CN108764540B - Water supply network pressure prediction method based on parallel LSTM series DNN - Google Patents

Water supply network pressure prediction method based on parallel LSTM series DNN Download PDF

Info

Publication number
CN108764540B
CN108764540B CN201810465536.9A CN201810465536A CN108764540B CN 108764540 B CN108764540 B CN 108764540B CN 201810465536 A CN201810465536 A CN 201810465536A CN 108764540 B CN108764540 B CN 108764540B
Authority
CN
China
Prior art keywords
model
dnn
water supply
pressure
data
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.)
Active
Application number
CN201810465536.9A
Other languages
Chinese (zh)
Other versions
CN108764540A (en
Inventor
徐哲
李玉全
陈晖�
何必仕
陈云
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Hangzhou Dianzi University
Original Assignee
Hangzhou Dianzi University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Hangzhou Dianzi University filed Critical Hangzhou Dianzi University
Priority to CN201810465536.9A priority Critical patent/CN108764540B/en
Publication of CN108764540A publication Critical patent/CN108764540A/en
Application granted granted Critical
Publication of CN108764540B publication Critical patent/CN108764540B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

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—Systems or methods specially adapted for specific business sectors, e.g. utilities or tourism
    • G06Q50/06—Electricity, gas 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

Abstract

The invention discloses a water supply network pressure prediction method based on parallel LSTM series DNN. The method firstly determines input and output items and establishes a prediction model based on parallel LSTM series DNN. Secondly, preprocessing the data and establishing a pressure prediction database. The predictive model is then trained. And finally, performing online pressure prediction. The method realizes the advantage complementation of the LSTM and the DNN, prevents the overfitting of the model by using a Dropout technology, accelerates the convergence speed of the model by using a Relu activation function, reduces the randomness and the calculated amount by using a small-batch gradient descent method, selects the RMSprop as an optimization algorithm of the random gradient descent method, and improves the anti-interference performance and the precision of the water supply network pressure prediction method.

Description

Water supply network pressure prediction method based on parallel LSTM series DNN
Technical Field
The invention belongs to the field of urban water supply, and particularly relates to a water supply network pressure prediction method based on parallel LSTM series DNN.
Background
The water supply pipe network system is a nonlinear dynamic system with a complex structure, a large scale and strong water use randomness, can quickly and accurately simulate and predict the operation condition of an outlet pipe network, and is the key for optimizing and scheduling the water supply pipe network. During production and operation of the water supply network, a dispatcher generally observes the operation condition of the pipe network according to pressure measured data. Therefore, the pressure of the monitoring point is subjected to predictive analysis, so that the scheduling personnel can predict in advance and make production command.
At present, water supply network pressure prediction methods are generally divided into three categories, namely a time series method, a structural analysis method and a systematic method. Wherein the time series method comprises a moving average method, an exponential smoothing method, a trend extrapolation method and the like; structural analysis methods include regression analysis methods and the like; the system method comprises grey prediction, artificial neural network and the like. The time sequence model has better prediction precision and simple data processing, but only can utilize short-term time sequence data of the measuring points, and is not suitable for the condition with larger change; the regression analysis method model is simple and convenient, but is difficult to select when being comprehensively influenced by various factors; the system method has the advantages of self-learning capability, nonlinear processing and the like, but has the problems of complex model, long training time and the like. The methods are all easy to be interfered by noise, and the prediction precision is difficult to ensure.
Disclosure of Invention
Aiming at the highly complex nonlinear characteristic of a water supply network and the defects of the prior art method, the invention provides a deep learning model based on parallel LSTM (long short term memory neural network) and series DNN (deep neural network) for predicting the pressure of the water supply network, and the prediction precision is improved.
Because the state quantity and the control quantity of the water supply network are two different types of characteristic information, if one LSTM model is simply adopted for characteristic extraction, the different influences of the two types of characteristic information on the model cannot be highlighted, therefore, the single LSTM model is expanded into a parallel LSTM model, and the two types of different characteristic information are respectively extracted and learned. And because the LSTM is good at processing data based on time series, the DNN is suitable for mapping the characteristic information to a higher space, the LSTM and the DNN are combined to be used as a unified framework by utilizing the respective advantages of the LSTM and the DNN, so that advantage complementation is realized, namely, the output results of the two paths of LSTMs are fused and then output through the DNN, and the pressure of a measuring point at the next moment is predicted. Therefore, the invention provides a water supply network pressure prediction method based on a parallel LSTM series DNN deep learning model.
In order to realize the purposes of high anti-interference performance and high prediction precision, the invention adopts the following steps:
1. determining input and output items, and establishing a prediction model based on parallel LSTM series DNN
Considering that the water supply network system is a multi-input multi-output nonlinear time-lag system, longer historical state quantity (measuring point pressure information) [ x (t), x (t-1), …, x (t-n) is selecteds)]And control amounts (inlet pressure and flow rate) [ u (t-1), u (t-2), …, u (t-n)u)]As an input item to compensate for the deficiency of "the water supply network status quantity is only partially known"; determining the output item as the output y at the moment of the pressure measuring point t +1m(t + 1). Here, n iss、nuIs a historical time window.
Establishing a deep learning model based on parallel LSTM series DNN:
a. respectively adopting LSTM model to make state quantities [ x (t), x (t-1), … and x (t-n)s)]And control amounts [ u (t-1), u (t-2), …, u (t-n)u)]And (5) carrying out feature extraction and learning.
Figure BDA0001661957570000021
Respectively are the output values of the state variable and the control variable of the water supply network through an LSTM model.
b. The deep neural network DNN model is adopted
Figure BDA0001661957570000022
Performing fusion processing to obtain output ym(t +1), which can be described by formula (1):
Figure BDA0001661957570000023
ym(t +1) is the prediction output of the deep learning model. []Which represents the merging of two moments having the same dimension in the time dimension, H () is the activation function of the DNN model. WDNN、bDNNThe weights and thresholds of the DNN model are respectively.
2. Preprocessing data, creating pressure prediction database
(1) Data pre-processing
Data filling: aiming at the problem of data loss of data acquired from an SCADA system on site, linear, parabolic or cubic curve interpolation is adopted to complement the missing data. If the missing data is excessive, the historical data of the period is abandoned.
Denoising data: aiming at the problem of large amount of noise interference in field data, wavelet transformation is adopted to remove noise. The wavelet transformation transducer carries out local signal analysis in time domain and frequency domain, can realize multi-resolution analysis, judges noise and mutation signals and determines effective signals.
Dimensionless processing: aiming at the problem that the pressure and the flow of a water supply network have different physical dimensions and magnitude, data are normalized, namely the input and the output are limited to [0, l ], so that the input and the output participate in model training and prediction in the same level, and the specific formula is shown as (2)
Figure BDA0001661957570000031
Wherein X represents the data to be normalized, Max (X), Min (X) represent the minimum value and the maximum value respectively, and XnorRepresents the normalized data.
(2) Building a pressure prediction database
Establishing a water supply network pressure prediction database: the data items include, in addition to the time stamp, the node (watch point or entry): (1) pressure and flow values of the measuring points, pressure and flow values of the inlet and the like are extracted/cleaned/converted from the SCADA in real time and stored as input items of the model; (2) the predicted pressure of the measuring point is predicted from the model and is an output item of the model; (3) and error data items are used for statistically analyzing the prediction precision.
3. Training a predictive model
(1) Determining training samples
DMA partitioning around a large water supply network, or a small water supply network, determines the input sample as { X (n)s),U(nu) Y, where X (n)s) Is an i-dimensional state variable, U (n)u) Is 2j dimension control variable, i is number of monitoring points, j is number of entries.
To guarantee training, the time span of the input sample data must guarantee Max (n)s,nu) The time interval is continuous, and continuous effective { X (n) of more than 1 hour is ensureds),U(nu) Y data, valid samples are no less than 12x24x15 ═ 4320.
(2) Determining basic structure of model, setting initial values of other parameters, and training model
And determining the value range of the parameter according to experience or the effect of preliminary parameter adjustment. n isu、nsE {1,2, … 12}, the time step t is 5 minutes, i.e. the maximum span of the history information is 60 minutes, the longer history information will cause the inputRedundancy is added, and the improvement of the prediction precision is not influenced greatly; the number of hidden Layers is in the range of {1,2, …,5}, the number of Layers of hidden Layers can be increased to improve the feature extraction and learning ability, but the model becomes more and more complex due to multiple Layers; the corresponding neuron number Neurons E [0,300 ∈ ]]The number of neurons determines the degree of non-linearity of the network training.
In order to prevent the deep learning model from generating an overfitting phenomenon, the invention randomly updates network parameters after Dropout is introduced into each layer, and the generalization capability of the model is increased. The Dropout technique specifically discards a certain proportion of hidden layer nodes at random (but weights are preserved and are not updated only temporarily) during model training, and restores full connectivity when the model is used. And a node discarding proportion dropout rate is in an element of [01,0.5], and for the selection of the node discarding proportion, if the proportion is too low, the effect cannot be achieved, and if the proportion is too high, the model can be subjected to under-learning.
The invention adopts a Mini-batch gradient descent method to optimize each parameter in the model, the method divides data into a plurality of batches, and the parameters are updated according to the batches (batch), thus, a group of data in one batch jointly determines the update of the parameters, and the randomness and the calculated amount are reduced. The sample size Mini _ batch ∈ [5,50] for the short run gradient descent method. The training round number epoch belongs to [100,200], the training frequency is too small to achieve the training effect, and too much training time is increased instead of improving the prediction precision.
The activation function is used in two parts of the model, one for computing the LSTM layer for the input values and the other for the output layer after feature fusion. The gradient vanishing problem can be brought by traditional saturation activation functions such as sigmoid and tanh, and the convergence speed of the model can be increased by unsaturated activation functions such as ReLU and the like relative to the saturation activation functions; the depth model using ReLU does not require pre-training before supervised training to achieve similar or even better results. The activation function of the present invention selects ReLU.
(3) Training iterations
During model training, the prediction value of the model is generally used
Figure BDA0001661957570000041
And the root mean square is obtained as a model error from the measured value y, as shown in equation (3). Wherein n is the number of neuron nodes in the output layer.
Figure BDA0001661957570000042
When loss<Error target epsilon [ 0.2%, 0.5%]And the training requirement is met, and the iteration is finished. Adjusting each parameter of the model when the error is larger, if the error does not meet the convergence condition and is not reduced any more, changing the basic structure of the model, namely, re-setting one { n }u,nsAnd Layers, adjusting other parameters according to each basic structure, and repeating the iterative training.
4. Online pressure prediction
Sequentially inputting continuous effective data such as pressure and flow values of measuring points in a pressure prediction database, pressure and flow values of inlets and the like into a model, and giving a pressure prediction value y at the moment of t +1 by the modelm(t +1), may be provided to the dispatcher for reference by about 5 minutes in advance.
At the same time, the pressure is predicted to be ym(t +1) is stored in a database, and is compared with the measured value y (t +1) at the time of t +1, and Δ ═ y is calculatedm(t +1) -y (t + 1). Let the allowable prediction error be σ ∈ [ 5%, 10%]If Δ is three times in succession>σ x y (t +1), then return to step 3, retrain the model, update the model parameters with the recent data.
The invention has the beneficial effects that: the invention provides a deep learning model based on parallel LSTM and DNN in series, which realizes advantage complementation of LSTM and DNN, prevents overfitting of the model by using a Dropout technology, accelerates the convergence speed of the model by using a Relu activation function, reduces randomness and calculated amount by using a small-batch gradient descent method, selects RMSprop as an optimization algorithm of the random gradient descent method, and improves the anti-interference performance and the precision of a water supply network pressure prediction method.
Drawings
FIG. 1: a prediction model based on parallel LSTM tandem DNNs;
FIG. 2: a water supply network monitoring point pressure prediction system application framework.
Detailed Description
In order to make the technical means and the creative features of the implementation of the present invention easily apparent, the implementation of the present invention is further described in detail below with reference to the accompanying drawings and the embodiments, and the scope of the right of the present invention is not limited.
Taking a water supply network in a Y district of a city as an example, the area of the area is about 106.7km2The daily water supply amount is about 150000m3The monitoring points comprise a water inlet flowmeter pressure measuring point, a pipe network pressure measuring point, a water outlet flowmeter point and a middle pipeline measuring point.
The information of the important pressure monitoring points in the pipe network is shown in the table 1
TABLE 1 pipe network 17 important pressure monitoring point information
Figure BDA0001661957570000051
1. Determining input and output items, and establishing a prediction model based on parallel LSTM series DNN
Considering that the water supply network system is a multi-input multi-output nonlinear time-lag system, longer historical state quantities [ x (t), x (t-1), …, x (t-n) are selecteds)]And control amounts [ u (t-1), u (t-2), …, u (t-n)u)]As an input item to compensate for the deficiency of "the water supply network status quantity is only partially known"; determining the output item as the output y at the moment of the pressure measuring point t +1m(t + 1). Wherein n iss、nuIs a historical time window.
Here, the water supply pressure at 4 water inlets, the historical control information of the water supply amount, the historical pressure history of 17 measuring points and the current state information of the city Y district water supply pipe network system are used as input items, and the pressure at the next moment of 17 monitoring points is used as an output item.
Establishing a depth learning model based on parallel LSTM tandem DNN (as shown in FIG. 1):
a. respectively adopting LSTM model to make state quantities [ x (t), x (t-1), … and x (t-n)s)]And control amounts [ u (t-1), u (t-2), …, u (t-n)u)]And (5) carrying out feature extraction and learning.
Figure BDA0001661957570000052
Respectively are the output values of the state variable and the control variable of the water supply network through an LSTM model.
b. The deep neural network DNN model is adopted
Figure BDA0001661957570000053
Performing fusion processing to obtain output ym(t +1), which can be described by formula (4):
Figure BDA0001661957570000054
ym(t +1) is the prediction output of the deep learning model. []Which represents the merging of two moments having the same dimension in the time dimension, H () is the activation function of the DNN model. WDNN、bDNNThe weights and thresholds of the DNN model are respectively.
2. Preprocessing data, creating pressure prediction database
(1) Data pre-processing
Data filling: aiming at the problem of data loss of data acquired from an SCADA system on site, linear, parabolic or cubic curve interpolation is adopted to complement the missing data. If the missing data is excessive, the historical data of the period is abandoned.
Denoising data: aiming at the problem of large amount of noise interference in field data, wavelet transformation is adopted to remove noise. The wavelet transformation transducer carries out local signal analysis in time domain and frequency domain, can realize multi-resolution analysis, judges noise and mutation signals and determines effective signals.
Dimensionless processing: aiming at the problem that the pressure and the flow of a water supply network have different physical dimensions and magnitude, data are normalized, namely the input and the output are limited to [0, l ], so that the input and the output participate in model training and prediction in the same level, and the specific formula is shown as (5)
Figure BDA0001661957570000061
Wherein X represents the data to be normalized, Max (X), Min (X) represent the minimum value and the maximum value respectively, and XnorRepresents the normalized data.
(2) Building a pressure prediction database
Establishing a water supply network pressure prediction database: the data items include, in addition to the time stamp, the node (watch point or entry): (1) pressure and flow values of the measuring points, pressure and flow values of the inlet and the like are extracted/cleaned/converted from the SCADA in real time and stored as input items of the model; (2) the predicted pressure of the measuring point is predicted from the model and is an output item of the model; (3) and error data items are used for statistically analyzing the prediction precision.
3. Training a predictive model
(1) Determining training samples
Surrounding a water supply network in a city Y region, determining the input sample as X (n)s),U(nu) Y, where X (n)s) Is a 17-dimensional state variable, U (n)u) For 8-dimensional control variables, 17 monitoring points, 4 entries, and t is the time step.
To guarantee training, the time span of the input sample data must guarantee Max (n)s,nu) The time interval is continuous, and continuous effective { X (n) of more than 1 hour is ensureds),U(nu) Y data, valid samples are not less than 12x24x15 ═ 4320
Here, the sample data set is data from 2016, 27 th month to 2016, 28 th month, 6 th month, 35 days, the sampling interval is 5 minutes, and the sample size is 10080. With data from 5 months, 28 days, 6 months to 26 days used for training.
(2) Determining basic structure of model, setting initial values of other parameters, and training model
And determining the value range of the parameter according to experience or the effect of preliminary parameter adjustment. n isu、nsE {1,2, … 12}, wherein the time step t is 5 minutes, that is, the maximum span of the historical information is 60 minutes, because longer historical information causes input redundancy and has little influence on improving the prediction accuracy; the number of hidden layers of LSTM is 2, the number of Neurons is 100, and the number of DNN layers is 1002, corresponding Neurons 96.
In order to prevent the deep learning model from generating an overfitting phenomenon, the invention randomly updates network parameters after Dropout is introduced into each layer, and the generalization capability of the model is increased. In this embodiment, the regularized node discard ratio droout rate is 0.3.
In this embodiment, each parameter in the model is optimized by a Mini-batch (Mini-batch) gradient descent method, where the sample size Mini _ batch of the Mini-batch gradient descent method is 32, and the training round number epoch is 100.
The activation function of this embodiment selects ReLU.
(3) Training iterations
During model training, the prediction value of the model is generally used
Figure BDA0001661957570000071
And the measured value y, the root mean square is obtained as a model error, as shown in equation (6). Wherein n is the number of neuron nodes in the output layer.
Figure BDA0001661957570000072
When loss<And the error target epsilon is 0.3 percent, the training requirement is met, and the iteration is finished. Adjusting each parameter of the model when the error is larger, if the error does not meet the convergence condition and is not reduced any more, changing the basic structure of the model, namely, re-setting one { n }u,nsAnd Layers, adjusting other parameters according to each basic structure, and repeating the iterative training.
4. Online pressure prediction
The pressure and flow values of the measuring points in the pressure prediction database, the pressure and flow values of the inlet and other continuous effective data are sequentially input into the model, and the model gives a pressure prediction value y at the moment of t +1m(t +1), may be provided to the dispatcher for reference by about 5 minutes in advance.
At the same time, the pressure is predicted to be ym(t +1) is stored in a database, and is compared with the measured value y (t +1) at the time of t +1, and Δ ═ y is calculatedm(t +1) -y (t + 1). Let the allowable prediction error be σ5%, if three times in succession>σ x y (t +1), model retraining is initiated (step 3), with the recent data, updating the model parameters.
In the embodiment, data from 27 days at 6 months to 28 days at 6 months are used for testing, the RMSE (root mean square error) and the MAPE (mean absolute percentage error) are used as evaluation performance indexes, and the performance indexes of each monitoring point are given in table 2, so that the prediction precision is high.
TABLE 2 predicted Performance indices for each monitoring Point
Figure BDA0001661957570000073
Figure BDA0001661957570000081
For comparison with the conventional model, the BP neural network, the SVM support vector machine, VARX and NARX were calculated respectively, and the comparison result with the conventional prediction method was obtained (see table 3).
TABLE 3 comparison with conventional prediction methods
Prediction method BP SVM VARX NARX The method of the invention
RMSE 100 (average value of each measuring point) 0.56 0.43 0.30 0.28 0.17
From table 3, the prediction results based on the parallel LSTM serial DNN deep learning model are significantly better than the conventional prediction model.
The method can be used for actual production according to the application framework of the pressure prediction system of the monitoring point of the water supply network shown in FIG. 2.

Claims (7)

1. The water supply network pressure prediction method based on parallel LSTM series DNN is characterized by comprising the following steps:
determining input and output items, and establishing a prediction model based on parallel LSTM series DNN, specifically:
considering that the water supply network system is a multi-input multi-output nonlinear time-lag system, the historical state quantity [ x (t), x (t-1), …, x (t-n) is selecteds)]And control amounts [ u (t-1), u (t-2), …, u (t-n)u)]As an input item; determining the output item as the output y at the moment of the pressure measuring point t +1m(t +1), where ns、nuIs a historical time window;
establishing a deep learning model based on parallel LSTM series DNN:
a. respectively adopting LSTM model to make state quantities [ x (t), x (t-1), … and x (t-n)s)]And control amounts [ u (t-1), u (t-2), …, u (t-n)u)]Performing feature extraction and learning while setting
Figure FDA0003226574610000011
Respectively outputting the state quantity and the control quantity of the water supply pipe network through an LSTM model;
b. the deep neural network DNN model is adopted
Figure FDA0003226574610000012
Performing fusion processing to obtain output ym(t+1)
Figure FDA0003226574610000013
Wherein]Represents the merging of two moments with the same dimension in the time dimension, H () being the activation function of the DNN model; wDNN、bDNNRespectively are the weight value and the threshold value of the DNN model;
step (2) data preprocessing is carried out, and a pressure prediction database is established
(2-1) data preprocessing
Data filling: aiming at the problem of data loss of data acquired from an SCADA system on site, linear, parabolic or cubic curve interpolation is adopted to complement the missing data;
denoising data: aiming at the problem of large noise interference of field data, wavelet transformation is adopted to remove noise;
dimensionless processing: aiming at the problems that the pressure and the flow of a water supply network have different physical dimensions and magnitude, data are normalized, namely the input and the output are limited to [0, l ], so that the input and the output participate in model training and prediction in the same level;
(2-2) building a pressure prediction database
The data items include, in addition to the time stamp and the node: (1) the pressure and flow values of the measuring points and the pressure and flow values of the inlet are extracted, cleaned, converted and stored in real time from the SCADA system and are used as input items of the model; (2) the predicted pressure of the measuring point is predicted from the model and is an output item of the model; (3) error data item for statistical analysis of prediction accuracy;
step (3) training prediction model
(3-1) determining training samples
DMA partitioning around a Large Water supply network or Small Water supply network, determining the input sample as { X (n)s),U(nu) Y, where X (n)s) Is an i-dimensional state quantity, U (n)u) Is 2j dimension control quantity, i is monitoring point number, j is entranceCounting;
(3-2) determining the basic structure of the model, setting the initial values of the other parameters, and starting to train the model
Determining the value range of the parameter according to experience or the effect of preliminary parameter adjustment; n isu、nsE, e {1,2,. 12}, wherein the time step t is 5 minutes, namely the maximum span of the historical information is 60 minutes; the number of hidden Layers is Layers ∈ {1, 2.., 5 }; the corresponding neuron number Neurons E [0,300 ∈ ]]；
(3-3) training iterations
Model prediction value in model training
Figure FDA0003226574610000021
And the measured value y is used to obtain the root mean square as the model error, when the error is<Error target epsilon [ 0.2%, 0.5%]When the training requirement is met, the iteration is finished; when the error does not meet the requirement, adjusting each parameter of the model, if the error does not meet the convergence condition and is not reduced, changing the basic structure of the model, namely, resetting one { n }u,nsLayers, then adjusting other parameters according to each basic structure, and repeating iterative training;
step (4) on-line pressure prediction
Pressure and flow values of measuring points in the pressure prediction database and pressure and flow values of inlets are sequentially input into the model, and the model gives a pressure prediction value y at the moment of t +1m(t + 1); at the same time, the pressure is predicted to be ym(t +1) is stored in a database, and is compared with the measured value y (t +1) at the time of t +1, and Δ ═ y is calculatedm(t +1) -y (t + 1); let the allowable prediction error be σ ∈ [ 5%, 10%]If Δ is three times in succession>And sigma x y (t +1), returning to the step (3), retraining the model, and updating the model parameters by using the recent data.
2. The parallel LSTM tandem DNN based water supply network pressure prediction method of claim 1, wherein: in the step (3-2), in order to prevent the overfitting phenomenon of the deep learning model, network parameters are randomly updated after a Dropout technology is introduced into each layer, and the generalization capability of the model is increased.
3. The water supply network pressure prediction method based on parallel LSTM tandem DNN of claim 2 wherein: the Dropout technology specifically comprises the steps that a certain proportion of hidden layer nodes are randomly abandoned during model training, but the weights are saved, only temporary updating is not carried out, and full connection is recovered when the model is used; the node discard ratio is between 0.1 and 0.5.
4. The parallel LSTM tandem DNN based water supply network pressure prediction method of claim 1, wherein: and (3) optimizing each parameter in the model by adopting a small batch gradient descent method in the step (3-2).
5. The method of water supply network pressure prediction based on parallel LSTM tandem DNN of claim 4 wherein: the small batch gradient descent method divides data into a plurality of batches and updates parameters according to the batches; the sample size of the small batch gradient descent method is between 5 and 50; the number of training rounds is between 100 and 200.
6. The parallel LSTM tandem DNN based water supply network pressure prediction method of claim 1, wherein: and (4) selecting a ReLU activation function as the activation function of the model in the step (3-2).
7. The parallel LSTM tandem DNN based water supply network pressure prediction method of claim 1, wherein: predicting the pressure y in the step (4)m(t +1) is provided 5 minutes ahead of time to the dispatcher for reference.
CN201810465536.9A 2018-05-16 2018-05-16 Water supply network pressure prediction method based on parallel LSTM series DNN Active CN108764540B (en)

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 CN108764540A (en) 2018-11-06
CN108764540B true 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)

Families Citing this family (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109930658B (en) * 2019-03-27 2021-02-05 杭州电子科技大学 Water supply network monitoring point arrangement method based on system visibility
CN110545208B (en) * 2019-09-23 2021-10-15 电子科技大学 Network traffic prediction method based on LSTM
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
CN111210093B (en) * 2020-03-05 2023-05-09 重庆森鑫炬科技有限公司 Daily water consumption prediction method based on big data
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
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
CN112561174B (en) * 2020-12-18 2023-05-02 西南交通大学 LSTM and MLP-based superimposed neural network prediction geothermal energy production method
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
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
CN114358244B (en) * 2021-12-20 2023-02-07 淮阴工学院 Big data intelligent detection system of pressure based on thing networking

Family Cites Families (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103258235B (en) * 2013-05-13 2016-01-27 杭州电子科技大学 A kind of water supply network reorganization and expansion optimization method based on improving artificial fish-swarm algorithm
US10783900B2 (en) * 2014-10-03 2020-09-22 Google Llc 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
KR102033411B1 (en) * 2016-08-12 2019-10-17 한국전자통신연구원 Apparatus and Method for Recognizing speech By Using Attention-based Context-Dependent Acoustic Model

Also Published As

Publication number Publication date
CN108764540A (en) 2018-11-06

Similar Documents

Publication Publication Date Title
CN108764540B (en) Water supply network pressure prediction method based on parallel LSTM series DNN
CN108900346B (en) Wireless network flow prediction method based on LSTM network
CN111563706A (en) Multivariable logistics freight volume prediction method based on LSTM network
Lu et al. Generalized radial basis function neural network based on an improved dynamic particle swarm optimization and AdaBoost algorithm
CN106022521B (en) Short-term load prediction method of distributed BP neural network based on Hadoop architecture
CN102622418B (en) Prediction device and equipment based on BP (Back Propagation) nerve network
CN113361777B (en) Runoff prediction method and system based on VMD decomposition and IHHO optimization LSTM
Yu et al. Error correction method based on data transformational GM (1, 1) and application on tax forecasting
CN112364975A (en) Terminal operation state prediction method and system based on graph neural network
CN112288193A (en) Ocean station surface salinity prediction method based on GRU deep learning of attention mechanism
CN111310965A (en) Aircraft track prediction method based on LSTM network
CN111860787A (en) Short-term prediction method and device for coupling directed graph structure flow data containing missing data
CN112257847A (en) Method for predicting geomagnetic Kp index based on CNN and LSTM
CN112163671A (en) New energy scene generation method and system
CN114548591A (en) Time sequence data prediction method and system based on hybrid deep learning model and Stacking
CN113095484A (en) Stock price prediction method based on LSTM neural network
CN109408896B (en) Multi-element intelligent real-time monitoring method for anaerobic sewage treatment gas production
CN113052373A (en) Monthly runoff change trend prediction method based on improved ELM model
Sulistiyo et al. Evolution strategies for weight optimization of Artificial Neural Network in time series prediction
Bucur et al. An adaptive fuzzy neural network for traffic prediction
CN111414927A (en) Method for evaluating seawater quality
CN113095489A (en) Inductive multi-level graph network representation learning method
CN110829434B (en) Method for improving expansibility of deep neural network tidal current model
CN112990618A (en) Prediction method based on machine learning method in industrial Internet of things
Hu et al. Application on crude oil output forecasting based on gray neural network

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