Water supply network pipe burst early warning method based on gated neural network

Abstract

The invention discloses a gated neural network-based pipe burst early warning method for a water supply network, which comprises the following steps: S1, acquiring the temperature at the current moment, and inputting the temperature at the current moment into a trained gated neural network to obtain a predicted value of water consumption at the current moment; s2, inputting the water consumption predicted value into a pipe network hydraulic model to obtain a pipeline pressure calculated value of the water supply pipe network at the current moment; s3, calculating the pipe explosion risk degree at the current moment according to the pipeline pressure calculation value; and S4, judging whether the pipe explosion risk degree at the current moment is greater than a preset value, if so, triggering a pipe explosion alarm, and otherwise, returning to the step S1 to continue monitoring.

Description

Technical Field

[0001] The invention relates to a water supply pipeline monitoring technology, and in particular to a water supply pipeline network burst warning method based on a gated neural network. Background Art

[0002] The urban water supply network is a complex underground project. When its pipelines burst, it may cause problems such as waste of water resources, water pollution, and large-scale water outages. In serious cases, it may bring major disasters or affect the social economy, and even affect people’s lives and property.

[0003] At present, water supply companies in some large cities will establish their own monitoring systems to monitor whether a pipe burst has occurred. During monitoring, the monitoring system will collect the pressure and flow of the pipeline network. If some of the monitoring values ​​of the pressure and flow of the pipeline network change drastically, it is determined that a pipe burst has occurred. At this time, inspectors will be sent out to inspect. This method is an operation to locate the pipe burst after it occurs, and the inspection range of the pipeline network pipe burst monitoring will be relatively large, which will take a long time and will not reduce the losses caused by the pipe burst.

[0004] Some water supply companies will establish an online water supply network hydraulic model based on the monitoring system, and use the difference between the monitored value and the model calculated value to make a judgment. If the difference suddenly reaches a certain level, it is suspected that a pipe burst has occurred, and then inspectors are sent to check the suspicious location. This method, due to the hydraulic model of the pipe network, can limit the location of the pipe burst to a few locations to narrow the inspection scope. However, it is still an operation to locate the pipe burst after it occurs. It can only minimize the losses caused by the pipe burst, and cannot predict the risk of pipe burst in advance. Summary of the invention

[0005] In view of the above-mentioned deficiencies in the prior art, the water supply pipe network burst warning method based on gated neural network provided by the present invention solves the problem that the prior art cannot provide early warning of the risk of water supply pipe network burst.

[0006] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is:

[0007] A water supply pipe burst warning method based on a gated neural network is provided, which comprises the following steps:

[0008] S1. Obtain the temperature at the current moment, and input the temperature at the current moment into the trained gated neural network to obtain the predicted value of water consumption at the current moment;

[0009] S2. Input the water consumption prediction value into the pipe network hydraulic model to obtain the calculated value of the pipe pressure of the water supply pipe network at the current moment;

[0010] S3. Calculate the risk of pipe burst at the current moment based on the calculated value of pipeline pressure:

[0011]

[0012] Among them, R 预测 is the pipe burst risk at the current moment; C i is the value of the i-th pipe burst risk index; Q i C i The weight of 预测 is the calculated value of pipeline pressure; P 允许 Design the maximum pressure value for the pipeline;

[0013] S4. Determine whether the pipe burst risk at the current moment is greater than a preset value. If so, trigger a pipe burst alarm. Otherwise, return to step S1 to continue monitoring.

[0014] Furthermore, the training method of the gated neural network includes:

[0015] S11, obtaining the air temperature and user water consumption at multiple times as a training set, and obtaining preset proportions of samples from the training set as a parameter adjustment training set and a parameter adjustment test set;

[0016] S12, obtaining the value ranges of multiple hyperparameters in the gated neural network, and randomly generating several groups of parameter solutions within the value ranges, and then dividing the parameter solutions into hyperparameter matrix A and hyperparameter matrix B in equal proportions;

[0017] S13, inputting each set of parameter solutions of the hyperparameter matrix A and the hyperparameter matrix B into an independent gated neural network, and training each gated neural network a preset number of times using the parameter adjustment training set;

[0018] S14, using the parameter adjustment test set to predict the model accuracy of each gated neural network trained, selecting the parameter solution with the highest model accuracy in the two matrices as the optimal solution, and marking the remaining parameter solutions as inferior solutions;

[0019] S15, judging whether the current number of cycles is greater than the preset number of cycles or whether the current optimal solution is equal to the optimal solution of the previous cycle, if any one of them is yes, proceeding to step S17, if both are no, proceeding to step S16;

[0020] S16, adjusting the inferior solution according to the two optimal solutions, using the two optimal solutions and the adjusted inferior solution to update the corresponding parameter solutions in the hyperparameter matrix A and the hyperparameter matrix B, and then returning to step S13;

[0021] S17, selecting a parameter solution with the highest model accuracy in the current cycle, and using the selected parameter solution as a hyperparameter of the gated neural network to obtain the gated neural network after parameter optimization;

[0022] S18. Using the training set to train the gated neural network after parameter optimization, to obtain the final gated neural network.

[0023] Furthermore, the method for adjusting the inferior solution according to the two optimal solutions includes:

[0024] A1. In the solution space, using the solution parameters as the position vector, calculate the distance between each parameter solution in the hyperparameter matrix B and the closest parameter solution in the hyperparameter matrix A.

[0025] A2. Calculate the ratio of the distance corresponding to each parameter solution in the hyperparameter matrix B to the maximum distance in the hyperparameter matrix B, and select the adjustment strategy of the corresponding parameter solution according to the ratio;

[0026] A3. For each inferior solution in the hyperparameter matrix A, randomly adjust each inferior solution based on the position of the optimal solution in the hyperparameter matrix A;

[0027] A4. For each inferior solution in the hyperparameter matrix B, adjust the inferior solution according to the selected adjustment strategy.

[0028] The beneficial effect of the above technical solution is: this solution enables the hyperparameter matrix A and the hyperparameter matrix B to perform parameter optimization according to the above logic. The hyperparameter matrix A will gradually occupy the local optimal solution found in the solution space, and the hyperparameter matrix B will continue to explore areas far away from the local optimal solution, thereby improving the efficiency of finding the global optimal solution of the GRU model hyperparameters.

[0029] Furthermore, an adjustment strategy for the corresponding parameter solution is selected according to the ratio:

[0030] When 0.75<k≤1, the selected adjustment strategy is:

[0031]

[0032] When 0.5<k≤0.75, the selected adjustment strategy is:

[0033]

[0034] When 0.25<k≤0.5, the selected adjustment strategy is:

[0035]

[0036] When 0<k≤0.25, the selected adjustment strategy is:

[0037]

[0038] Among them, k is the ratio; and are the lth inferior solutions in the hyperparameter matrix B at the t+1th and tth cycles respectively; is the optimal solution in the hyperparameter matrix B in the tth cycle; is the hyperparameter matrix A with The closest parametric solution.

[0039] The beneficial effect of the above technical solution is: this solution affects the hyperparameter matrix B through the hyperparameter matrix A, and combined with the above four adjustment strategies, it can not only effectively help the optimization process of the hyperparameter matrix B to escape the local optimal solution, but also improve the convergence speed of the hyperparameter optimization algorithm.

[0040] The closer the solution of hyperparameter matrix B is to the solution of hyperparameter matrix A, the more aggressive the triggered adjustment strategy is, and the greater the probability that the solution of hyperparameter matrix B will escape the local optimal trap, which can effectively reduce the computing resources required by the algorithm.

[0041] Furthermore, in step A3, the expression for randomly adjusting each inferior solution is:

[0042]

[0043] in, and are the lth inferior solutions in the hyperparameter matrix A at the t+1th and tth cycles respectively; is a random vector between [0,1]; is the optimal solution in the hyperparameter matrix A in the tth cycle.

[0044] The beneficial effect of the above technical solution is: this solution continuously reduces the distance between the inferior solution and the superior solution in the hyperparameter matrix A, so that the inferior solutions in the hyperparameter matrix A are distributed around the local optimal solution discovered by the current algorithm. This can not only realize the intensive search of the hyperparameter matrix A for the space around the local optimal solution, but also make the adjustment strategy for selecting the solution close to the local optimal solution in the hyperparameter matrix B more aggressive, thereby increasing the probability of the solution of the hyperparameter matrix B jumping out of the local optimal trap.

[0045] Furthermore, between step S15 and step S16, it is also included to judge whether the current loop number is a multiple of ten. If so, the hyperparameter matrix A and the hyperparameter matrix B share the optimal solution in the next loop; otherwise, the optimal solution is not shared.

[0046] The beneficial effect of the above technical solution is: by sharing the optimal solution, the two matrices can know the optimal solution parameters discovered by the other party, which is convenient for further enabling the solution matrix to escape the local optimal solution trap and explore the unknown solution space.

[0047] Furthermore, the multiple hyperparameters are the number of hidden units, sliding window sequence length, batch size, learning rate, optimizer type, number of layers, and dropout rate.

[0048] Furthermore, the pipe burst risk index includes the ratio of pipeline pressure to allowable pressure C1, the ratio of service life to design life C2, the number of historical pipe bursts C3, whether there are important users downstream of the pipeline C4, the economic loss caused by the pipe burst C5, the degree of social traffic interruption caused by the pipe burst C6, the time required for maintenance C7 and whether there are other water supply sources downstream of the pipeline C8; the weight of the pipe burst risk index is determined according to the value of the pipe burst risk index using the fuzzy hierarchical analysis method.

[0049] The beneficial effects of the present invention are as follows: the present invention firstly adopts a gated neural network to predict the water consumption of the water supply network according to the temperature at the current moment, then adopts a network hydraulic model to obtain a calculated value of the pipeline pressure, and finally predicts the pipe burst risk degree in combination with the pipe burst risk degree formula constructed by the present invention. The pipe burst risk degree can be used to know whether there is a risk in the water supply network at the current moment, so as to give an early warning, so that the management personnel can prepare a response strategy in advance to avoid pipe bursts in the water supply network, thereby reducing the operation and maintenance cost of the water supply network, and at the same time, it can also avoid the impact of pipe bursts on the normal water use of users.

[0050] When training the gated neural network, this scheme adjusts the hyperparameters of the gated neural network through the hyperparameter matrix A and the hyperparameter matrix B and multiple parameter adjustment strategies, which can improve the training speed and prediction accuracy of the gated neural network, thereby ensuring the accuracy of the water supply network pressure prediction and the accuracy of the early warning. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 This is a flow chart of a water supply network burst warning method based on a gated neural network.

[0052] Figure 2 This is a comparison chart of the prediction results of the genetic algorithm-based gated neural network after 1 hour of iterative optimization.

[0053] Figure 3 Comparison of prediction results of the water demand prediction model based on the improved GRU after 1-hour iterative optimization.

[0054] Figure 4 This is a color-coded diagram of pipe burst warning levels based on water consumption forecasts. DETAILED DESCRIPTION

[0055] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0056] refer to Figure 1 , Figure 1 The flowchart of the water supply network burst warning method based on the gated neural network is shown; Figure 1 As shown, the method S includes steps S1 to S4.

[0057] In step S1, the temperature at the current moment is obtained, and the temperature at the current moment is input into the trained gated neural network to obtain the predicted value of water consumption at the current moment;

[0058] In step S2, the water consumption prediction value is input into the pipe network hydraulic model to obtain the calculated value of the pipe pressure of the water supply pipe network at the current moment; the pipe network hydraulic model can be foreign Epanet, WaterGEMS software, or domestic WaterDesk software.

[0059] In step S3, the pipe burst risk at the current moment is calculated according to the calculated value of the pipeline pressure:

[0060]

[0061] Among them, R 预测 is the pipe burst risk at the current moment; C i is the value of the i-th pipe burst risk index; Q i C i The weight of 预测 is the calculated value of pipeline pressure; P 允许 Design the maximum pressure value for the pipeline;

[0062] The weight of the pipe burst risk index of this scheme is determined according to the value of the pipe burst risk index using the fuzzy hierarchical analysis method. The pipe burst risk index and weight can be referred to Table 1.

[0063] Table 1 Pipe burst risk indicators and weights

[0064]

[0065] In step S4, it is determined whether the risk of pipe burst at the current moment is greater than the preset value. If so, the pipe burst alarm is triggered, otherwise it returns to step S1 to continue monitoring. The early warning method of this scheme is theoretically always running, and it continuously or every preset time period collects the temperature at the current time to issue a pipe burst warning for the water supply network; it can also be restarted after running for a period of time.

[0066] In one embodiment of the present invention, the training method of the gated neural network includes:

[0067] S11. Obtain the air temperature and user water consumption at multiple moments as training sets, and obtain preset proportions of samples in the training sets as parameter adjustment training sets and parameter adjustment test sets. The preset proportions here are very small, so that fewer training samples can speed up the convergence of the model during hyperparameter optimization.

[0068] S12. Obtain the value ranges of multiple hyperparameters in the gated neural network, and randomly generate several groups of parameter solutions within the value ranges, and then divide the parameter solutions into hyperparameter matrix A and hyperparameter matrix B in equal proportions; the preferred hyperparameters in this scheme are the number of hidden units, sliding window sequence length, batch size, learning rate, optimizer type, number of layers, and dropout rate.

[0069] To facilitate the understanding of the hyperparameter matrix A and the hyperparameter matrix B, the following shows a schematic diagram of only 5 sets of parameter solutions in the hyperparameter matrix A:

[0070]

[0071] in, They are the first to fifth groups of parameter solutions, that is, each group of parameter solutions includes 7 hyperparameters. The first subscript in x indicates which group of parameter solutions it belongs to. The second subscripts 1 to 7 represent the number of hidden units, sliding window sequence length, batch size, learning rate, optimizer type, number of layers, and dropout rate, respectively.

[0072] S13. Input each set of parameter solutions of the hyperparameter matrix A and the hyperparameter matrix B into an independent gated neural network, and use the parameter adjustment training set to train each gated neural network for a preset number of times; assuming that there are five sets of parameter solutions in the hyperparameter matrix A, then each parameter solution serves as a hyperparameter of an independent gated neural network, that is, the hyperparameter matrix A corresponds to five gated neural networks.

[0073] S14, using the parameter adjustment test set to predict the model accuracy of each gated neural network trained, selecting the parameter solution with the highest model accuracy in the two matrices as the optimal solution, and marking the remaining parameter solutions as inferior solutions;

[0074] S15, judging whether the current number of cycles is greater than the preset number of cycles or whether the current optimal solution is equal to the optimal solution of the previous cycle, if any one of them is yes, proceeding to step S17, if both are no, proceeding to step S16;

[0075] S16, adjusting the inferior solution according to the two optimal solutions, using the two optimal solutions and the adjusted inferior solution to update the corresponding parameter solutions in the hyperparameter matrix A and the hyperparameter matrix B, and then returning to step S13;

[0076] During implementation, the method for adjusting the inferior solution based on the two optimal solutions preferably includes:

[0077] A1. In the solution space, take the solution parameters as the position vector and calculate the distance between each parameter solution in the hyperparameter matrix B and the closest parameter solution in the hyperparameter matrix A:

[0078]

[0079] in, is the distance in the hyperparameter matrix A The nearest parameter solution; is the lth inferior solution in the hyperparameter matrix B.

[0080] A2. Calculate the ratio of the distance corresponding to each parameter solution in the hyperparameter matrix B to the maximum distance in the hyperparameter matrix B, and select the adjustment strategy of the corresponding parameter solution according to the ratio:

[0081] When 0.75<k≤1, the selected adjustment strategy is:

[0082]

[0083] When 0.5<k≤0.75, the selected adjustment strategy is:

[0084]

[0085] When 0.25<k≤0.5, the selected adjustment strategy is:

[0086]

[0087] When 0<k≤0.25, the selected adjustment strategy is:

[0088]

[0089] Among them, k is the ratio; and are the lth inferior solutions in the hyperparameter matrix B at the t+1th and tth cycles respectively; is the optimal solution in the hyperparameter matrix B in the tth cycle; is the hyperparameter matrix A with The closest parametric solution.

[0090] A3. For each inferior solution in the hyperparameter matrix A, randomly adjust each inferior solution based on the position of the optimal solution in the hyperparameter matrix A:

[0091]

[0092] in, and are the lth inferior solutions in the hyperparameter matrix A at the t+1th and tth cycles respectively; is a random vector between [0,1]; is the optimal solution in the hyperparameter matrix A in the tth cycle.

[0093] A4. For each inferior solution in the hyperparameter matrix B, adjust the inferior solution according to the selected adjustment strategy.

[0094] S17, selecting a parameter solution with the highest model accuracy in the current cycle, and using the selected parameter solution as a hyperparameter of the gated neural network to obtain the gated neural network after parameter optimization;

[0095] S18. Using the training set to train the gated neural network after parameter optimization, to obtain the final gated neural network.

[0096] This solution uses the above method to optimize the hyperparameters of the gated neural network, which has faster convergence efficiency and better global search capability, can more effectively escape the local optimal trap, and can significantly reduce the resources required for algorithm calculation.

[0097] During implementation, the preferred step of this scheme also includes determining whether the current number of loops is a multiple of ten between step S15 and step S16. If so, the hyperparameter matrix A and the hyperparameter matrix B share the optimal solution in the next loop; otherwise, the optimal solution is not shared. This can further facilitate the solution matrix to escape the local optimal solution trap and explore the unknown solution space.

[0098] This scheme also combines the collected data on pipe bursts in water supply networks in various places, and establishes a mapping relationship between the pipe burst risk R and the pipe burst warning level through the fuzzy hierarchy process FAHP combined with the actual management implementation difficulty. In addition, each pipeline is color-coded in the pipe burst warning system layer according to different warning levels to provide eye-catching pipe burst prompts, as shown in Table 2.

[0099] Table 2 Mapping relationship between pipe burst risk and pipe burst warning level

[0100] The percentage of pipe burst risk to the warning threshold Pipe burst warning level Layer Color ≤70％ Low probability of pipe burst, no warning information green 70％～80％ There is a medium probability of pipe bursting. It is recommended to dispatch and reduce the pressure. blue 80％～90％ There is a high probability of pipe bursting, and managers should pay attention to it orange color ≥100％ There is a high probability of pipe burst, triggering pipe burst warning red

[0101] The effect of the water supply pipe network burst warning method of this scheme is described below in conjunction with specific embodiments:

[0102] This method is applied to the water supply network in a county in Fujian Province to build a scientific early warning system for pipe burst in the water supply network. The county's tap water company has built a complete water supply marketing system to obtain users' long-term historical water consumption data. In addition, the users in the county are classified according to the collection frequency of water load metering, and a water consumption prediction model (gated neural network) with corresponding accuracy is established for users with different collection frequencies. For example, some large users such as factories and shopping malls in the county are equipped with real-time remote intelligent water meters, which can provide high-frequency historical water consumption data. Taking this county as an example, the effect of the gated neural network water supply network pipe burst early warning method based on parameter optimization is explored in practical applications.

[0103] A user is equipped with an online smart water meter with a collection frequency of 15 minutes. 4588 historical water consumption data of the user in continuous historical periods are selected, and 1250 random time periods of water consumption data are taken as the model test set, and the water consumption data of other data time periods are used as the training data set.

[0104] Within the time limit of 15 minutes for the update step length of the county's water supply network burst warning system, the gated neural network based on genetic algorithm optimization and the improved GRU model (gated neural network) proposed by the hyperparameter optimization of this scheme are used to compare the prediction results of the random water demand test set with the true value of the test set. The comparison results are as follows: Figure 2 , Figure 3 As shown. Figure 2 and Figure 3 It can be seen that within the time required by the pipe burst warning system, the predicted value of the water demand prediction model based on genetic algorithm training almost completely deviates from the true value, the error is very large, and it cannot provide scientific guidance for pipe burst warning. However, thanks to the proposed fast and effective hyperparameter optimization algorithm, the predicted value of the improved GRU model (the gated neural network after parameter optimization in this scheme) almost coincides with the true value, with a small error. It can be seen that the accuracy can meet the needs of pipe burst warning.

[0105] Taking a pipe burst record of the county's water supply network as an example, the historical water consumption data of the month before the pipe burst event was used to train the gated neural network after parameter optimization of this scheme, and the water consumption of each pipeline at the prediction time was obtained. Then, based on the water consumption and the hydraulic model of the pipe network, the pressure value of the water supply network was obtained. After that, the values ​​of the influencing factors of each pipeline on the pipe burst were collected, and the pipe burst risk of each pipeline at the prediction time was calculated in combination with the pipe burst risk calculation formula proposed by this method. Through the mapping relationship between the pipe burst risk and the pipe burst warning level, a pipe burst warning color annotation layer was established in the offline pipe network model, and different colors were used to provide eye-catching annotations for pipe network dispatching managers, such as Figure 4 shown.

[0106] exist Figure 4 In the figure, the black circle indicates the pipe with a burst risk greater than 100%, reaching an extremely high probability of burst, triggering a burst warning. At the same time, this is also the actual location where the burst occurs in reality, which shows that the water supply network burst warning method of this method can effectively warn of burst-risk pipelines, provide a scientific basis for network operation and maintenance, and provide strong support for reducing social property losses and maintaining economic development.

[0107] This method was tested by simulating multiple historical pipe burst conditions in a southeastern pipe network, and successfully completed the pipe burst event warning. Compared with other pipe burst warning methods, this method has the advantages of being stable, fast, and scientific. Practice has shown that this method can effectively realize pipe burst warning in water supply pipe networks and reduce or even avoid losses caused by pipe bursts.

Claims

1. A water supply pipe burst warning method based on a gated neural network, characterized in that: Includes steps: S1. Obtain the temperature at the current moment, and input the temperature at the current moment into the trained gated neural network to obtain the predicted value of water consumption at the current moment; S2. Input the water consumption prediction value into the pipe network hydraulic model to obtain the calculated value of the pipe pressure of the water supply pipe network at the current moment; S3. Calculate the risk of pipe burst at the current moment based on the calculated value of pipeline pressure: Among them, R 预测 is the pipe burst risk at the current moment; C i is the value of the i-th pipe burst risk index; Q i C i The weight of 预测 is the calculated value of pipeline pressure; P 允许 Design the maximum pressure value for the pipeline; S4. Determine whether the pipe burst risk at the current moment is greater than a preset value. If so, trigger a pipe burst alarm. Otherwise, return to step S1 to continue monitoring.

2. The water supply network burst warning method based on gated neural network according to claim 1 is characterized in that: Gated neural network training methods include: S11, obtaining the temperature and user water consumption at multiple times as a training set, and obtaining samples of preset proportions from the training set as a parameter adjustment training set and a parameter adjustment test set; S12, obtaining the value ranges of multiple hyperparameters in the gated neural network, and randomly generating several groups of parameter solutions within the value ranges, and then dividing the parameter solutions into hyperparameter matrix A and hyperparameter matrix B in equal proportions; S13, inputting each set of parameter solutions of the hyperparameter matrix A and the hyperparameter matrix B into an independent gated neural network, and training each gated neural network a preset number of times using the parameter adjustment training set; S14, using the parameter adjustment test set to predict the model accuracy of each gated neural network trained, selecting the parameter solution with the highest model accuracy in the two matrices as the optimal solution, and marking the remaining parameter solutions as inferior solutions; S15, judging whether the current number of cycles is greater than the preset number of cycles or whether the current optimal solution is equal to the optimal solution of the previous cycle, if any one of them is yes, proceeding to step S17, if both are no, proceeding to step S16; S16, adjusting the inferior solution according to the two optimal solutions, using the two optimal solutions and the adjusted inferior solution to update the corresponding parameter solutions in the hyperparameter matrix A and the hyperparameter matrix B, and then returning to step S13; S17, selecting a parameter solution with the highest model accuracy in the current cycle, and using the selected parameter solution as a hyperparameter of the gated neural network to obtain the gated neural network after parameter optimization; S18. Using the training set to train the gated neural network after parameter optimization, to obtain the final gated neural network.

3. The water supply network burst warning method based on gated neural network according to claim 2 is characterized in that: Methods for adjusting the inferior solution based on the two optimal solutions include: A1. In the solution space, using the solution parameters as the position vector, calculate the distance between each parameter solution in the hyperparameter matrix B and the closest parameter solution in the hyperparameter matrix A. A2. Calculate the ratio of the distance corresponding to each parameter solution in the hyperparameter matrix B to the maximum distance in the hyperparameter matrix B, and select the adjustment strategy of the corresponding parameter solution according to the ratio; A3. For each inferior solution in the hyperparameter matrix A, randomly adjust each inferior solution based on the position of the optimal solution in the hyperparameter matrix A; A4. For each inferior solution in the hyperparameter matrix B, adjust the inferior solution according to the selected adjustment strategy.

4. The water supply pipe network burst warning method based on gated neural network according to claim 3 is characterized in that: The adjustment strategy of the corresponding parameter solution is selected according to the ratio: When 0.75<k≤1, the selected adjustment strategy is: When 0.5<k≤0.75, the selected adjustment strategy is: When 0.25<k≤0.5, the selected adjustment strategy is: When 0<k≤0.25, the selected adjustment strategy is: Among them, k is the ratio; and are the lth inferior solutions in the hyperparameter matrix B at the t+1th and tth cycles respectively; is the optimal solution in the hyperparameter matrix B in the tth cycle; is the hyperparameter matrix A with The closest parametric solution.

5. The water supply pipe network burst warning method based on gated neural network according to claim 3 is characterized in that: The expression for randomly adjusting each inferior solution in step A3 is: in, and are the lth inferior solutions in the hyperparameter matrix A at the t+1th and tth cycles respectively; is a random vector between [0,1]; is the optimal solution in the hyperparameter matrix A in the tth cycle.

6. The water supply network burst warning method based on gated neural network according to claim 3 is characterized in that: The steps between step S15 and step S16 also include: Determine whether the current number of loops is a multiple of ten. If so, the hyperparameter matrix A and the hyperparameter matrix B share the optimal solution in the next loop; otherwise, the optimal solution is not shared.

7. The water supply pipe burst warning method based on gated neural network according to any one of claims 2 to 5, characterized in that: The multiple hyperparameters are the number of hidden units, sliding window sequence length, batch size, learning rate, optimizer type, number of layers, and dropout rate.

8. The water supply network burst warning method based on gated neural network according to any one of claims 1 to 5, characterized in that: The pipe burst risk indicators include the ratio of pipeline pressure to allowable pressure C1, the ratio of service life to design life C2, the number of historical pipe bursts C3, whether there are important users downstream of the pipeline C4, the economic losses caused by the pipe burst C5, the degree of social traffic interruption caused by the pipe burst C6, the time required for maintenance C7 and whether there are other water supply sources downstream of the pipeline C8; The weight of the pipe burst risk index is determined according to the value of the pipe burst risk index by using the fuzzy hierarchical analysis method.

Images