CN108764540A - Water supply network pressure prediction method based on parallel LSTM series connection DNN - Google Patents
Water supply network pressure prediction method based on parallel LSTM series connection DNN Download PDFInfo
- Publication number
- CN108764540A CN108764540A CN201810465536.9A CN201810465536A CN108764540A CN 108764540 A CN108764540 A CN 108764540A CN 201810465536 A CN201810465536 A CN 201810465536A CN 108764540 A CN108764540 A CN 108764540A
- Authority
- CN
- China
- Prior art keywords
- model
- dnn
- prediction
- water supply
- supply network
- 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
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 title claims abstract description 50
- 238000000034 method Methods 0.000 title claims abstract description 29
- 230000004913 activation Effects 0.000 claims abstract description 13
- 238000011478 gradient descent method Methods 0.000 claims abstract description 11
- 238000005516 engineering process Methods 0.000 claims abstract description 6
- 238000012549 training Methods 0.000 claims description 31
- 238000012544 monitoring process Methods 0.000 claims description 13
- 238000013136 deep learning model Methods 0.000 claims description 12
- 238000013528 artificial neural network Methods 0.000 claims description 7
- 238000000605 extraction Methods 0.000 claims description 7
- 210000002569 neuron Anatomy 0.000 claims description 7
- 230000000694 effects Effects 0.000 claims description 6
- 238000012545 processing Methods 0.000 claims description 6
- 230000006870 function Effects 0.000 claims description 5
- 241000208340 Araliaceae Species 0.000 claims description 4
- 235000005035 Panax pseudoginseng ssp. pseudoginseng Nutrition 0.000 claims description 4
- 235000003140 Panax quinquefolius Nutrition 0.000 claims description 4
- 235000008434 ginseng Nutrition 0.000 claims description 4
- 230000001537 neural effect Effects 0.000 claims description 3
- 230000003252 repetitive effect Effects 0.000 claims description 3
- 238000007619 statistical method Methods 0.000 claims description 3
- 230000009466 transformation Effects 0.000 claims description 3
- 239000000284 extract Substances 0.000 claims description 2
- 238000004364 calculation method Methods 0.000 abstract description 4
- 238000005457 optimization Methods 0.000 abstract description 2
- 239000013589 supplement Substances 0.000 abstract description 2
- 238000004458 analytical method Methods 0.000 description 8
- 230000008859 change Effects 0.000 description 4
- 238000013459 approach Methods 0.000 description 3
- 230000007812 deficiency Effects 0.000 description 3
- 238000004519 manufacturing process Methods 0.000 description 3
- 238000006243 chemical reaction Methods 0.000 description 2
- 238000004140 cleaning Methods 0.000 description 2
- 230000014509 gene expression Effects 0.000 description 2
- 230000004807 localization Effects 0.000 description 2
- 238000010606 normalization Methods 0.000 description 2
- 238000003860 storage Methods 0.000 description 2
- 230000002463 transducing effect Effects 0.000 description 2
- 101001095088 Homo sapiens Melanoma antigen preferentially expressed in tumors Proteins 0.000 description 1
- 102100037020 Melanoma antigen preferentially expressed in tumors Human genes 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000000295 complement effect Effects 0.000 description 1
- 230000002950 deficient Effects 0.000 description 1
- 230000008034 disappearance Effects 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000013213 extrapolation Methods 0.000 description 1
- 230000004927 fusion Effects 0.000 description 1
- 238000009499 grossing Methods 0.000 description 1
- 230000036039 immunity Effects 0.000 description 1
- 230000007787 long-term memory Effects 0.000 description 1
- 238000005312 nonlinear dynamic Methods 0.000 description 1
- 238000000611 regression analysis Methods 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
- 238000004088 simulation Methods 0.000 description 1
- 239000008400 supply water Substances 0.000 description 1
- 238000012706 support-vector machine Methods 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
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"
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- 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
- G06N3/048—Activation functions
-
- 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
- G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
- G06Q50/06—Energy or water supply
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/02—Preprocessing
- G06F2218/04—Denoising
- G06F2218/06—Denoising by applying a scale-space analysis, e.g. using wavelet analysis
Landscapes
- Engineering & Computer Science (AREA)
- Business, Economics & Management (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Economics (AREA)
- General Physics & Mathematics (AREA)
- Strategic Management (AREA)
- Health & Medical Sciences (AREA)
- Data Mining & Analysis (AREA)
- Human Resources & Organizations (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- General Business, Economics & Management (AREA)
- General Engineering & Computer Science (AREA)
- Tourism & Hospitality (AREA)
- Marketing (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Development Economics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Evolutionary Biology (AREA)
- Computational Linguistics (AREA)
- Game Theory and Decision Science (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Entrepreneurship & Innovation (AREA)
- Computing Systems (AREA)
- Operations Research (AREA)
- Quality & Reliability (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Public Health (AREA)
- Water Supply & Treatment (AREA)
- Primary Health Care (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
- Feedback Control In General (AREA)
Abstract
The invention discloses a kind of water supply network pressure prediction methods based on parallel LSTM series connection DNN.The present invention determines input and output item first, establishes the prediction model for the DNN that connects based on parallel LSTM.Secondly data prediction, establishes pressure prediction database.Then prediction model is trained.Finally carry out online pressure prediction.The present invention realizes the mutual supplement with each other's advantages of LSTM and DNN, it is used in combination Dropout technologies to prevent model over-fitting, Relu activation primitives accelerate model convergence rate, small lot gradient descent method reduces randomness and calculation amount, RMSprop is selected as the optimization algorithm of stochastic gradient descent method, improves the anti-interference and precision of water supply network pressure prediction method.
Description
Technical field
The invention belongs to urban water supply field, specifically a kind of water supply network pressure based on parallel LSTM series connection DNN is pre-
Survey method.
Background technology
Water supply pipe net system be one it is complicated, in large scale, with the strong nonlinear dynamic system of water randomness, can
Quick and precisely simulation and the operating condition for predicting pipe network, are the key that carry out water supply network Optimized Operation.Water supply network is given birth to
In production operation, dispatcher generally observes the operation conditions of pipe network with pressure measured data.Therefore, monitoring point pressure is carried out
Forecast analysis contributes to dispatcher to prejudge in advance, carries out production commander.
It is big to be generally divided into time series method, structured analysis method and systems approach three for water supply network pressure prediction method at present
Class.Wherein time series method includes the method for moving average, exponential smoothing, trend extrapolation etc.;Structured analysis method includes returning to divide
Analysis method etc.;Systems approach includes gray prediction, artificial neural network etc..Time series models precision of prediction preferably, data processing
Simply, but the short-term time series data of measuring point can only be utilized, the case where being not particularly suited for changing greatly;Regression analysis model letter
Single, convenient but by many factors combined influence when, is difficult to select;Systems approach has self-learning capability, Nonlinear Processing etc. excellent
Point, but there are models it is complicated, the training time is long the problems such as.Easily by noise jamming, precision of prediction is difficult to ensure the above method.
Invention content
For the highly complex nonlinear characteristic of water supply network and the deficiency of art methods, the present invention proposes a kind of base
Ductwork pressure is carried out in the deep learning model of parallel LSTM (shot and long term Memory Neural Networks) series connection DNN (deep neural network)
Prediction improves precision of prediction.
Since the quantity of state and controlled quentity controlled variable of water supply network are two distinct types of characteristic informations, if simply using one
LSTM models carry out feature extraction, and two category feature information will be unable to highlight to the Different Effects of model, for this purpose, by single LSTM moulds
Type is extended to parallel LSTM models, extracts respectively, learns the different characteristic information of two classes.Again due to when LSTM, which is good at processing, to be based on
Between sequence data, DNN is suitable for characteristic information being mapped to higher space, utilizes LSTM and the advantages of DNN respectively, general
LSTM and DNN join together the framework unified as one, realize and have complementary advantages, i.e., will be after the fusion of the output result of two-way LSTM
It is exported by DNN, the measuring point pressure of subsequent time is predicted in realization.The present invention provides one kind based on parallel as a result,
The water supply network pressure prediction method of LSTM series connection DNN deep learning models.
To realize that high noise immunity, the high purpose of high precision of prediction, the present invention take following steps:
1, it determines input and output item, establishes the prediction model for the DNN that connects based on parallel LSTM
It is the nonlinear systems with delay of a multiple-input and multiple-output in view of water supply pipe net system, selects longer historic state
Amount (measuring point pressure information) [x (t), x (t-1) ..., x (t-ns)] and controlled quentity controlled variable (water inlet pressure and flow) [u (t-1), u
(t-2),…,u(t-nu)] it is used as input item, with the deficiency of compensation " water supply network quantity of state is only partly understood ";Determine output item
For the output y at pressure-measuring-point t+1 momentm(t+1).Here, ns、nuFor historical time window.
Establish the deep learning model for the DNN that connects based on parallel LSTM:
A. be respectively adopted LSTM models to quantity of state [x (t), x (t-1) ..., x (t-ns)] and controlled quentity controlled variable [u (t-1), u
(t-2),…,u(t-nu)] carry out feature extraction and study.The respectively state variable of water supply network and control
Output valve of the variable processed through LSTM models.
B. use deep-neural-network DNN models willFusion treatment is carried out, output y is obtainedm(t+1),
Formula (1) can be used to describe:
ym(t+1) it is exported for the prediction of deep learning model.[] indicates there is two kinds on time dimension same dimension
Square merge, H () be DNN models activation primitive.WDNN、bDNNThe respectively weights and threshold value of DNN models.
2, data prediction establishes pressure prediction database
(1) data prediction
Data are filled a vacancy:For the data from SCADA system collection in worksite, there are data loss problems, using linear, throwing
The data of object line or cubic curve interpolation completion missing.If missing data is excessive, then the historical data of this period is abandoned.
Data de-noising:For field data, there are much noise interference problems, and noise is removed using wavelet transformation.Small echo becomes
Transducing carries out localization signal analysis, it can be achieved that multiresolution analysis, judges noise and jump signal in time domain and frequency domain, determines
Useful signal.
Dimensionless processing:There is different physics dimension sum number magnitude problems, logarithm for water supply network pressure and flow
According to normalized is made, i.e., input is limited to [0, l] with output, make they with identical grade participate in model training with it is pre-
It surveys, specific formula is such as shown in (2)
In formula, X expressions need normalization data, Max (X), Min (X) to respectively represent minimum value and maximum value, XnorRepresentative is returned
Data after one change.
(2) pressure prediction database is established
Establish water supply network pressure prediction database:Data item is other than timestamp, node (monitoring point or entrance), packet
It includes:(1) pressure of measuring point, flow value, pressure, flow value of entrance etc., extraction/cleaning/conversion in real time from SCADA, and deposit
Storage, the input item as model;(2) forecast pressure of measuring point comes from model prediction, is the output item of model;(3) error information
, it is used for statistical analysis precision of prediction.
3, training prediction model
(1) training sample is determined
DMA subregions around large-scale water supply network or small-sized water supply network determine that input sample is { X (ns),U(nu),
Y }, wherein X (ns) it is that i ties up state variable, U (nu) it is 2j dimension control variables, i is that monitoring is counted, and j is entrance number,.
To ensure to train, input sample data time span must assure that Max (ns,nu) a period is continuous, must generally it ensure
{ X (the n of 1 hour or more continuous effectives),U(nu), Y } data, it is a that effective sample is no less than 12x24x15=4320.
(2) it determines model basic structure, remaining initial parameter values is set, start training pattern
Rule of thumb or just the effect of step ginseng determines the value range of parameter.nu、ns∈ { 1,2 ... 12 }, time step
T=5 minutes long, i.e., historical information maximum span is 60 minutes, since longer historical information can make input redundancy, and to improving
Precision of prediction has no much influences;Hidden layer number Layers ∈ { 1,2 ..., 5 }, spy can be improved by increasing the number of plies of hidden layer
Sign extraction and learning ability, but multilayer can make model become to become increasingly complex;Corresponding neuronal quantity Neurons ∈ [0,300],
The quantity of neuron determines the nonlinear degree of network training.
There is over-fitting in deep learning model in order to prevent, and the present invention can be random after introducing Dropout at each layer
Ground updates network parameter, increases the generalization ability of model.Dropout technology specific practices are to abandon one at random in model training
The hidden layer node (but weight can preserve, only temporarily without update) of certainty ratio, and restore full connection when model uses.Section
Point abandons ratio dropout rate ∈ [01,0.5], and the selection of ratio is abandoned for node, is not had if ratio is too low
Effect, ratio too Gao Zehui lead to the deficient study of model.
The present invention is using small lot (Mini-batch) gradient descent method come the parameters in Optimized model, this method
Several batches are divided the data into, carry out undated parameter by batch (batch), in this way, one group of data in one batch have codetermined ginseng
Several updates, reduces randomness and calculation amount.The sample size Mini_batch ∈ [5,50] of small lot gradient descent method.Instruction
Practice wheel number epoch ∈ [100,200], frequency of training is not achieved training effect, excessively can't improve precision of prediction instead very little
Increase the training time.
Two parts use activation primitive in a model, first, calculating LSTM layers for input value, are also useful for feature and melt
Output layer after conjunction.Traditional saturation activation function, if sigmoid and tanh can bring gradient disappearance problem, and ReLU etc. is non-
Saturation activation function can accelerate model convergence rate relative to saturation activation function;Depth model using ReLU is having prison
Close even preferably result can be obtained by not needing pre-training before supervising and instructing white silk.Invention activation function chooses ReLU.
(3) training iteration
In model training, generally pass through the predicted value of modelRoot mean square is found out with measured value y as model error, such as
Shown in formula (3).Wherein, n is output layer neuron node number.
Work as loss<Error target ε ∈ [0.2%, 0.5%], reach training requirement, iteration terminates.The tune when error is larger
The parameters of integral mould change model basic structure, i.e., give again if error does not meet the condition of convergence and no longer reduces
One { nu,ns, Layers }, other parameters are adjusted further according to each basic structure, again repetitive exercise.
4, online pressure prediction
By the pressure of measuring point, flow value in pressure prediction database, the continuous effectives data such as pressure, flow value of entrance,
It is sequentially inputted to model, model then provides the pressure prediction value y at t+1 momentm(t+1), it can be supplied within about t=5 minutes in advance
Dispatcher refers to.
Meanwhile by pressure prediction value ym(t+1) it is stored in database, is compared with the measured value y (t+1) at t+1 moment, is counted
Calculate Δ=ym(t+1)-y(t+1).If allowing to predict that error is σ ∈ [5%, 10%], if continuous Δ three times>σ * y (t+1), then return
Step 3 is returned, re -training model updates model parameter with most recent data.
Beneficial effects of the present invention:The present invention proposes a kind of deep learning model for the DNN that connects based on parallel LSTM, realizes
The mutual supplement with each other's advantages of LSTM and DNN is used in combination Dropout technologies to prevent model over-fitting, Relu activation primitives from accelerating model convergence speed
Degree, small lot gradient descent method reduce randomness and calculation amount, and RMSprop is selected to be calculated as the optimization of stochastic gradient descent method
Method improves the anti-interference and precision of water supply network pressure prediction method.
Description of the drawings
Fig. 1:Prediction model based on parallel LSTM series connection DNN;
Fig. 2:Water supply network monitoring point pressure prediction system application framework.
Specific implementation mode
Technological means to make the present invention realize is readily apparent from creation characteristic, with reference to the accompanying drawings and examples, to this
The realization method of invention is described in further detail, and is not intended to limit the interest field of the present invention.
The areas Y of the cities Xian Yimou water supply network is example, which is roughly equal to 106.7km2, day water supply be about 150000m3, prison
Measuring point includes water inlet flow meter pressure measuring point, ductwork pressure measuring point, water outlet flow measuring point, intermediate conduit measuring point.
Important pressure monitoring point information is shown in Table 1 inside specific pipe network
1 pipe network of table, 17 important pressure monitoring point information
1, it determines input and output item, establishes the prediction model for the DNN that connects based on parallel LSTM
It is the nonlinear systems with delay of a multiple-input and multiple-output in view of water supply pipe net system, selects longer historic state
Amount [x (t), x (t-1) ..., x (t-ns)] and controlled quentity controlled variable [u (t-1), u (t-2) ..., u (t-nu)] it is used as input item, with compensation
The deficiency of " water supply network quantity of state is only partly understood ";Determine that output item is the output y at pressure-measuring-point t+1 momentm(t+1).Its
In, ns、nuFor historical time window.
Here, with the pressure of supply water of 4 water inlets of the areas Y of city water supply pipe net system, the history control information of water supply
It is input item with 17 measuring point pressure histories and current state information, using the pressure of 17 monitoring point subsequent times as output
?.
Establish the deep learning model (as shown in Figure 1) for the DNN that connects based on parallel LSTM:
A. be respectively adopted LSTM models to quantity of state [x (t), x (t-1) ..., x (t-ns)] and controlled quentity controlled variable [u (t-1), u
(t-2),…,u(t-nu)] carry out feature extraction and study.The respectively state variable of water supply network and control
Output valve of the variable processed through LSTM models.
B. use deep-neural-network DNN models willFusion treatment is carried out, output y is obtainedm(t+1),
Formula (4) can be used to describe:
ym(t+1) it is exported for the prediction of deep learning model.[] indicates there is two kinds on time dimension same dimension
Square merge, H () be DNN models activation primitive.WDNN、bDNNThe respectively weights and threshold value of DNN models.
2, data prediction establishes pressure prediction database
(1) data prediction
Data are filled a vacancy:For the data from SCADA system collection in worksite, there are data loss problems, using linear, throwing
The data of object line or cubic curve interpolation completion missing.If missing data is excessive, then the historical data of this period is abandoned.
Data de-noising:For field data, there are much noise interference problems, and noise is removed using wavelet transformation.Small echo becomes
Transducing carries out localization signal analysis, it can be achieved that multiresolution analysis, judges noise and jump signal in time domain and frequency domain, determines
Useful signal.
Dimensionless processing:There is different physics dimension sum number magnitude problems, logarithm for water supply network pressure and flow
According to normalized is made, i.e., input is limited to [0, l] with output, make they with identical grade participate in model training with it is pre-
It surveys, specific formula is such as shown in (5)
In formula, X expressions need normalization data, Max (X), Min (X) to respectively represent minimum value and maximum value, XnorRepresentative is returned
Data after one change.
(2) pressure prediction database is established
Establish water supply network pressure prediction database:Data item is other than timestamp, node (monitoring point or entrance), packet
It includes:(1) pressure of measuring point, flow value, pressure, flow value of entrance etc., extraction/cleaning/conversion in real time from SCADA, and deposit
Storage, the input item as model;(2) forecast pressure of measuring point comes from model prediction, is the output item of model;(3) error information
, it is used for statistical analysis precision of prediction.
3, training prediction model
(1) training sample is determined
Around the areas Y of certain city water supply network, determine that input sample is for { X (ns),U(nu), Y }, wherein X (ns) it is 17 dimension shapes
State variable, U (nu) it is that 8 dimensions control variable, 17 monitoring points, 4 entrances, t is time step.
To ensure to train, input sample data time span must assure that Max (ns,nu) a period is continuous, must generally it ensure
{ X (the n of 1 hour or more continuous effectives),U(nu), Y } data, it is a that effective sample is no less than 12x24x15=4320
Here, sample data set be on June 28,27 days to 2016 May in 2016, totally 35 days data, sampling interval be
5 minutes, continuous effective, sample size 10080.Wherein the data in June May 28 to 26 days are for training.
(2) it determines model basic structure, remaining initial parameter values is set, start training pattern
Rule of thumb or just the effect of step ginseng determines the value range of parameter.nu、ns∈ { 1,2 ... 12 }, time step
T=5 minutes long, i.e., historical information maximum span is 60 minutes, since longer historical information can make input redundancy, and to improving
Precision of prediction has no much influences;Hidden layer the number layer=2, corresponding neuronal quantity Neurons=100, DNN of LSTM
Layer=2, corresponding Neurons=96.
There is over-fitting in deep learning model in order to prevent, and the present invention can be random after introducing Dropout at each layer
Ground updates network parameter, increases the generalization ability of model.In the present embodiment, the node of regularization abandons ratio dropout rate
=0.3.
The present embodiment is using small lot (Mini-batch) gradient descent method come the parameters in Optimized model, small lot
The sample size Mini_batch=32 of gradient descent method, exercise wheel number epoch=100.
The present embodiment activation primitive chooses ReLU.
(3) training iteration
In model training, generally pass through the predicted value of modelRoot mean square is found out with measured value y as model error, such as
Shown in formula (6).Wherein, n is output layer neuron node number.
Work as loss<Error target ε=0.3%, reaches training requirement, and iteration terminates.Model is adjusted when error is larger
Parameters change model basic structure if error does not meet the condition of convergence and no longer reduces, i.e., give { a n againu,
ns, Layers }, other parameters are adjusted further according to each basic structure, again repetitive exercise.
4, online pressure prediction
The pressure of measuring point, flow value in pressure prediction database, the continuous effectives data such as pressure, flow value of entrance, according to
Secondary to be input to model, model then provides the pressure prediction value y at t+1 momentm(t+1), tune can be supplied within about t=5 minutes in advance
Degree personnel refer to.
Meanwhile by pressure prediction value ym(t+1) it is stored in database, is compared with the measured value y (t+1) at t+1 moment, is counted
Calculate Δ=ym(t+1)-y(t+1).If allowing to predict that error is σ=5%, if continuous Δ three times>σ * y (t+1), then Boot Model
Re -training (step 3) updates model parameter with most recent data.
The data in June 27 to June 28 for testing, are passed through RMSE (root-mean-square error) and MAPE by the present embodiment
(average absolute percent error) provides each monitoring point performance indicator as evaluation performance indicator, table 2, has very high prediction essence
Degree.
2 each monitoring point estimated performance index of table
In order to be compared with conventional model, respectively tentative calculation BP neural network, SVM support vector machines, VARX, NARX,
It has obtained and traditional prediction method comparison result (being shown in Table 3).
The comparison of table 3 and traditional prediction method
Prediction technique | BP | SVM | VARX | NARX | The method of the present invention |
RMSE*100 (each measuring point average value) | 0.56 | 0.43 | 0.30 | 0.28 | 0.17 |
It is obtained by table 3, Classical forecast mould will be significantly better than based on parallel LSTM serial Ds NN deep learning model predictions result
Type.
The method of the present invention can be used for reality according to pressure prediction system application framework in water supply network monitoring point shown in Fig. 2
Production.
Claims (7)
1. the water supply network pressure prediction method for the DNN that connected based on parallel LSTM, it is characterised in that this method is specifically:
Step (1) determines input and output item, establishes the prediction model for the DNN that connects based on parallel LSTM, specifically:
It is the nonlinear systems with delay of a multiple-input and multiple-output in view of water supply pipe net system, selects historic state amount [x (t), x
(t-1),…,x(t-ns)] and controlled quentity controlled variable [u (t-1), u (t-2) ..., u (t-nu)] it is used as input item;Determine that output item is pressure
The output y at measuring point t+1 momentm(t+1), n heres、nuFor historical time window;
Establish the deep learning model for the DNN that connects based on parallel LSTM:
A. be respectively adopted LSTM models to quantity of state [x (t), x (t-1) ..., x (t-ns)] and controlled quentity controlled variable [u (t-1), u (t-
2),…,u(t-nu)] feature extraction and study are carried out, it sets simultaneouslyRespectively the quantity of state of water supply network and
Output valve of the controlled quentity controlled variable through LSTM models;
B. use deep-neural-network DNN models willFusion treatment is carried out, output y is obtainedm(t+1)
Wherein [] indicates that the square with same dimension merges on time dimension by two kinds, and H () is the activation of DNN models
Function;WDNN、bDNNThe respectively weights and threshold value of DNN models;
Step (2) data prediction, establishes pressure prediction database
(2-1) data prediction
Data are filled a vacancy:For the data from SCADA system collection in worksite, there are data loss problems, using linear, parabola
Or the data of cubic curve interpolation completion missing;
Data de-noising:For field data, there are much noise interference problems, and noise is removed using wavelet transformation;
Dimensionless processing:There is different physics dimension sum number magnitude problems for water supply network pressure and flow, data are made
Input is limited to [0, l] with output, them is made to participate in model training and prediction with identical grade by normalized;
(2-2) establishes pressure prediction database
Data item further includes other than timestamp, node:(1) pressure of measuring point, flow value, the pressure of entrance, flow value, from
It extracts, clean, convert in real time in SCADA system and store, the input item as model;(2) forecast pressure of measuring point, comes from
Model prediction is the output item of model;(3) error information item is used for statistical analysis precision of prediction;
Step (3) trains prediction model
(3-1) determines training sample
DMA subregions around large-scale water supply network or small-sized water supply network determine that input sample is { X (ns),U(nu), Y },
Middle X (ns) it is that i ties up quantity of state, U (nu) it is that 2j ties up controlled quentity controlled variable, i is that monitoring is counted, and j is entrance number;
(3-2) determines model basic structure, remaining initial parameter values is arranged, and starts training pattern
Rule of thumb or just the effect of step ginseng determines the value range of parameter;nu、ns∈ { 1,2 ... 12 }, time step t=
5 minutes, i.e. historical information maximum span is 60 minutes;Hidden layer number Layers ∈ 1,2 ..., 5 };Corresponding neuronal quantity
Neurons∈[0,300];
(3-3) trains iteration
In model training, the predicted value of model is crossedRoot mean square is found out as model error loss with measured value y, works as loss<Accidentally
Poor target ε ∈ [0.2%, 0.5%], reach training requirement, iteration terminates;The each of model is adjusted when error is unsatisfactory for requiring
Parameter changes model basic structure if error does not meet the condition of convergence and no longer reduces, i.e., gives { a n againu,ns,
Layers }, other parameters are adjusted further according to each basic structure, again repetitive exercise;
Step (4) online pressure prediction
By the pressure of measuring point, flow value in pressure prediction database, the pressure of entrance, flow value are sequentially inputted to model, model
Then provide the pressure prediction value y at t+1 momentm(t+1);Meanwhile by pressure prediction value ym(t+1) it is stored in database, with the t+1 moment
Measured value y (t+1) be compared, calculate Δ=ym(t+1)-y(t+1);If allowing to predict that error is σ ∈ [5%, 10%],
If continuous Δ three times>σ * y (t+1), then return to step (3), re -training model update model parameter with most recent data.
2. the water supply network pressure prediction method of the DNN according to claim 1 that connected based on parallel LSTM, feature are existed
In:There is over-fitting in deep learning model in order to prevent in step (3-2), at each layer introduce Dropout technologies after with
Network parameter is updated to machine, increases the generalization ability of model.
3. the water supply network pressure prediction method of the DNN according to claim 2 that connected based on parallel LSTM, feature are existed
In:The Dropout technologies are specifically to abandon a certain proportion of hidden layer node at random in model training, but weight can protect
It deposits, only temporarily without update, and restores full connection when model uses;Node abandons ratio between 0.1 to 0.5.
4. the water supply network pressure prediction method of the DNN according to claim 1 that connected based on parallel LSTM, feature are existed
In:Using small lot gradient descent method come the parameters in Optimized model in step (3-2).
5. the water supply network pressure prediction method of the DNN according to claim 1 that connected based on parallel LSTM, feature are existed
In:The small lot gradient descent method is to divide the data into several batches, carrys out undated parameter by batch;Small lot gradient descent method
Sample size be 5 to 50 between;Between exercise wheel number is 100 to 200.
6. the water supply network pressure prediction method of the DNN according to claim 1 that connected based on parallel LSTM, feature are existed
In:The activation primitive of model selects ReLU activation primitives in step (3-2).
7. the water supply network pressure prediction method of the DNN according to claim 1 that connected based on parallel LSTM, feature are existed
In:Pressure prediction value y in step (4)m(t+1) shifting to an earlier date 5 minutes is supplied to dispatcher to refer to.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810465536.9A CN108764540B (en) | 2018-05-16 | 2018-05-16 | Water supply network pressure prediction method based on parallel LSTM series DNN |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810465536.9A CN108764540B (en) | 2018-05-16 | 2018-05-16 | Water supply network pressure prediction method based on parallel LSTM series DNN |
Publications (2)
Publication Number | Publication Date |
---|---|
CN108764540A true CN108764540A (en) | 2018-11-06 |
CN108764540B CN108764540B (en) | 2021-11-30 |
Family
ID=64007993
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810465536.9A Active CN108764540B (en) | 2018-05-16 | 2018-05-16 | Water supply network pressure prediction method based on parallel LSTM series DNN |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108764540B (en) |
Cited By (13)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109930658A (en) * | 2019-03-27 | 2019-06-25 | 杭州电子科技大学 | A kind of water supply network monitoring point method for arranging based on System Observability |
CN110545208A (en) * | 2019-09-23 | 2019-12-06 | 电子科技大学 | Network traffic prediction method based on LSTM |
CN111210093A (en) * | 2020-03-05 | 2020-05-29 | 重庆森鑫炬科技有限公司 | Daily water consumption prediction method based on big data |
CN112115999A (en) * | 2020-09-15 | 2020-12-22 | 燕山大学 | Wind turbine generator fault diagnosis method of space-time multi-scale neural network |
CN112252292A (en) * | 2020-10-15 | 2021-01-22 | 河北工业大学 | Real-time highway compactness monitoring method based on artificial neural network |
CN112434847A (en) * | 2020-11-17 | 2021-03-02 | 上海东普信息科技有限公司 | Express delivery quantity prediction method, device, equipment and storage medium based on LSTM model |
CN112561174A (en) * | 2020-12-18 | 2021-03-26 | 西南交通大学 | Method for predicting geothermal energy production based on LSTM and MLP superimposed neural network |
CN112594553A (en) * | 2020-12-07 | 2021-04-02 | 熊猫智慧水务有限公司 | Pipe network pressure regulation and control method based on pressure target curve |
CN112818495A (en) * | 2021-02-22 | 2021-05-18 | 成都四为电子信息股份有限公司 | Novel dynamic correction method for pipeline pressure drop measurement and calculation algorithm parameters |
CN112857658A (en) * | 2021-01-13 | 2021-05-28 | 淮阴工学院 | Pressure detection system based on cloud platform |
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 |
CN113378335A (en) * | 2021-05-07 | 2021-09-10 | 广州观必达数据技术有限责任公司 | Water supply network pressure prediction method and system based on machine learning |
CN114358244A (en) * | 2021-12-20 | 2022-04-15 | 淮阴工学院 | Pressure big data intelligent detection system based on Internet of things |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103258235A (en) * | 2013-05-13 | 2013-08-21 | 杭州电子科技大学 | Water supply network reorganization and expansion optimization method based on improved artificial fish school algorithm |
US20160099010A1 (en) * | 2014-10-03 | 2016-04-07 | Google Inc. | Convolutional, long short-term memory, fully connected deep neural networks |
CN105868852A (en) * | 2016-03-28 | 2016-08-17 | 广州地理研究所 | Method for predicting daily water consumption of urban dweller |
US20180047389A1 (en) * | 2016-08-12 | 2018-02-15 | Electronics And Telecommunications Research Institute | Apparatus and method for recognizing speech using attention-based context-dependent acoustic model |
-
2018
- 2018-05-16 CN CN201810465536.9A patent/CN108764540B/en active Active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103258235A (en) * | 2013-05-13 | 2013-08-21 | 杭州电子科技大学 | Water supply network reorganization and expansion optimization method based on improved artificial fish school algorithm |
US20160099010A1 (en) * | 2014-10-03 | 2016-04-07 | Google Inc. | Convolutional, long short-term memory, fully connected deep neural networks |
CN105868852A (en) * | 2016-03-28 | 2016-08-17 | 广州地理研究所 | Method for predicting daily water consumption of urban dweller |
US20180047389A1 (en) * | 2016-08-12 | 2018-02-15 | Electronics And Telecommunications Research Institute | Apparatus and method for recognizing speech using attention-based context-dependent acoustic model |
Cited By (19)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109930658A (en) * | 2019-03-27 | 2019-06-25 | 杭州电子科技大学 | A kind of water supply network monitoring point method for arranging based on System Observability |
CN110545208A (en) * | 2019-09-23 | 2019-12-06 | 电子科技大学 | Network traffic prediction method based on LSTM |
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 |
CN111210093A (en) * | 2020-03-05 | 2020-05-29 | 重庆森鑫炬科技有限公司 | Daily water consumption prediction method based on big data |
CN112115999A (en) * | 2020-09-15 | 2020-12-22 | 燕山大学 | Wind turbine generator fault diagnosis method of space-time multi-scale neural network |
CN112115999B (en) * | 2020-09-15 | 2022-07-01 | 燕山大学 | Wind turbine generator fault diagnosis method of space-time multi-scale neural network |
CN112252292A (en) * | 2020-10-15 | 2021-01-22 | 河北工业大学 | Real-time highway compactness monitoring method based on artificial neural network |
CN112252292B (en) * | 2020-10-15 | 2021-05-28 | 河北工业大学 | Real-time highway compactness monitoring method based on artificial neural network |
CN112434847A (en) * | 2020-11-17 | 2021-03-02 | 上海东普信息科技有限公司 | Express delivery quantity prediction method, device, equipment and storage medium based on LSTM model |
CN112594553A (en) * | 2020-12-07 | 2021-04-02 | 熊猫智慧水务有限公司 | Pipe network pressure regulation and control method based on pressure target curve |
CN112561174A (en) * | 2020-12-18 | 2021-03-26 | 西南交通大学 | Method for predicting geothermal energy production based on LSTM and MLP superimposed neural network |
CN112857658A (en) * | 2021-01-13 | 2021-05-28 | 淮阴工学院 | Pressure detection system based on cloud platform |
CN112857658B (en) * | 2021-01-13 | 2022-06-07 | 淮阴工学院 | Pressure detection system based on cloud platform |
CN112818495A (en) * | 2021-02-22 | 2021-05-18 | 成都四为电子信息股份有限公司 | Novel dynamic correction method for pipeline pressure drop measurement and calculation algorithm parameters |
CN112818495B (en) * | 2021-02-22 | 2024-09-06 | 成都四为电子信息股份有限公司 | Dynamic correction method for novel pipeline pressure drop calculation algorithm parameters |
CN113378335A (en) * | 2021-05-07 | 2021-09-10 | 广州观必达数据技术有限责任公司 | Water supply network pressure prediction method and system based on machine learning |
CN114358244A (en) * | 2021-12-20 | 2022-04-15 | 淮阴工学院 | Pressure big data intelligent detection system based on Internet of things |
CN114358244B (en) * | 2021-12-20 | 2023-02-07 | 淮阴工学院 | Big data intelligent detection system of pressure based on thing networking |
Also Published As
Publication number | Publication date |
---|---|
CN108764540B (en) | 2021-11-30 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108764540A (en) | Water supply network pressure prediction method based on parallel LSTM series connection DNN | |
Kisi et al. | Forecasting daily lake levels using artificial intelligence approaches | |
Castillo et al. | Hybrid intelligent systems for time series prediction using neural networks, fuzzy logic, and fractal theory | |
CN103280114B (en) | Signal lamp intelligent control method based on BP-PSO fuzzy neural network | |
CN108108832A (en) | A kind of oil truck oil and gas leakage intelligent monitor system based on wireless sensor network | |
CN103489039B (en) | There is the freeway traffic flow amount fusion forecasting method of online self-tuning optimization ability | |
CN106815782A (en) | A kind of real estate estimation method and system based on neutral net statistical models | |
CN108053052B (en) | A kind of oil truck oil and gas leakage speed intelligent monitor system | |
CN115099519B (en) | Oil well yield prediction method based on multi-machine learning model fusion | |
Lv et al. | An improved long short-term memory neural network for stock forecast | |
CN112163671A (en) | New energy scene generation method and system | |
CN108446771A (en) | A method of preventing Sale Forecasting Model over-fitting | |
CN111311577A (en) | Intelligent water seepage detection method based on generation of confrontation network and reinforcement learning | |
Buragohain | Adaptive network based fuzzy inference system (ANFIS) as a tool for system identification with special emphasis on training data minimization | |
Robati et al. | Inflation rate modeling: adaptive neuro-fuzzy inference system approach and particle swarm optimization algorithm (ANFIS-PSO) | |
CN109408896B (en) | Multi-element intelligent real-time monitoring method for anaerobic sewage treatment gas production | |
Rizal et al. | Sequence to sequence analysis with long short term memory for tourist arrivals prediction | |
Wu et al. | Automatic fuzzy model identification for short-term load forecast | |
Zhang et al. | A network traffic prediction model based on quantum inspired PSO and neural network | |
Borysenko et al. | Intelligent forecasting in multi-criteria decision-making. | |
CN114707613B (en) | Layered depth strategy gradient network-based power grid regulation and control method | |
Cheng et al. | Prediction of Tourist Flow Based on Deep Belief Network and Echo State Network. | |
Agustyawan et al. | Combination of Backpropagation Neural Network and Particle Swarm Optimization for Prediction of Water Production in Municipal Waterworks | |
Petrovic‐Lazarevic et al. | Neuro‐fuzzy support of knowledge management in social regulation | |
Banik et al. | Modeling chaotic behavior of Dhaka stock market index values using the neuro-fuzzy model |
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 |