated
Net carries out leakage loss identification, and the science decision of Added Management person finds and safeguard in time pipe network model region, has important economy
Meaning and realistic meaning.
The model for being commonly used in identification pipe network model is broadly divided into two major classes：One kind is the physics based on laboratory test
Mechanism study model；One kind is the data mining model based on computer simulation, including statistical model, probabilistic model, engineering
Practise model etc., it usually needs substantial amounts of historical data.With the arriving of cloud era, big data has attracted more and more concerns,
Good achievement is achieved in numerous areas, therefore the second class model also becomes current research hotspot, especially machine
Learning model.
China's public supply mains data volume is big, and basic data is of low quality, causes the factor of pipe network model numerous and closes
System is complicated.Existing pipe network model identification model is mostly based on normal discharge or pressure of some classical traditional algorithms to pipe network
Power is predicted, such as multiple linear regression, exponential smoothing, backpropagation (Back Propagation, abbreviation BP) nerve
Network etc., then by comparing predicted value and actual monitoring value identification leakage accident.These traditional Forecasting Methodologies are big in processing
Ability in terms of amount complex nonlinear data is limited, it is impossible to which the preferable hidden feature excavated between abstract data, there are pre-
The problem of precision is not high is surveyed, this is but also the accuracy of leakage accident identification has no way of ensureing.
The content of the invention
In order to save water resource, the leak rate of water supply network is reduced, reduces economic loss, meets the reason of sustainable development
It reads, it would be desirable to the hidden feature of water supply network data is deeply excavated using the prior art, it is good to establish precision height, stability
Model identifies pipe network model region.In view of the deficiencies of the prior art, it is the present invention is intended to provide a kind of new based on gating cycle
The neural network model and its training method of unit and application are different from the new god of traditional neural network model by structure
Through network, mass data can be handled, excavate non-linear relation and Integrated Evaluation Model effect, be applied to water supply network
Leakage accident identifies, can improve the accuracy rate of leakage loss identification, so that daily management person can have found pipe network model region in time,
Economic loss is reduced, saves water resource, auxiliary Running-water Company makes scientific and reasonable decision-making.
One aspect of the present invention provides a kind of neural network model based on gating cycle unit, for identifying water supply network
Leakage accident, including：
Multiple gating cycle unit (Gated Recurrent Unit, abbreviation GRU) layers are configured to receive feed pipe
Net is in traffic characteristic in different time periods；
First full articulamentum, composition input layer in parallel with the multiple gating cycle elementary layer, the described first full connection
Layer is configured to receive the Meteorological Characteristics of water supply network；
The merging layer being connected with the input layer with the pattern that tensor is connected；
With the merging layer with the second full articulamentum that the pattern that tensor connect is connected to the full articulamentums of M, M for more than
Or the integer equal to 2；And
The output layer being connected with the full articulamentums of M with the pattern that tensor is connected is configured to output water supply network
In the predicted flow rate of subsequent time.
Gating cycle neutral net (Gated Recurrent Unit Network, abbreviation GRUN) provided by the present invention
It is the improvement to existing Recognition with Recurrent Neural Network.The characteristics of this network, is the implicit node using memory module replacement commonly,
Ensure that gradient will not disappear or explode after transferring across many time steps, so as to overcome conventional recycle neural metwork training
In the difficulty that runs into, the memory module that the present invention uses is gating cycle unit (GRU).Fig. 1 shows the reality of the present invention
Apply the topology diagram of the neural network model based on gating cycle unit in mode.
Another aspect of the present invention provides a kind of training method of above-mentioned neural network model, including,
Build the neural network model；
Historical traffic feature in different time periods and history Meteorological Characteristics are randomly divided into training set in proportion and verification collects；
Historical traffic feature in different time periods in the training set is input in the multiple gating cycle elementary layer,
And the history Meteorological Characteristics in the training set are input in the described first full articulamentum, it is defeated to obtain the neural network model
The training flow gone out；
Historical traffic feature in different time periods is concentrated to be input in the multiple gating cycle elementary layer the verification,
And the history Meteorological Characteristics for concentrating the verification are input in the described first full articulamentum, and it is defeated to obtain the neural network model
The verification flow gone out；
It is tested based on the training set and training traffic generating training curve, and based on the verification collection and verification traffic generating
Curve is demonstrate,proved, when the mean square error value stabilization of the training curve and verification curve is in steady state value, completes the neutral net mould
The training of type.
The training method has taken into full account the generalization ability of the randomness of gradient descent direction and model in the training process,
Improve the convergence rate and training effectiveness of model, the effective information between abundant mining data.
In the preferred embodiment of the present invention, it is described in different time periods that above-mentioned training method further includes acquisition
Historical traffic feature, including：
Obtain the historical traffic data of water supply network to be measured；And
Processing is extracted to the historical traffic data and normalized obtains the historical traffic feature.
According to the present invention, the historical traffic data can be obtained from the independent measure region (district of Running-water Company
Metered areas, referred to as DMA) water supply network historical traffic data.
In the preferred embodiment of the present invention, extracting processing to the historical traffic data includes,
In the historical traffic data of water supply network, respectively from the first time period, second time period, in the 3rd period with etc. between
The data on flows of number is specified every extraction, wherein,
First time period is chosen before moment t to be predicted is closed on,
Second time period is chosen before the initial time of the first time period,
Chose for the 3rd period before the initial time of the second time period.
According to the present invention, the tendency of flow-time sequence is considered, extraction closes on the flow number of moment t moment to be predicted
According to the i.e. data on flows of first time period.Consider the periodicity of flow-time sequence, extraction is separated by one section with moment t to be predicted
The data on flows of time, it is contemplated that be daily divided into two periods, i.e. second time period and the data on flows of the 3rd period.
In the specific embodiment of the present invention, including first time period, second time period and the 3rd time
In the entire period of section, average sample is carried out with k times/h of sample frequency, then the data on flows included in the entire period includes
The data on flows at t-24kd, t-24kd+1 ... t-2, t-1 moment, wherein d represent number of days, when therefrom obtaining t-m to t-1
The value range of the data on flows at quarter, wherein m is the integer of 2-24, is preferably 5, the data on flows as first time period；It obtains
Take t-h1To t-h2The data on flows at moment, wherein h1And h2Value range be 24k-2 to 24k+20 integer, it is preferable that h1
For 24k+2, h2For 24k-2；Obtain t-h3To t-h4The data on flows at moment, wherein h3And h4Value range be 48k-2 to 48k
+ 20 integer, it is preferable that h3For 48k+2, h4For 48k-2.
Preferably, k is the integer of 1-12, is preferably 3-5.
In another preferred embodiment of the present invention, above-mentioned training method, which further includes, obtains the meteorological spy of the history
Sign, including：
Obtain the history meteorological data of water supply network to be measured；And
Processing is extracted to the history meteorological data and normalized obtains the history Meteorological Characteristics.
According to the present invention, the history meteorological data can be obtained from the history meteorological data of China Meteorological data network, including
Maximum temperature, minimum temperature, relative humidity, precipitation etc..
The present invention a preferred embodiment in, to the history meteorological data extract processing including pair
The history meteorological data and the historical traffic data carry out Pearson correlation analysis, choose related coefficient and are more than or wait
In the history meteorological data of specified value V, the history meteorological data after extraction process is obtained.
In the preferred embodiment of the present invention, the specified value V is 0.80-0.99.
According to the present invention, the normalized includes, and the numerical value of initial data is transformed into the range of [0,1], normalizing
Shown in the formula such as formula (1) for changing processing,
In formula (1), y represents the data after normalized, and x represents the initial data of input, xmaxAnd xminIt represents respectively
The maximum and minimum value of input data.
According to the present invention, activation primitive selection tanh, ReLU or Linear of each network layer.
Further aspect of the present invention provides a kind of recognition methods of the leakage accident of water supply network, including,
Feature extraction step in the stipulated time T that leakage accident does not occur, obtains the different time of water supply network to be measured
Section reference flow feature and with reference to Meteorological Characteristics；
The reference flow feature in different time periods is input to the neural network model by predicted flow rate obtaining step
Multiple gating cycle elementary layers in, and by the first full connection that the neural network model is input to reference to Meteorological Characteristics
In layer, the predicted flow rate of the neural network model output is obtained；
Reference data set determines step, obtains the measured discharge in the stipulated time T, and every group in mutually in the same time
Predicted flow rate forms a time series vector with measured discharge, by all time series vector structures in the stipulated time T
Into matrix be used as with reference to data set；
Identification of accidental events step calculates each time sequence that newly-increased time series vector is concentrated compared with the reference data
The COS distance of column vector, the number n of time series vector of the statistics COS distance less than distance threshold D, when n is less than or equal to
During amount threshold N, determine that the leakage accident has occurred.
According to the present invention, the newly-increased time series vector is that the period beyond the stipulated time T is according to described
Feature extraction step, predicted flow rate obtaining step, reference data value determine the time series vector that step obtains.
According to the present invention, concentrated by formula (2) calculating newly-increased time series vector compared with the reference data every
The COS distance of a time series vector.
In formula (2), i is newly-increased time series vector, and j is the time series vector that reference data is concentrated, and ‖ i ‖ and ‖ j ‖ are
Vector field homoemorphism.
According to the present invention, the reference flow feature in different time periods of water supply network to be measured and the acquisition with reference to Meteorological Characteristics
In mode and training method above, historical traffic feature is consistent with the acquisition modes of history Meteorological Characteristics.
According to the present invention, in above-mentioned recognition methods, set the time T as 2 days, i.e., 48 it is small when, set adopting for data
Integrate frequency as k times/h (i.e. obtaining k time series vector in a hour), then time series included in reference data set
The number of vector is 48k.
In another preferred embodiment of the present invention, the distance threshold D is to be directed to the reference data set
In calculate 0.1-0.5 times, preferably 0.4-0.5 times of the medians of all COS distances of all time series vectors.
Compared with existing public supply mains leakage loss recognition methods, the advantage of recognition methods of the invention is：
First, be different from traditional neural network model, the present invention multiple GRU layer with full articulamentum by parallel connection or
The mode of series connection connects, and forms a complicated deep neural network.GRUN has the ability of stronger processing nonlinear data,
It is stronger especially for the processing capacity of sequence data, the memory state to past data can be generated, and establishes different periods
Dependence between data.Secondly, the capability of fitting of GRUN is stronger, is easier to restrain in learning process, is not easy to be absorbed in office
The minimum state in portion.So can more accurately predict pipe network flow using GRUN, so as to which pipe network model be identified, improve
The accuracy rate of pipe network model identification.
Second, the present invention uses the rejecting outliers method based on COS distance after the accurate predicted flow rates of GRUN are used
Identify leakage accident, the COS distance judgement compared by analysis between time series vector (being made of predicted value and measured value) is
No generation leakage accident.On the one hand, the use of COS distance eliminates pipe network monitoring data fluctuations scope greatly to rejecting outliers
Caused by influence；On the other hand, it is remaining between multiple time series vectors compared to the absolute difference between analysis predicted value and measured value
The comparison of chordal distance avoids subjectivity when judging leakage accident, accuracy higher.
Description of the drawings
Fig. 1 is the topological structure of the neural network model based on gating cycle unit in an embodiment of the invention
Figure.
Fig. 2 is the flow of neural network model of the training based on gating cycle unit in an embodiment of the invention
Figure.
Fig. 3 is the figure for representing the mean square error of training curve and verification curve in the embodiment of the present invention 1.
Fig. 4 is the leakage accident knowledge that water supply network is carried out using the neural network model in an embodiment of the invention
Other flow chart.
Specific embodiment
To be better understood from and implementing the present invention, the present invention is explained in detail below in conjunction with the drawings and specific embodiments
It states.Although it should be appreciated that embodiments of the present invention are illustrated, but it is clear that the present invention is not limited to above-mentioned
Embodiment can carry out various modifications in the range of its purport is not departed from.
In the following embodiments, by the use of 2.7 softwares of Python as the development platform of model, and using Numpy and
Pandas storehouses read, store, analyze data, and the visualization of data is done using Matplotlib storehouses, is taken using Keras storehouses
Neural network model is built, substantially increases development efficiency.
Embodiment 1 trains neural network model
1) neural network model is built according to Fig. 1
Input layer (the section of 3 GRU layers of setting and the first full articulamentum is formed by 3 GRU layers and the first full articulamentum parallel connection
Points are respectively that 48,32,32,8, GRU layers of activation primitive is tanh and ReLU, and the activation primitive of the first full articulamentum is
ReLU)；
The pattern connected with tensor makes 3 GRU layer and the first full articulamentum be connected respectively with merging (Merge) layer；
Merging layer and the second full articulamentum, the 3rd full articulamentum, the 4th full articulamentum, the 5th are made with the pattern that tensor is connected
Full articulamentum, the 6th full articulamentum, the 7th full articulamentum connection (number of nodes of setting the second full articulamentum to the 7th full articulamentum
For 64,32,16,8,4,2, activation primitive ReLU)；
The pattern connected using tensor make the 7th full articulamentum be connected with output layer (activation primitive of output layer as Linear,
Learning rate is 0.02, batch_size 60).
2) generate training set and verification collects
2-1) to the CZ cities DMA water supply networks of Running-water Company from 2 1st, 2016 to 31 days January in 2017
Data on flows carries out average sample by 4 times/h of frequency acquisition, that is, the data on flows every 15min is obtained, to the data of acquisition
It is pre-processed：Cause the data on flows of accident including cleaning non-natural factor (third party, artificial)；Typing mistake is corrected, clearly
Wash apparent abnormal data etc..Then, first time period, second time period, the historical traffic data of the 3rd period are therefrom extracted,
And be normalized according to formula (1), historical traffic feature in different time periods is obtained, wherein,
The data on flows of first time period includes t-75min, t-60min, t-45min, t-30min, t-15min moment
Data on flows；
The data on flows of second time period includes t-24.5h, t-24.25h, t-24h, t-23.75h to the t-23.5h moment
Data on flows；
The data on flows of 3rd period includes t-48.5h, t-48.25h, t-48h, t-47.75h to the t-47.5h moment
Data on flows.
2-2) from Chinese meteorological data net obtain history meteorological data, including maximum temperature, minimum temperature, relative humidity and
Precipitation pre-processes the data of acquisition：Cause the flow number of accident including cleaning non-natural factor (third party, artificial)
According to；Typing mistake is corrected, cleans apparent abnormal data etc..By itself and step 2-1) in historical traffic data carry out Pearson
Correlation analysis finds that maximum temperature, minimum temperature and the correlation of flow are notable, and related coefficient is more than 0.8, therefore by highest
Temperature, the normalized of minimum temperature are as a result, as history Meteorological Characteristics.
It is first sample 2-3) by sample needed for historical traffic feature in different time periods and the composition modeling of history Meteorological Characteristics
Upset at random, sample is then divided into 10 parts, randomly select 1 part as verification collection, remaining 9 parts are used as training set, training set
22500 samples, verification 2500 samples of collection.
3) training pattern
The training of neural network model is carried out according to the flow chart of Fig. 2, specifically, by training set and the history stream of verification collection
Measure feature and history Meteorological Characteristics are inputted respectively in the input layer of neural network model, and the training flow and verification for obtaining output flow
Amount.It often completes a wheel to train, the sample in training set all can at random be upset once.
Table 1
4) model is verified
It is tested based on the training set and training traffic generating training curve, and based on the verification collection and verification traffic generating
Demonstrate,prove curve.As shown in figure 3, abscissa represents the wheel number of training, each round trains iteration 375 times；Ordinate represents training set
With the mean square error of verification collection.As can be seen that when the mean square error stabilization of training curve and verification curve is in steady state value, explanation
Model stability, fitting is preferable, is suitable for the neural network model of identification.
On the other hand, if the mean square error of two curves is not stablized in steady state value, illustrate that model is unstable, be fitted compared with
Difference, training data is very little or model parameter also needs to optimize.
In this example, trained obtained neural network model can be illustratively expressed as：
Qt=Qt-i×W+b
In above formula, QtFor the water requirement of moment t to be predicted；
Qt-iFor the history water requirement at t-i moment, i=1,2,3 ...；
W is weight matrix；
B is bias term.
2 leakage accident of embodiment identifies
It is known that any leakage accident did not occurred 1 day 2 months to 10 days 2 months 2017, with 2 days -2 months on the 1st 2 months 2017
Data on flows builds reference data set, and for identifying the leakage accident of 3 days 2 months to 10 days 2 months 2017.
In order to verify recognition effect, the artificial data on flows for giving 4,5 and 6 days 2 months 2017 increases the DMA respectively
Daily average water discharge (79m3/ h) 5%, 10% and 15%.
The identification of leakage accident is carried out according to flow chart shown in Fig. 4, specifically,
Step 1, obtained in the way of step 2) in embodiment 1 do not occur on 2 1st, 2017 of leakage accident-
Reference flow feature in the 48h of on 2 2nd, 2017 and with reference to Meteorological Characteristics.
The reference flow feature in different time periods that step 1 obtains is input to embodiment 1 and trains completion by step 2
In multiple gating cycle elementary layers of neural network model, and the reference Meteorological Characteristics that step 1 is obtained are input to described first
In full articulamentum, the predicted flow rate of output is obtained.
Step 3 obtains the measured discharge data in the 48h of on 2 2nd, 1 on the 1st 2 months 2017, by each actual measurement
The predicted flow rates that the step of flow is with being in mutually in the same time two obtains form a time series vectors.So, at this two days
In, 192 time series vectors (48h, frequency acquisition are 4 times/h) are formed altogether, this 192 time series vectors are formed
Matrix is used as with reference to data set, and table 2 is the example of a part for reference data set.
Table 2
Step 4, in the way of above-mentioned steps one to step 3, by the data on flows of 3 days 2 months to 10 days 2 months 2017
It handles as 768 newly-increased time series vectors (192h, frequency acquisition are 4 times/h).
Based on formula (2), the every of the reference data concentration that each newly-increased time series vector is obtained compared with step 3 is calculated
The COS distance of a time series vector, in 192 COS distances calculated in each newly-increased time series vector, statistics
The number n of time series vector of the COS distance less than distance threshold D.In this example, distance threshold D is 192 COS distances
0.45 times of median, i.e., 0.00133, amount threshold N is 42.
When n is more than amount threshold N, assert that the newly-increased time series vector is normal vector, representing the time recognizes
There is no leakage accident；When n is below amount threshold N, assert that the newly-increased time series vector is exception vector, representing should
Time recognizes leakage accident.
3 recognition result of table
For theoretically, in 768 newly-increased time serieses that 3 days 2 months to 10 days 2 months 2017 amount to 8 days, Ying You
480 normal vectors and 288 exception vectors.According to the recognition result of table 3, the neural network model of embodiment 1 can be with
Identify the smaller leakage loss of outflow (+5% DMA daily average water discharges), but it is insensitive；And for the larger leakage loss of flow (+10% ,+
15% DMA daily average water discharges) it can realize and accurately identify, accuracy rate is more than 85%.
It these results suggest that, based on the neural network model of gating cycle unit, can accurately identify water supply
The leakage accident of pipe network, and the practicability of this method is stronger.The present invention extends grinding for existing pipe network model identification model
Study carefully content, making scientific and reasonable decision-making for Running-water Company provides a kind of new thinking.
Although the present invention has been described in detail, it will be understood by those skilled in the art that in spirit and scope of the invention
Modification will be apparent.However, it should be understood that each side of the invention recorded, different specific embodiments
Each several part and the various features enumerated can be combined or all or part of exchange.In above-mentioned each specific embodiment, that
A little embodiments with reference to another embodiment can be combined suitably with other embodiment, this is will be by this field skill
Art personnel are to understand.In addition, it will be understood to those of skill in the art that the description of front is only exemplary mode, not purport
In the limitation present invention.
Claims (9)
1. a kind of neural network model based on gating cycle unit, for identifying the leakage accident of water supply network, including：
Multiple gating cycle elementary layers are configured to receive water supply network in traffic characteristic in different time periods；
First full articulamentum, composition input layer in parallel with the multiple gating cycle elementary layer, the first full articulamentum quilt
It is configured to receive the Meteorological Characteristics of water supply network；
The merging layer being connected with the input layer with the pattern that tensor is connected；
With the merging layer with the second full articulamentum that the pattern that tensor is connected is connected to the full articulamentums of M, M is to be more than or wait
In 2 integer；And
The output layer being connected with the full articulamentums of M with the pattern that tensor is connected is configured to output water supply network under
The predicted flow rate at one moment.
2. a kind of training method of neural network model described in claim 1, including：
Build the neural network model；
Historical traffic feature in different time periods and history Meteorological Characteristics are randomly divided into training set in proportion and verification collects；
Historical traffic feature in different time periods in the training set is input in the multiple gating cycle elementary layer, and will
History Meteorological Characteristics in the training set are input in the described first full articulamentum, obtain the neural network model output
Training flow；
Historical traffic feature in different time periods is concentrated to be input in the multiple gating cycle elementary layer the verification, and will
The history Meteorological Characteristics that the verification is concentrated are input in the described first full articulamentum, obtain the neural network model output
Verify flow；
Based on the training set and training traffic generating training curve, and it is bent based on the verification collection and verification traffic generating verification
Line when the mean square error value stabilization of the training curve and verification curve is in steady state value, completes the neural network model
Training.
3. training method according to claim 2, which is characterized in that further include and obtain the history stream in different time periods
Measure feature, including：
Obtain the historical traffic data of water supply network to be measured；And
Processing is extracted to the historical traffic data and normalized obtains the historical traffic feature.
4. training method according to claim 3, which is characterized in that processing bag is extracted to the historical traffic data
Include, in the historical traffic data of water supply network, respectively from the first time period, second time period, in the 3rd period with
The data on flows for specifying number is extracted at equal intervals, wherein,
First time period is chosen before moment t to be predicted is closed on,
Second time period is chosen before the initial time of the first time period,
Chose for the 3rd period before the initial time of the second time period.
5. the training method according to claim 3 or 4, which is characterized in that it further includes and obtains the history Meteorological Characteristics, bag
It includes：
Obtain the history meteorological data of water supply network to be measured；And
Processing is extracted to the history meteorological data and normalized obtains the history Meteorological Characteristics.
6. training method according to claim 5, which is characterized in that processing bag is extracted to the history meteorological data
It includes：Pearson correlation analysis is carried out to the history meteorological data and the historical traffic data, related coefficient is chosen and is more than
Or the history meteorological data equal to specified value, obtain the history meteorological data after extraction process.
7. training method according to claim 6, which is characterized in that the specified value is 0.80-0.99.
8. a kind of recognition methods of the leakage accident of water supply network, including,
Feature extraction step within the stipulated time that leakage accident does not occur, obtains the in different time periods of water supply network to be measured
Reference flow feature and with reference to Meteorological Characteristics；
The reference flow feature in different time periods is input to nerve described in claim 1 by predicted flow rate obtaining step
In multiple gating cycle elementary layers of network model, and the of the neural network model is input to reference to Meteorological Characteristics by described
In one full articulamentum, the predicted flow rate of the neural network model output is obtained；
Reference data set determines step, obtains the measured discharge within the stipulated time, every group in prediction mutually in the same time
Flow forms a time series vector with measured discharge, the square that all time series vectors in the stipulated time are formed
Battle array is as with reference to data set；
Identification of accidental events step, calculate newly-increased time series vector compared with each time series that the reference data is concentrated to
The COS distance of amount, statistics COS distance are less than the number of the time series vector of distance threshold, be less than when the number of statistics or
During equal to amount threshold, determine that the leakage accident has occurred；
Wherein, the newly-increased time series vector is that the period beyond the stipulated time obtains according to the feature
Step, predicted flow rate obtaining step, reference data value determine the time series vector that step obtains.
9. recognition methods according to claim 8, which is characterized in that the distance threshold is for the reference data set
In calculate 0.1-0.5 times of the medians of all COS distances of all time series vectors.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710998436.8A CN108051035B (en) | 2017-10-24 | 2017-10-24 | The pipe network model recognition methods of neural network model based on gating cycle unit |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710998436.8A CN108051035B (en) | 2017-10-24 | 2017-10-24 | The pipe network model recognition methods of neural network model based on gating cycle unit |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108051035A true CN108051035A (en) | 2018-05-18 |
| CN108051035B CN108051035B (en) | 2019-08-09 |
Family
ID=62119600
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710998436.8A Active CN108051035B (en) | 2017-10-24 | 2017-10-24 | The pipe network model recognition methods of neural network model based on gating cycle unit |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108051035B (en) |
Cited By (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108900446A (en) * | 2018-05-28 | 2018-11-27 | 南京信息工程大学 | Coordinate transform norm blind balance method based on gating cycle unit neural network |
| CN109359698A (en) * | 2018-10-30 | 2019-02-19 | 清华大学 | Leakage loss recognition methods based on long Memory Neural Networks model in short-term |
| CN109522716A (en) * | 2018-11-15 | 2019-03-26 | 中国人民解放军战略支援部队信息工程大学 | A kind of network inbreak detection method and device based on timing neural network |
| CN110070175A (en) * | 2019-04-12 | 2019-07-30 | 北京市商汤科技开发有限公司 | Image processing method, model training method and device, electronic equipment |
| CN110599468A (en) * | 2019-08-30 | 2019-12-20 | 中国信息通信研究院 | No-reference video quality evaluation method and device |
| CN110837933A (en) * | 2019-11-11 | 2020-02-25 | 重庆远通电子技术开发有限公司 | Leakage identification method, device, equipment and storage medium based on neural network |
| CN111062476A (en) * | 2019-12-06 | 2020-04-24 | 重庆大学 | Water quality prediction method based on gated circulation unit network integration |
| CN112101400A (en) * | 2019-12-19 | 2020-12-18 | 国网江西省电力有限公司电力科学研究院 | Industrial control system abnormality detection method, equipment, server and storage medium |
| CN112118143A (en) * | 2020-11-18 | 2020-12-22 | 迈普通信技术股份有限公司 | Traffic prediction model, training method, prediction method, device, apparatus, and medium |
| CN113280265A (en) * | 2020-02-20 | 2021-08-20 | 中国石油天然气股份有限公司 | Working condition identification method and device, computer equipment and storage medium |
| CN113944888A (en) * | 2021-11-03 | 2022-01-18 | 北京软通智慧科技有限公司 | Gas pipeline leakage detection method, device, equipment and storage medium |
| CN115031776A (en) * | 2022-05-12 | 2022-09-09 | 浙江中控信息产业股份有限公司 | Method for monitoring and analyzing siltation of drainage pipe network |
| CN115654381A (en) * | 2022-10-24 | 2023-01-31 | 电子科技大学 | Water supply pipeline leakage detection method based on graph neural network |
| CN117490002A (en) * | 2023-12-28 | 2024-02-02 | 成都同飞科技有限责任公司 | Water supply network flow prediction method and system based on flow monitoring data |
Citations (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20130066568A1 (en) * | 2010-04-15 | 2013-03-14 | Julio Roberto Alonso | Integrated system with acoustic technology, mass imbalance and neural network for detecting, locating and quantifying leaks in ducts |
| CN103530818A (en) * | 2013-10-12 | 2014-01-22 | 杭州电子科技大学 | Water supply pipe network modeling method based on BRB (belief-rule-base) system |
| CN104061445A (en) * | 2014-07-09 | 2014-09-24 | 中国石油大学(华东) | Pipeline leakage detection method based on neural network |
| CN105221933A (en) * | 2015-08-24 | 2016-01-06 | 哈尔滨工业大学 | A kind of pipeline network leak detecting method in conjunction with resistance identification |
| CN106022518A (en) * | 2016-05-17 | 2016-10-12 | 清华大学 | Pipe damage probability prediction method based on BP neural network |
| CN106287239A (en) * | 2016-08-16 | 2017-01-04 | 浙江大学 | Ball device and method is detected in the intelligence pipe of public supply mains leakage location |
| CN106352244A (en) * | 2016-08-31 | 2017-01-25 | 中国石油化工股份有限公司 | Pipeline leakage detection method based on hierarchical neural network |
| CN206130547U (en) * | 2016-07-07 | 2017-04-26 | 北京信息科技大学 | Gas transmission pipeline leak testing system under multiplex condition |
| CN107013812A (en) * | 2017-05-05 | 2017-08-04 | 西安科技大学 | A kind of THM coupling line leakage method |
| CN107239859A (en) * | 2017-06-05 | 2017-10-10 | 国网山东省电力公司电力科学研究院 | The heating load forecasting method of Recognition with Recurrent Neural Network is remembered based on series connection shot and long term |
-
2017
- 2017-10-24 CN CN201710998436.8A patent/CN108051035B/en active Active
Patent Citations (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20130066568A1 (en) * | 2010-04-15 | 2013-03-14 | Julio Roberto Alonso | Integrated system with acoustic technology, mass imbalance and neural network for detecting, locating and quantifying leaks in ducts |
| CN103530818A (en) * | 2013-10-12 | 2014-01-22 | 杭州电子科技大学 | Water supply pipe network modeling method based on BRB (belief-rule-base) system |
| CN104061445A (en) * | 2014-07-09 | 2014-09-24 | 中国石油大学(华东) | Pipeline leakage detection method based on neural network |
| CN105221933A (en) * | 2015-08-24 | 2016-01-06 | 哈尔滨工业大学 | A kind of pipeline network leak detecting method in conjunction with resistance identification |
| CN106022518A (en) * | 2016-05-17 | 2016-10-12 | 清华大学 | Pipe damage probability prediction method based on BP neural network |
| CN206130547U (en) * | 2016-07-07 | 2017-04-26 | 北京信息科技大学 | Gas transmission pipeline leak testing system under multiplex condition |
| CN106287239A (en) * | 2016-08-16 | 2017-01-04 | 浙江大学 | Ball device and method is detected in the intelligence pipe of public supply mains leakage location |
| CN106352244A (en) * | 2016-08-31 | 2017-01-25 | 中国石油化工股份有限公司 | Pipeline leakage detection method based on hierarchical neural network |
| CN107013812A (en) * | 2017-05-05 | 2017-08-04 | 西安科技大学 | A kind of THM coupling line leakage method |
| CN107239859A (en) * | 2017-06-05 | 2017-10-10 | 国网山东省电力公司电力科学研究院 | The heating load forecasting method of Recognition with Recurrent Neural Network is remembered based on series connection shot and long term |
Cited By (19)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108900446A (en) * | 2018-05-28 | 2018-11-27 | 南京信息工程大学 | Coordinate transform norm blind balance method based on gating cycle unit neural network |
| CN109359698A (en) * | 2018-10-30 | 2019-02-19 | 清华大学 | Leakage loss recognition methods based on long Memory Neural Networks model in short-term |
| CN109522716A (en) * | 2018-11-15 | 2019-03-26 | 中国人民解放军战略支援部队信息工程大学 | A kind of network inbreak detection method and device based on timing neural network |
| CN110070175A (en) * | 2019-04-12 | 2019-07-30 | 北京市商汤科技开发有限公司 | Image processing method, model training method and device, electronic equipment |
| CN110070175B (en) * | 2019-04-12 | 2021-07-02 | 北京市商汤科技开发有限公司 | Image processing method, model training method and device and electronic equipment |
| CN110599468A (en) * | 2019-08-30 | 2019-12-20 | 中国信息通信研究院 | No-reference video quality evaluation method and device |
| CN110837933A (en) * | 2019-11-11 | 2020-02-25 | 重庆远通电子技术开发有限公司 | Leakage identification method, device, equipment and storage medium based on neural network |
| CN111062476A (en) * | 2019-12-06 | 2020-04-24 | 重庆大学 | Water quality prediction method based on gated circulation unit network integration |
| CN112101400A (en) * | 2019-12-19 | 2020-12-18 | 国网江西省电力有限公司电力科学研究院 | Industrial control system abnormality detection method, equipment, server and storage medium |
| CN113280265A (en) * | 2020-02-20 | 2021-08-20 | 中国石油天然气股份有限公司 | Working condition identification method and device, computer equipment and storage medium |
| CN113280265B (en) * | 2020-02-20 | 2022-08-05 | 中国石油天然气股份有限公司 | Working condition identification method and device, computer equipment and storage medium |
| CN112118143A (en) * | 2020-11-18 | 2020-12-22 | 迈普通信技术股份有限公司 | Traffic prediction model, training method, prediction method, device, apparatus, and medium |
| CN112118143B (en) * | 2020-11-18 | 2021-02-19 | 迈普通信技术股份有限公司 | Traffic prediction model training method, traffic prediction method, device, equipment and medium |
| CN113944888A (en) * | 2021-11-03 | 2022-01-18 | 北京软通智慧科技有限公司 | Gas pipeline leakage detection method, device, equipment and storage medium |
| CN113944888B (en) * | 2021-11-03 | 2023-12-08 | 北京软通智慧科技有限公司 | Gas pipeline leakage detection method, device, equipment and storage medium |
| CN115031776A (en) * | 2022-05-12 | 2022-09-09 | 浙江中控信息产业股份有限公司 | Method for monitoring and analyzing siltation of drainage pipe network |
| CN115654381A (en) * | 2022-10-24 | 2023-01-31 | 电子科技大学 | Water supply pipeline leakage detection method based on graph neural network |
| CN117490002A (en) * | 2023-12-28 | 2024-02-02 | 成都同飞科技有限责任公司 | Water supply network flow prediction method and system based on flow monitoring data |
| CN117490002B (en) * | 2023-12-28 | 2024-03-08 | 成都同飞科技有限责任公司 | Water supply network flow prediction method and system based on flow monitoring data |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108051035B (en) | 2019-08-09 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108051035B (en) | The pipe network model recognition methods of neural network model based on gating cycle unit | |
| EP4216117B1 (en) | Method and apparatus for training water-and-sediment prediction model for reservoir and method and apparatus for predicting water-and-sediment in reservoir | |
| CN106530715B (en) | Road network traffic state prediction method based on fuzzy Markov process | |
| CN109630095B (en) | A kind of rod-pumped well operating mode's switch method and system based on multi-angle of view study | |
| CN112464584B (en) | Method for estimating water level and flow of free surface flow | |
| CA2392063A1 (en) | Method and program for simulating a physical system using object-oriented programming | |
| CN111664823B (en) | Method for detecting thickness of scale layer of voltage-sharing electrode based on difference of medium heat conduction coefficients | |
| CN104992010B (en) | A kind of more section joint parameter estimation methods based on topological subregion | |
| CN103914594A (en) | Concrete thermodynamic parameter intelligent recognition method based on support vector machine | |
| CN112149873A (en) | Low-voltage transformer area line loss reasonable interval prediction method based on deep learning | |
| CN106677763A (en) | Oil well dynamic liquid level prediction method based on dynamic integrated modeling | |
| Li et al. | A k-nearest neighbor locally weighted regression method for short-term traffic flow forecasting | |
| CN106257541A (en) | A kind of modification method of bridge finite element model | |
| CN103823988A (en) | Method for predicating and analyzing water quantity and quality coupling simulation in oversized river basin | |
| CN111199298A (en) | Flood forecasting method and system based on neural network | |
| Wang et al. | Roof pressure prediction in coal mine based on grey neural network | |
| Xiyun et al. | Wind power probability interval prediction based on bootstrap quantile regression method | |
| CN103020346B (en) | Test method for physical design similarity of circuit | |
| CN103353295A (en) | Method for accurately predicating vertical deformation of dam body | |
| Li et al. | Jaya-ICSM: A rapid inverse method driven by monitoring data for concrete-faced rockfill dams static displacement simulation | |
| Zhang et al. | BP-PSO-based intelligent case retrieval method for high-rise structural form selection | |
| Huy et al. | Short-term load forecasting in power system using CNN-LSTM neural network | |
| Fang et al. | SAW: A hybrid prediction model for parking occupancy under the environment of lacking real-time data | |
| CN106934729A (en) | Building Testing and appraisal method and device | |
| Jiang et al. | Discharge estimation based on machine learning |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |