CN108051035A - The pipe network model recognition methods of neural network model based on gating cycle unit - Google Patents

Abstract

The invention provides a neural network model based on a gated cycle unit and its training method and application. Wherein, the neural network model based on the gated recurrent unit includes: multiple gated recurrent unit layers, which are configured to receive the flow characteristics of the water supply network in different time periods; the first fully connected layer, which is connected with the multiple The gated recurrent unit layer is connected in parallel to form an input layer, and the first fully connected layer is configured to receive the meteorological characteristics of the water supply pipe network; a merge layer connected with the input layer in a tensor series mode; both connected with the merge layer The second fully connected layer to the Mth fully connected layer connected in tensor series mode, M is an integer greater than or equal to 2; and the output layer connected with the Mth fully connected layer in tensor series mode, which Configured to output the predicted flow of the water supply network at the next moment.

Description

Pipe Network Leakage Identification Method Based on Gated Recurrent Unit Neural Network Model

technical field

The invention relates to the field of neural networks, in particular to a neural network model based on a gated recurrent unit and its training method and application.

Background technique

Under the influence of factors such as aging pipelines, limited technical investment, and backward regulatory systems, the leakage of urban water supply networks in my country is relatively common. Leakage of water supply pipe network not only brings huge economic losses to water supply companies, but also causes waste of energy and resources. Most of the water supply companies in my country lack scientific and effective management methods for the water supply pipe network, and the basic data of the pipe network are not perfect. These problems have always restricted the management efficiency and service level of the water supply company. Therefore, it is of great economic and practical significance to identify the leakage of the pipeline network to be tested, assist the manager to make scientific decisions, and discover and maintain the leakage area of the pipeline network in time.

The common models used to identify pipeline network leakage are mainly divided into two categories: one is the physical mechanism research model based on laboratory experiments; the other is the data mining model based on computer simulation, including statistical models, probability models, and machine models. Learning models, etc., usually requires a large amount of historical data. With the advent of the cloud era, big data has attracted more and more attention, and has achieved good results in many fields. Therefore, the second type of model has become a current research hotspot, especially the machine learning model.

my country's urban water supply network has a large amount of data, and the quality of the basic data is not high. There are many factors and complex relationships that lead to network leakage. Most of the existing pipeline network leakage identification models are based on some classic traditional algorithms to predict the normal flow or pressure of the pipeline network, such as multiple linear regression, exponential smoothing method, back propagation (Back Propagation, BP) neural network, etc. , and then identify leakage accidents by comparing predicted values with actual monitored values. These traditional forecasting methods have limited ability to deal with a large amount of complex nonlinear data, and cannot well dig out the hidden features between abstract data, and there is a problem of low prediction accuracy, which also makes the accuracy of leakage accident identification There is no guarantee.

Contents of the invention

In order to save water resources, reduce the leakage rate of the water supply network, reduce economic losses, and conform to the concept of sustainable development, we need to use existing technologies to dig out the hidden characteristics of the water supply network data, and establish high precision and good stability. model to identify leakage areas in the pipeline network. Aiming at the deficiencies of the prior art, the present invention aims to provide a new neural network model based on gated recurrent units and its training method and application. By constructing a new neural network different from the traditional neural network model, a large amount of data can be processed , excavate the nonlinear relationship, and comprehensively evaluate the model effect, and apply it to the leakage accident identification of the water supply network, which can improve the accuracy of leakage identification, so that daily managers can find the leakage area of the pipeline network in time and reduce economic losses , save water resources, and assist water companies to make scientific and reasonable decisions.

One aspect of the present invention provides a neural network model based on gated cyclic units for identifying leakage accidents in water supply pipe networks, including:

A plurality of gated recurrent unit (Gated Recurrent Unit, GRU for short) layers, which are configured to receive flow characteristics of the water supply network in different time periods;

A first fully connected layer, which is connected in parallel with the multiple gated recurrent unit layers to form an input layer, and the first fully connected layer is configured to receive meteorological features of the water supply network;

a pooling layer connected in tensor concatenation with said input layer;

The second fully-connected layer to the Mth fully-connected layer, all of which are connected to the merging layer in tensor series, where M is an integer greater than or equal to 2; and

An output layer connected to the Mth fully connected layer in tensor series, configured to output the predicted flow rate of the water supply network at the next moment.

The Gated Recurrent Unit Network (GRUN for short) provided by the present invention is an improvement to the existing recurrent neural network. The characteristic of this kind of network is to use the memory module to replace the ordinary hidden nodes, to ensure that the gradient will not disappear or explode after passing through many time steps, so as to overcome the difficulties encountered in the traditional cycle neural network training. The memory module used in the present invention is a gated recurrent unit (GRU). FIG. 1 shows a topology diagram of a neural network model based on gated recurrent units in an embodiment of the present invention.

Another aspect of the present invention provides a training method for the aforementioned neural network model, including:

Construct the neural network model;

Randomly divide the historical flow characteristics and historical meteorological characteristics of different time periods into a training set and a verification set;

Input the historical traffic characteristics of different time periods in the training set into the multiple gated recurrent unit layers, and input the historical meteorological characteristics in the training set into the first fully connected layer, and obtain the neural The training traffic output by the network model;

Input the historical traffic characteristics of different time periods in the verification set into the multiple gated recurrent unit layers, and input the historical meteorological characteristics in the verification set into the first fully connected layer to obtain the neural Verification traffic output by the network model;

Generate a training curve based on the training set and training traffic, and generate a verification curve based on the verification set and verification traffic, when the mean square error value of the training curve and verification curve is stable at a constant value, complete the neural network model training.

This training method fully considers the randomness of the gradient descent direction and the generalization ability of the model during the training process, improves the convergence speed and training efficiency of the model, and fully mines the effective information between the data.

In a preferred embodiment of the present invention, the above-mentioned training method also includes acquiring the historical flow characteristics of the different time periods, including:

Obtain historical flow data of the water supply network to be tested; and

Extraction processing and normalization processing are performed on the historical traffic data to obtain the historical traffic features.

According to the present invention, the historical flow data may be obtained from historical flow data of independent metered areas (districtmetered areas, DMA for short) water supply pipe network of the water company.

In a more preferred embodiment of the present invention, extracting the historical flow data includes, in the historical flow data of the water supply pipeline network, respectively starting from the first time period, the second time period, and the third time period Extract the specified number of traffic data at equal intervals in the segment, where,

Select the first time period before the time t to be predicted,

selecting a second time period before the starting moment of the first time period,

A third time period is selected before the starting moment of the second time period.

According to the present invention, considering the trend of the flow time series, the flow data near the time t to be predicted is extracted, that is, the flow data in the first time period. Considering the periodicity of the flow time series, extracting the flow data separated by a period of time from the time t to be predicted can be divided into two time periods by day, namely the flow data of the second time period and the third time period.

In a specific embodiment of the present invention, during the entire time period including the first time period, the second time period and the third time period, average sampling is performed at a sampling frequency of k times/h, then the whole time period includes The flow data of t-24k·d, t-24k·d+1...t-2, t-1 time flow data, where d represents the number of days, from which to obtain the flow data from tm to t-1 time, where m The value range is an integer of 2-24, preferably 5, as the flow data of the first time period; obtain the flow data of th ₁ to th ₂ moments, wherein the value range of h ₁ and h ₂ is 24k-2 to An integer of 24k+20, preferably, h ₁ is 24k+2, h ₂ is 24k-2; obtain the flow data from th ₃ to th ₄ , where the values of h ₃ and h ₄ range from 48k-2 to 48k An integer of +20, preferably, h ₃ is 48k+2, and h ₄ is 48k-2.

Preferably, k is an integer of 1-12, preferably 3-5.

In another preferred embodiment of the present invention, the above-mentioned training method also includes obtaining the historical meteorological features, including:

Obtain historical weather data of the water supply network to be tested; and

Extraction processing and normalization processing are performed on the historical meteorological data to obtain the historical meteorological features.

According to the present invention, the historical meteorological data can be obtained from the historical meteorological data of China Meteorological Data Network, including maximum temperature, minimum temperature, relative humidity, precipitation and the like.

In a more preferred embodiment of the present invention, extracting the historical meteorological data includes performing Pearson correlation analysis on the historical meteorological data and the historical flow data, and selecting the correlation coefficient greater than or equal to the specified value V The historical meteorological data is obtained by extracting and processing the historical meteorological data.

In a more preferred embodiment of the present invention, the specified value V is 0.80-0.99.

According to the present invention, the normalization process includes converting the numerical value of the original data into the range of [0,1], and the formula of the normalization process is as shown in formula (1),

In formula (1), y represents the normalized data, x represents the input original data, x _max and x _min represent the maximum value and minimum value of the input data respectively.

According to the present invention, the activation function of each network layer is selected from tanh, ReLU or Linear.

Another aspect of the present invention provides a method for identifying leakage accidents of a water supply pipe network, including:

The feature acquisition step is to acquire the reference flow characteristics and reference meteorological characteristics in different time periods of the water supply network to be tested within the specified time T when no leakage accident occurs;

The forecast flow acquisition step is to input the reference flow characteristics of the different time periods into the multiple gated recurrent unit layers of the neural network model, and input the reference meteorological characteristics into the first full network of the neural network model In the connection layer, the predicted traffic output by the neural network model is obtained;

Refer to the step of determining the data set to obtain the measured flow within the specified time T, each group of predicted flow and measured flow at the same time constitutes a time series vector, and a matrix composed of all time series vectors within the specified time T as a reference dataset;

Accident identification step, calculating the cosine distance of the newly added time series vector relative to each time series vector in the reference data set, counting the number n of time series vectors whose cosine distance is less than the distance threshold D, when n is less than or equal to the number When the threshold N is reached, it is determined that the leakage accident has occurred.

According to the present invention, the newly added time series vector is a time series vector obtained according to the feature acquisition step, forecast flow acquisition step, and reference data value determination step during a time period other than the specified time T.

According to the present invention, formula (2) is used to calculate the cosine distance of the newly added time series vector relative to each time series vector in the reference data set.

In formula (2), i is the new time series vector, j is the time series vector in the reference data set, and ‖i‖ and ‖j‖ are the modulus of the vector.

According to the present invention, the acquisition method of the reference flow characteristics and the reference meteorological characteristics in different time periods of the water supply network to be tested is consistent with the acquisition method of the historical flow characteristics and historical meteorological characteristics in the above training method.

According to the present invention, in the above identification method, the time T is set to be 2 days, i.e. 48 hours, and the data collection frequency is set to be k times/h (that is, k time series vectors are obtained within one hour), then refer to The number of time series vectors contained in the data set is 48k.

In another preferred embodiment of the present invention, the distance threshold D is 0.1-0.5 times the median of all cosine distances calculated for all time series vectors in the reference data set, preferably 0.4-0.5 times .

Compared with the existing urban water supply network leakage identification method, the identification method of the present invention has the advantages of:

First, different from the traditional neural network model, the present invention connects multiple GRU layers and fully connected layers in parallel or in series to form a complex deep neural network. GRUN has a strong ability to deal with nonlinear data, especially for sequence data. It can generate a memory state for past data and establish dependencies between data in different periods. Secondly, GRUN has a strong fitting ability, it is easier to converge during the learning process, and it is not easy to fall into a local minimum state. Therefore, the use of GRUN can predict the flow of the pipeline network more accurately, thereby identifying the leakage of the pipeline network, and improving the accuracy of the identification of the leakage of the pipeline network.

Second, after using GRUN to accurately predict the flow rate, the present invention uses an outlier detection method based on cosine distance to identify leakage accidents, and determines whether leakage occurs by analyzing and comparing the cosine distance between time series vectors (consisting of predicted values and measured values). damage accident. On the one hand, the use of cosine distance eliminates the impact of the large fluctuation range of pipe network monitoring data on outlier detection; on the other hand, compared with the absolute difference between the analysis predicted value and the measured value, the cosine The comparison of the distance avoids the subjectivity when judging the leakage accident, and the accuracy is higher.

Description of drawings

FIG. 1 is a topological structure diagram of a neural network model based on a gated recurrent unit in an embodiment of the present invention.

FIG. 2 is a flow chart of training a neural network model based on gated recurrent units in an embodiment of the present invention.

FIG. 3 is a graph showing mean square errors of training curves and verification curves in Example 1 of the present invention.

Fig. 4 is a flow chart of using the neural network model in an embodiment of the present invention to recognize leakage accidents of the water supply pipe network.

Detailed ways

In order to better understand and implement the present invention, the present invention will be described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be understood that although the embodiments of the present invention have been described, it is obvious that the present invention is not limited to the above-described embodiments, and various modifications can be made without departing from the gist thereof.

In the following embodiments, Python 2.7 software is used as the model development platform, Numpy and Pandas libraries are used to read, store, and analyze data, Matplotlib library is used to visualize data, and Keras library is used to build neural network models. Improve development efficiency.

Embodiment 1 training neural network model

1) Construct the neural network model according to Figure 1

The input layer is composed of 3 GRU layers and the first fully connected layer in parallel (set the number of nodes of the 3 GRU layers and the first fully connected layer to be 48, 32, 32, 8 respectively, and the activation functions of the GRU layer are tanh and ReLU , the activation function of the first fully connected layer is ReLU);

Connect the 3 GRU layers and the first fully connected layer to the Merge layer in a tensor series mode;

Connect the merged layer with the second fully connected layer, the third fully connected layer, the fourth fully connected layer, the fifth fully connected layer, the sixth fully connected layer, and the seventh fully connected layer in the tensor series mode (set the first The number of nodes from the second fully connected layer to the seventh fully connected layer is 64, 32, 16, 8, 4, 2, and the activation function is ReLU);

The seventh fully connected layer is connected to the output layer in the tensor series mode (the activation function of the output layer is Linear, the learning rate is 0.02, and the batch_size is 60).

2) Generate training set and validation set

2-1) The flow data of a DMA water supply network in CZ City of the water company from February 1, 2016 to January 31, 2017 was averagely sampled at a collection frequency of 4 times/h, that is, every 15 minutes Preprocessing of the acquired data: including cleaning the traffic data caused by unnatural factors (third party, man-made) accidents; correcting input errors, cleaning obviously abnormal data, etc. Then, extract the historical flow data of the first time period, the second time period, and the third time period, and perform normalization processing according to the formula (1) to obtain the historical flow characteristics of different time periods, where,

The flow data in the first time period include flow data at t-75min, t-60min, t-45min, t-30min, and t-15min;

The flow data in the second time period includes the flow data from t-24.5h, t-24.25h, t-24h, t-23.75h to t-23.5h;

The flow data in the third time period includes the flow data at time t-48.5h, t-48.25h, t-48h, t-47.75h to t-47.5h.

2-2) Obtain historical meteorological data from the China Meteorological Data Network, including maximum temperature, minimum temperature, relative humidity and precipitation, and preprocess the obtained data: including cleaning the flow of accidents caused by unnatural factors (third parties, man-made) data; correct input errors, clean obviously abnormal data, etc. Perform Pearson correlation analysis with the historical flow data in step 2-1), and find that the correlation between the highest temperature and the lowest temperature and the flow rate is significant, and the correlation coefficient is greater than 0.8, so the normalized processing results of the highest temperature and the lowest temperature , as a historical meteorological feature.

2-3) The historical flow characteristics and historical meteorological characteristics of different time periods are used to form the samples required for modeling. First, the samples are randomly disrupted, and then the samples are divided into 10 parts. One part is randomly selected as the verification set, and the remaining 9 parts are used as the verification set. The training set has 22500 samples in the training set and 2500 samples in the verification set.

3) Training model

The training of the neural network model is carried out according to the flow chart in Figure 2. Specifically, the historical flow characteristics and historical meteorological characteristics of the training set and the verification set are respectively input into the input layer of the neural network model, and the output training flow and verification flow are obtained. Every time a round of training is completed, the samples in the training set will be randomly shuffled once.

Table 1

4) Verify the model

A training curve is generated based on the training set and training traffic, and a verification curve is generated based on the verification set and verification traffic. As shown in Figure 3, the abscissa indicates the number of rounds of training, and each round of training iterates 375 times; the ordinate indicates the mean square error of the training set and the verification set. It can be seen that when the mean square error of the training curve and the verification curve is stable at a constant value, it indicates that the model is stable and fits well, and is suitable as a neural network model for recognition.

On the other hand, if the mean square error of the two curves is not stable at a constant value, it means that the model is unstable, the fitting is poor, there are too few training data or the model parameters need to be optimized.

In this example, the trained neural network model can be exemplarily expressed as:

Q _t =Q _ti ×W+b

In the above formula, _Qt is the water demand at the time t to be predicted;

Q _ti is the historical water demand at time ti, i=1, 2, 3...;

W is the weight matrix;

b is a bias term.

Embodiment 2 Leakage Accident Identification

It is known that no leakage accidents occurred from February 1, 2017 to February 10, 2017. The reference data set was constructed with the flow data from February 1 to February 2, 2017, and used to identify February 3, 2017 Leakage accident until February 10.

In order to verify the identification effect, the flow data on February 4, 5 and 6, 2017 were artificially increased by 5%, 10% and 15% of the DMA daily average flow (79m ³ /h) respectively.

Carry out leakage accident identification according to the flow chart shown in Figure 4, specifically,

Step 1: Obtain the reference flow characteristics and reference meteorological characteristics within 48 hours from February 1, 2017 to February 2, 2017, according to the method of step 2) in Example 1.

Step 2, input the reference flow characteristics obtained in step 1 in different time periods into the multiple gated recurrent unit layers of the neural network model trained in embodiment 1, and input the reference meteorological characteristics obtained in step 1 into the first In a fully connected layer, the output prediction flow is obtained.

Step 3: Acquire the measured flow data within 48 hours from February 1, 2017 to February 2, 2017, and form a time series vector with each measured flow and the predicted flow obtained in step 2 at the same time. Then, in these two days, a total of 192 time series vectors (48h, the acquisition frequency is 4 times/h) are formed, and the matrix formed by these 192 time series vectors is used as a reference data set. Table 2 shows the reference data set Part of the example.

Table 2

Step 4: Process the flow data from February 3rd to February 10th, 2017 into 768 new time series vectors (192h, the collection frequency is 4 times/h) according to the above steps 1 to 3.

Based on formula (2), calculate the cosine distance of each new time series vector relative to each time series vector in the reference data set obtained in step 3, among the 192 cosine distances calculated by each new time series vector , the number n of time series vectors whose cosine distance is less than the distance threshold D is counted. In this example, the distance threshold D is 0.45 times the median of 192 cosine distances, that is, 0.00133, and the number threshold N is 42.

When n is greater than the number threshold N, the newly added time series vector is determined to be a normal vector, which means that no leakage accidents have been identified at this time; when n is below the number threshold N, the newly added time series vector is determined to be an abnormal vector, Indicates the time at which a leakage accident was identified.

Table 3 Recognition results

Theoretically, there should be 480 normal vectors and 288 abnormal vectors among the 768 new time series for 8 days from February 3 to February 10, 2017. According to the identification result of table 3, it can be known that the neural network model of embodiment 1 can identify the leakage (+5% DMA daily average flow) of the flow, but it is not sensitive; %, +15% DMA daily average traffic) can realize accurate identification, and the accuracy rate is above 85%.

The above results show that the neural network model based on the gated cyclic unit can more accurately identify the leakage accidents of the water supply network, and the method is more practical. The invention expands the research content of the existing pipeline network leakage identification model, and provides a new idea for water companies to make scientific and reasonable decisions.

While the invention has been described in detail, modifications within the spirit and scope of the invention will be readily apparent to those skilled in the art. In addition, it should be understood that various aspects described in the present invention, various parts of different specific embodiments, and various listed features may be combined or interchanged in whole or in part. In each of the specific embodiments described above, those embodiments that refer to another specific embodiment may be appropriately combined with other embodiments, as will be understood by those skilled in the art. Furthermore, those skilled in the art will appreciate that the foregoing description is by way of example only, and is not intended to limit the invention.

Claims (9)

1. A neural network model based on gated recurrent units for identifying leakage accidents in water supply networks, including: a plurality of layers of gated recirculation units configured to receive flow characteristics of the water distribution network at different time periods; A first fully connected layer, which is connected in parallel with the multiple gated recurrent unit layers to form an input layer, and the first fully connected layer is configured to receive meteorological features of the water supply network; a pooling layer connected in tensor concatenation with said input layer; The second fully-connected layer to the Mth fully-connected layer, all of which are connected to the merging layer in tensor series, where M is an integer greater than or equal to 2; and An output layer connected to the Mth fully connected layer in tensor series, configured to output the predicted flow rate of the water supply network at the next moment. 2. a training method of the neural network model claimed in claim 1, comprising: Construct the neural network model; Randomly divide the historical flow characteristics and historical meteorological characteristics of different time periods into a training set and a verification set; Input the historical traffic characteristics of different time periods in the training set into the multiple gated recurrent unit layers, and input the historical meteorological characteristics in the training set into the first fully connected layer, and obtain the neural The training traffic output by the network model; Input the historical traffic characteristics of different time periods in the verification set into the multiple gated recurrent unit layers, and input the historical meteorological characteristics in the verification set into the first fully connected layer to obtain the neural Verification traffic output by the network model; Generate a training curve based on the training set and training traffic, and generate a verification curve based on the verification set and verification traffic, when the mean square error value of the training curve and verification curve is stable at a constant value, complete the neural network model training. 3. The training method according to claim 2, further comprising obtaining historical traffic characteristics of the different time periods, including: Obtain historical flow data of the water supply network to be tested; and Extraction processing and normalization processing are performed on the historical traffic data to obtain the historical traffic features. 4. The training method according to claim 3, wherein extracting the historical flow data comprises, in the historical flow data of the water supply pipe network, respectively starting from the first time period and the second time period , Extract the specified number of traffic data at equal intervals in the third time period, where, Select the first time period before the time t to be predicted, selecting a second time period before the starting moment of the first time period, A third time period is selected before the starting moment of the second time period. 5. according to the described training method of claim 3 or 4, it is characterized in that, also comprise obtaining described historical meteorological characteristic, comprise: Obtain historical weather data of the water supply network to be tested; and Extraction processing and normalization processing are performed on the historical meteorological data to obtain the historical meteorological features. 6. training method according to claim 5, it is characterized in that, extracting described historical meteorological data comprises: carrying out Pearson correlation analysis to described historical meteorological data and described historical flow data, select correlation coefficient greater than or The historical meteorological data equal to the specified value is obtained to obtain the extracted and processed historical meteorological data. 7. The training method according to claim 6, characterized in that, the prescribed value is 0.80-0.99. 8. A method for identifying leakage accidents of a water supply pipe network, comprising: The feature acquisition step is to acquire the reference flow characteristics and reference meteorological characteristics of the water supply network to be tested in different time periods within the specified time period when no leakage accident occurs; Forecast flow acquisition step, input the reference flow characteristics of the different time periods into the multiple gated recurrent unit layers of the neural network model according to claim 1, and input the reference meteorological characteristics into the neural network model In the first fully connected layer, obtain the predicted traffic output by the neural network model; Refer to the step of determining the data set to obtain the measured flow within the specified time, each group of predicted flow and measured flow at the same time constitutes a time series vector, and use the matrix formed by all time series vectors within the specified time as a reference data set; The accident identification step is to calculate the cosine distance of the newly added time series vector relative to each time series vector in the reference data set, and count the number of time series vectors whose cosine distance is less than the distance threshold, when the number of statistics is less than or equal to When the quantity threshold value is reached, it is determined that the leakage accident has occurred; Wherein, the newly added time-series vector is a time-series vector obtained according to the feature acquisition step, forecast traffic acquisition step, and reference data value determination step in a time period other than the specified time. 9. The identification method according to claim 8, wherein the distance threshold is 0.1-0.5 times the median of all cosine distances calculated for all time series vectors in the reference data set.