Method for reconstructing non-monitored node liquid level data based on monitored node liquid level data

Abstract

The invention discloses a method for reconstructing non-monitored node liquid level data based on monitored node liquid level data, and belongs to the technical field of drainage system data restoration. According to the method, the problem that comprehensive data perception of the whole drainage pipe network cannot be realized due to the fact that unmonitored node data cannot be reconstructed in an existing method is solved. A pipe diameter, pipe length and pipe elevation information matrix is formed according to pipe network topological relation data, and physically constrained data transmission and aggregation are carried out on liquid level data of monitoring nodes according to the pipe diameter, pipe length and pipe elevation information matrix by means of a space-time diagram convolutional neural network. The spatial-temporal characteristic relation between the monitoring nodes and the unmonitored nodes is reserved, and the liquid level data of the unmonitored nodes are reconstructed more accurately and effectively. The liquid level data of all the unmonitored nodes in the drainage pipe network system can be accurately reconstructed only through the liquid level data of a small number of monitoring nodes, and the monitoring equipment cost and the follow-up operation and maintenance management cost are reduced. The method can be applied to drainage system data restoration.

Description

Technical Field

[0001] The present invention belongs to the technical field of drainage system data repair, and in particular relates to a method for reconstructing unmonitored node liquid level data based on monitored node liquid level data. Background Art

[0002] The drainage network is a core component of modern urban infrastructure, and its operating efficiency has a significant impact on the development of the urban economy. At present, the framework of my country's drainage network system is basically mature, but with the promotion of the global digital background, the functional requirements of the drainage system are constantly increasing, and it is urgent to achieve more intelligent and efficient management of the existing drainage network system through intelligent empowerment.

[0003] As the perception data of the drainage network system, the scientific, accurate, timely and reliable monitoring data can provide an intelligent basis for drainage planning and scheduling, network transformation and long-term operation and maintenance, and can play a role in ensuring the safety and efficiency of the network. However, the drainage network system has a late informatization construction, a complex monitoring environment, and difficult later operation and maintenance, and high management costs, resulting in low monitoring space density and unstable monitoring data quality. In this context, how to use the limited monitoring space density to infer the information of other unknown nodes in the drainage network system has become the key to solving monitoring problems and bottlenecks in intelligent development.

[0004] However, existing methods mainly focus on the anomaly repair of monitoring node data, fail to explore the reconstruction of unmonitored node data, and cannot achieve comprehensive data perception of the entire drainage network and the construction of an information supervision platform. Summary of the invention

[0005] The purpose of the present invention is to solve the problem that the existing methods cannot reconstruct the unmonitored node data, resulting in the inability to achieve comprehensive data perception of the entire drainage network, and propose a method for reconstructing the unmonitored node liquid level data based on the monitored node liquid level data.

[0006] The technical solution adopted by the present invention to solve the above technical problems is: a method for reconstructing the liquid level data of unmonitored nodes based on the liquid level data of monitored nodes, and the method specifically comprises the following steps:

[0007] Step 1: Collect the pipe network topology data of the drainage pipe network system and the historical liquid level data of each monitoring node;

[0008] Step 2: Obtain the node-pipeline relationship matrix, pipeline elevation information matrix, node relationship adjacency matrix, pipe diameter information matrix, and pipe length information matrix based on the pipe network topology data; obtain the liquid level data matrix of all nodes based on the historical liquid level data of each monitoring node;

[0009] Step 3: Build a spatiotemporal graph convolutional neural network. The working process of the spatiotemporal graph convolutional neural network is as follows:

[0010] In the spatiotemporal graph convolutional neural network, the input data passes through the input layer, feature extraction unit, information aggregation transmission unit, feature fusion layer, time series recursive unit, dense layer and output layer in sequence; that is,

[0011] The input data is first input to the feature extraction unit through the input layer, and then the output of the feature extraction unit is used as the input of the information aggregation transmission unit;

[0012] The output of the information aggregation transfer unit is used as the input of the feature fusion layer, and the output of the feature fusion layer is used as the input of the time series recursive unit;

[0013] The output of the time series recurrent unit is used as the input of the dense layer, and the output of the dense layer is output through the output layer;

[0014] Step 4: Use the node-pipeline relationship matrix, pipeline elevation information matrix, node relationship adjacency matrix, pipe diameter information matrix, pipe length information matrix and all node liquid level data matrices obtained in step 2 to train the spatiotemporal graph convolutional neural network;

[0015] Step 5. Collect the real-time liquid level monitoring data of each monitoring node in the drainage pipe network system, obtain the real-time liquid level data matrix based on the real-time liquid level monitoring data, input the real-time liquid level data matrix and the node-pipe relationship matrix, pipe elevation information matrix, node relationship adjacency matrix, pipe diameter information matrix and pipe length information matrix of the drainage pipe network system into the trained spatiotemporal graph convolutional network, and output the liquid level data reconstruction results of the unmonitored nodes through the trained spatiotemporal graph convolutional network.

[0016] Furthermore, the pipe network topology data includes the connection relationship between the inspection well, the pump station forebay and the river outlet and the pipeline, the pipeline diameter, the pipeline length and the pipeline elevation.

[0017] Furthermore, the node-pipeline relationship matrix is:

[0018] Take each inspection well, pump station forebay and river outlet as a node and number each node:

[0019]

[0020] Among them, G is the node-pipeline relationship matrix, n represents the number of nodes in the drainage network system, and m represents the number of pipes in the drainage network system;

[0021]

[0022] Among them, G ij It is the element in the i-th row and j-th column of the node-pipeline relationship matrix G.

[0023] Furthermore, the pipeline elevation information matrix is:

[0024]

[0025] Among them, H is the pipeline elevation information matrix;

[0026]

[0027] Among them, h j,us_invert represents the upstream invert elevation of pipe j, h j,ds_invert represents the downstream invert elevation of pipe j, H ij is the element in the i-th row and j-th column of the matrix H.

[0028] Furthermore, the node relationship adjacency matrix is:

[0029]

[0030] Among them, A is the node relationship adjacency matrix;

[0031]

[0032] Among them, A ij is the element in the i-th row and j-th column of the node relationship adjacency matrix A.

[0033] Furthermore, the pipe diameter information matrix is:

[0034] Step 1: Set the search depth δ, search all nodes within the search depth range with node i as the root node, and record the set of searched nodes as node data set N δ , computing node data set N δ The standard deviation d of the diameter of all pipes formed by the internal node connection std :

[0035]

[0036] Among them, d k is the node dataset N δ The diameter of the kth pipeline formed by the internal node connection, d is the node data set N δ The average diameter of all pipes formed by the internal node connection, N is the node data set N δ The total number of pipelines formed by connecting internal nodes;

[0037] Step 2: Node i to node data set N δ The pipe diameter weight f of internal node j w (d ij )for:

[0038] (1) If there is only one path from node i to node j, then

[0039]

[0040] Among them, α is the weight distribution parameter; d qj is the diameter of the pipe connecting node q and node j; q represents the node one depth level above node j on the path from node i to node j; e is the base of the natural logarithm; η qj is the depth level of the pipeline connecting node q to node j;

[0041] (2) If there are at least two paths from node i to node j, then

[0042]

[0043] Where q represents a node one depth level above node j on any path between node i and node j; d qj represents the diameter of the pipe connecting node q and node j; qj is the depth level of the pipeline connecting node j to node q; f w (d qj ) is the pipe diameter weight from node q to node j;

[0044]

[0045] Where Q is the set of nodes located one depth level above node j on all paths from node i to node j; |Q| is the number of nodes in set Q;

[0046] Step 3: D ij As the element of row i and column j in the pipe diameter information matrix D;

[0047] Step 4: After processing each node in the drainage pipe network system using the methods from step 1 to step 3, the pipe diameter information matrix D is obtained:

[0048]

[0049] Furthermore, the pipe length information matrix is:

[0050] Step 1: Set the search depth δ, search all nodes within the search depth range with node i as the root node, and record the set of searched nodes as node data set N δ , computing node data set N δ The standard deviation of the length of all pipes formed by the internal node connection is l std :

[0051]

[0052] Among them, l k is the node dataset N δ The length of the kth pipeline formed by the connection of the internal nodes, is the node dataset N δ The average length of all pipes formed by the internal node connection, N node data set N δ The total number of pipelines formed by connecting internal nodes;

[0053] Step 2: Node i to node data set N δ The tube length weight f of internal node j w (l ij )for:

[0054] (1) If there is only one path from node i to node j, then

[0055]

[0056] Among them, α is the weight distribution parameter; l qj is the length of the pipe connecting node q and node j; q represents the node one depth level above node j on the path from node i to node j; e is the base of the natural logarithm; η qj is the depth level of the pipeline connecting node q to node j;

[0057] (2) If there are at least two paths from node i to node j, then

[0058]

[0059] Where q represents a node one depth level above node j on any path between node i and node j; d qj represents the length of the pipe connecting node q and node j; qj is the depth level of the pipe connecting node j to node q;

[0060]

[0061] Where Q is the set of nodes located one depth level above node j on all paths from node i to node j; |Q| is the number of nodes in set Q;

[0062] Step 3: L ij As the element of row i and column j in the pipe length information matrix L;

[0063] Step 4: After processing each node in the drainage pipe network system using the methods of steps 1 to 3, the pipe length information matrix L is obtained:

[0064]

[0065] Furthermore, the liquid level data matrix of all nodes is obtained according to the historical liquid level data of each monitoring node, specifically:

[0066] Sort the historical liquid level data monitored by the i-th monitoring node in time series to form X p vector:

[0067] X i =[X i1 ,X i2 …,X iT’ ]

[0068] Among them, X it is the vector X i The t-th monitoring value in, that is, the t-th monitoring value monitored by the i-th monitoring node, T' is the total number of monitoring values, t = 1, 2, ..., T';

[0069] Then the data matrix of all node liquid levels is X:

[0070]

[0071] Among them, the superscript T represents the transpose of the matrix;

[0072] In the matrix X, the elements of the rows corresponding to the nodes that are not monitored are set to 0.

[0073] Furthermore, the specific process of step 4 is as follows:

[0074] Step 4: According to the number of monitoring nodes M, input the number of monitoring nodes M in : Number of monitoring nodes for loss calculation M loss =1:1 ratio, generate monitoring node combination library S;

[0075] That is, for each monitoring node combination in the monitoring node combination library S, the monitoring node is divided into two parts: input monitoring node and loss calculation monitoring node;

[0076] For each monitoring node combination in the monitoring node combination library S, a training sample is generated correspondingly, specifically:

[0077] For any monitoring node combination in the monitoring node combination library S, only the liquid level data corresponding to the input monitoring node selected by this monitoring node combination is retained in the new liquid level data matrix X′, and the liquid level data of other monitoring nodes are set to 0. A training sample is formed using the node-pipeline relationship matrix, pipeline elevation information matrix, node relationship adjacency matrix, pipe diameter information matrix, pipe length information matrix and the new liquid level data matrix X′;

[0078] Step 42: Slidingly segment each training sample generated in step 41 according to a time window of a specified size to obtain a training sample library;

[0079] Step 43: Initialize the maximum number of training cycles to process, and initialize the training sample update frequency of each training cycle to epoch;

[0080] Step 44: Set the number of training cycles p = 1;

[0081] Step 45: randomly select a monitoring node combination from the monitoring node combination library S;

[0082] Step 46: Initialize e=1;

[0083] Step 47: randomly obtain one training sample from the training sample set corresponding to the monitoring node combination selected in step 45, and use the selected current sample to train the spatiotemporal graph convolutional neural network;

[0084] The loss function calculation method during training is as follows: using the spatiotemporal graph convolutional neural network to output the predicted results of the liquid level data of the monitoring nodes that are set to 0 in the current sample, and calculating the loss function based on the predicted results and the actual liquid level data;

[0085] Step 48: Determine whether e=epoch is satisfied;

[0086] If satisfied, execute step 49;

[0087] If not satisfied, set e=e+1 and return to step 47;

[0088] Step 49: determine whether p=process is satisfied;

[0089] If satisfied, the entire training process ends and a trained spatiotemporal graph convolutional neural network is obtained;

[0090] If not satisfied, set p=p+1 and return to step 4 or 5.

[0091] Furthermore, the feature extraction unit includes a GNN and a fully connected layer, and the specific working process is as follows:

[0092] Take X′ as the input of the fully connected layer, and then take the output f(X′) of X′ after the fully connected layer as the input of GNN:

[0093] X′ D =GNN((W d *D+b d )⊙f(X′)),A)

[0094] X′ L =GNN((W l*L+b l )⊙f(X′)),A)

[0095] X′ H =GNN((W h *H+b h )⊙f(X′)),G)

[0096] Among them, W d , W l and W h is the weight of GNN, b d , b l and b h is the bias of GNN, * represents element-wise multiplication, ⊙ represents matrix product, X′ D , X′ L and X′ H is the output of the feature extraction unit.

[0097] The beneficial effects of the present invention are:

[0098] The present invention forms a pipe diameter, pipe length and pipeline elevation information matrix based on the pipe network topology relationship data, and uses a spatiotemporal graph convolutional neural network to perform physically constrained data transmission and aggregation on the liquid level data of the monitoring nodes according to the pipe diameter, pipe length and pipeline elevation information matrix, retaining the spatiotemporal characteristic relationship between the monitoring nodes and the unmonitored nodes, and reconstructing the liquid level data of the unmonitored nodes more accurately and effectively.

[0099] The method provided by the present invention can reduce the number of monitoring nodes to be deployed. Only the liquid level data of a small number of monitoring nodes are needed to accurately reconstruct the liquid level data of all unmonitored nodes in the drainage network system. It is not only simple to operate, but also can reduce the cost of monitoring equipment and subsequent operation and maintenance management costs.

[0100] The method provided by the present invention can realize comprehensive perception of the drainage pipe network system and build a drainage information supervision platform, so that the safe operation of the entire urban drainage pipeline can be grasped and controlled, provide scientific and advanced city-level hydraulic analysis capabilities, and more comprehensively evaluate the capabilities of the urban drainage pipe network system. BRIEF DESCRIPTION OF THE DRAWINGS

[0101] Figure 1 It is a flow chart of a method for reconstructing liquid level data of unmonitored nodes based on liquid level data of monitored nodes according to the present invention. DETAILED DESCRIPTION

[0102] Specific implementation method 1: Combination Figure 1 The present embodiment describes a method for reconstructing the liquid level data of unmonitored nodes based on the liquid level data of monitored nodes, and the method specifically comprises the following steps:

[0103] Step 1: Collect the pipe network topology data of the drainage pipe network system and the historical liquid level data of each monitoring node;

[0104] Step 2: Obtain the node-pipeline relationship matrix, pipeline elevation information matrix, node relationship adjacency matrix, pipe diameter information matrix, and pipe length information matrix based on the pipe network topology data; obtain the liquid level data matrix of all nodes based on the historical liquid level data of each monitoring node;

[0105] Step 3: Build a spatiotemporal graph convolutional neural network. The working process of the spatiotemporal graph convolutional neural network is as follows:

[0106] In the spatiotemporal graph convolutional neural network, the input data passes through the input layer, feature extraction unit, information aggregation transmission unit (GNN), feature fusion layer (Conv1D), time series recurrent unit (GRU), dense layer (Dense) and output layer in sequence; that is,

[0107] The input data is first input to the feature extraction unit through the input layer, and then the output of the feature extraction unit is used as the input of the information aggregation transmission unit;

[0108] The output of the information aggregation transfer unit is used as the input of the feature fusion layer, and the output of the feature fusion layer is used as the input of the time series recursive unit;

[0109] The output of the time series recurrent unit is used as the input of the dense layer, and the output of the dense layer is output through the output layer;

[0110] Step 4: Use the node-pipeline relationship matrix, pipeline elevation information matrix, node relationship adjacency matrix, pipe diameter information matrix, pipe length information matrix and all node liquid level data matrices obtained in step 2 to train the spatiotemporal graph convolutional neural network;

[0111] Step 5. Collect the real-time liquid level monitoring data of each monitoring node in the drainage pipe network system, and obtain the real-time liquid level data matrix according to the real-time liquid level monitoring data (the construction method is the same as that of matrix X, that is, the elements of the rows corresponding to the unmonitored nodes are set to 0), and input the real-time liquid level data matrix and the node-pipe relationship matrix, pipeline elevation information matrix, node relationship adjacency matrix, pipe diameter information matrix and pipe length information matrix of the drainage pipe network system into the trained spatiotemporal graph convolutional network. The reconstructed result of the liquid level data of the unmonitored nodes is output through the trained spatiotemporal graph convolutional network, that is, the liquid level data of the unmonitored nodes at the time corresponding to the collected real-time liquid level monitoring data is output.

[0112] Specific implementation method 2: This implementation method is different from specific implementation method 1 in that the pipe network topology data includes the connection relationship between the inspection well, the pump station forebay and the river outlet and the pipeline, the pipeline diameter, the pipeline length and the pipeline elevation.

[0113] The other steps and parameters are the same as those in the first embodiment.

[0114] Specific implementation method three: This implementation method is different from specific implementation methods one or two in that the node-pipeline relationship matrix is:

[0115] Take each inspection well, pump station forebay and river outlet as a node and number each node (number each node 1, 2, ..., n in sequence, random numbering is OK):

[0116]

[0117] Among them, G is the node-pipeline relationship matrix, n represents the number of nodes in the drainage network system, and m represents the number of pipes in the drainage network system;

[0118]

[0119] Among them, G ij It is the element in the i-th row and j-th column of the node-pipeline relationship matrix G.

[0120] The other steps and parameters are the same as those in the first or second embodiment.

[0121] Specific implementation method 4: This implementation method is different from one of the specific implementation methods 1 to 3 in that the pipeline elevation information matrix is:

[0122]

[0123] Among them, H is the pipeline elevation information matrix;

[0124]

[0125] Among them, h j,us_invert represents the upstream invert elevation of pipe j, h j,ds_invert represents the downstream invert elevation of pipe j, H ij is the element in the i-th row and j-th column of the matrix H.

[0126] The other steps and parameters are the same as those in Specific Embodiments 1 to 3.

[0127] Specific implementation method 5: This implementation method is different from one of the specific implementation methods 1 to 4 in that the node relationship adjacency matrix is:

[0128]

[0129] Among them, A is the node relationship adjacency matrix;

[0130]

[0131] Among them, Aij is the element in the i-th row and j-th column of the node relationship adjacency matrix A.

[0132] The other steps and parameters are the same as those in Specific Embodiments 1 to 4.

[0133] Specific implementation method 6: This implementation method is different from any one of the specific implementation methods 1 to 5 in that the pipe diameter information matrix is:

[0134] Step 1: Set the search depth δ, search all nodes within the search depth range with node i as the root node, and record the set of searched nodes as node data set N δ , computing node data set N δ The standard deviation d of the diameter of all pipes formed by the internal node connection std :

[0135]

[0136] Among them, d k is the node dataset N δ The diameter of the kth pipeline formed by the internal node connection, d is the node data set N δ The average diameter of all pipes formed by the internal node connection, N is the node data set N δ The total number of pipelines formed by connecting internal nodes;

[0137] Step 2: Node i to node data set N δ The pipe diameter weight f of internal node j w (d ij )for:

[0138] (1) If there is only one path from node i to node j, then

[0139]

[0140] Wherein, α is a weight distribution parameter (in the present invention, the value is α=0.5); d qj is the diameter of the pipe connecting node q and node j; q represents the node one depth level above node j on the path between node i and node j (q may be node i itself); e is the base of the natural logarithm; η qj is the depth level of the pipeline connecting node q and node j (in this case η qj =1, the pipeline between the root node and the first-level leaf node is considered to be the first depth level, the pipeline between the second-level leaf node and the third-level leaf node is considered to be the second depth level, and so on);

[0141] (2) If there are at least two paths from node i to node j, then

[0142]

[0143] Where q represents a node one depth level above node j on any path between node i and node j; d qj represents the diameter of the pipe connecting node q and node j; qj is the depth level of the pipeline connecting node j to node q; f w (d qj ) is the pipe diameter weight from node q to node j;

[0144]

[0145] Where Q is the set of nodes located one depth level above node j on all paths from node i to node j; |Q| is the number of nodes in set Q;

[0146] Step 3: D ij As the element of row i and column j in the pipe diameter information matrix D;

[0147] Step 4: After processing each node in the drainage pipe network system respectively using the methods of steps 1 to 3 (i.e., taking each node as the root node), obtain the pipe diameter information matrix D:

[0148]

[0149] The other steps and parameters are the same as those in Specific Implementation Methods 1 to 5.

[0150] The node search method in this embodiment is similar to a tree diagram. The search direction is the flow direction of the liquid. Starting from the root node, after obtaining all leaf nodes from the 1st level to the δth level, all the leaf nodes obtained are combined into a node data set N δ .

[0151] Specific implementation method 7: This implementation method is different from any one of specific implementation methods 1 to 6 in that the pipe length information matrix is:

[0152] Step 1: Set the search depth δ, search all nodes within the search depth range with node i as the root node, and record the set of searched nodes as node data set N δ , computing node data set N δ The standard deviation of the length of all pipes formed by the internal node connection is l std :

[0153]

[0154] Among them, l k is the node dataset N δThe length of the kth pipeline formed by the connection of the internal nodes, is the node dataset N δ The average length of all pipes formed by the internal node connection, N node data set N δ The total number of pipelines formed by connecting internal nodes;

[0155] Step 2: Node i to node data set N δ The tube length weight f of internal node j w (l ij )for:

[0156] (1) If there is only one path from node i to node j, then

[0157]

[0158] Wherein, α is a weight distribution parameter (the value in the present invention is α=0.5); l qj is the length of the pipe connecting node q and node j; q represents the node one depth level above node j on the path from node i to node j; e is the base of the natural logarithm; η qj is the depth level of the pipeline connecting node q and node j (in this case η qj =1);

[0159] (2) If there are at least two paths from node i to node j, then

[0160]

[0161] Where q represents a node one depth level above node j on any path between node i and node j; d qj represents the length of the pipe connecting node q and node j; qj is the depth level of the pipe connecting node j to node q;

[0162]

[0163] Where Q is the set of nodes located one depth level above node j on all paths from node i to node j; |Q| is the number of nodes in set Q;

[0164] Step 3: L ij As the element of row i and column j in the pipe length information matrix L;

[0165] Step 4: After processing each node in the drainage pipe network system respectively using the methods of steps 1 to 3 (i.e., taking each node as the root node), the pipe length information matrix L is obtained:

[0166]

[0167] The other steps and parameters are the same as those in Specific Embodiments 1 to 6.

[0168] Specific implementation eight: This implementation differs from one of specific implementations one to seven in that the liquid level data matrix of all nodes is obtained based on the historical liquid level data of each monitoring node, specifically:

[0169] Sort the historical liquid level data monitored by the i-th monitoring node in time series to form X p vector:

[0170] X i =[X i1 ,X i2 …,X iT’ ]

[0171] Among them, X it is the vector X i The t-th monitoring value in, that is, the t-th monitoring value monitored by the i-th monitoring node, T' is the total number of monitoring values, t = 1, 2, ..., T';

[0172] Then the data matrix of all node liquid levels is X:

[0173]

[0174] Among them, the superscript T represents the transpose of the matrix;

[0175] In the matrix X, the elements of the rows corresponding to the nodes that are not monitored are set to 0.

[0176] The other steps and parameters are the same as those in Specific Embodiments 1 to 7.

[0177] Specific implementation method 9: This implementation method is different from specific implementation methods 1 to 8 in that the specific process of step 4 is as follows:

[0178] Step 4: According to the number of monitoring nodes M, input the number of monitoring nodes M in : Number of monitoring nodes for loss calculation M loss =1:1 ratio, generate monitoring node combination library S (where m in ∪M loss =M,M in ∩M loss =0);

[0179] That is, for each monitoring node combination in the monitoring node combination library S, the monitoring node is divided into two parts: input monitoring node and loss calculation monitoring node, and the combination library S includes all possible monitoring node combinations;

[0180] For each monitoring node combination in the monitoring node combination library S, a training sample is generated correspondingly, specifically:

[0181] For any monitoring node combination in the monitoring node combination library S, only the liquid level data corresponding to the input monitoring node selected by this monitoring node combination is retained in the new liquid level data matrix X′, and the liquid level data of other monitoring nodes are set to 0. A training sample is formed using the node-pipeline relationship matrix, pipeline elevation information matrix, node relationship adjacency matrix, pipe diameter information matrix, pipe length information matrix and the new liquid level data matrix X′;

[0182] Step 42: Slidingly segment each training sample generated in step 41 according to a time window of a specified size to obtain a training sample library;

[0183] For example, for a node including 20 monitoring data, the time window length is set to 10 monitoring data, then the 1st monitoring data to the 10th monitoring data are used as a training sample, the 2nd monitoring data to the 11th monitoring data are used as a training sample, and so on, the 11th monitoring data to the 20th monitoring data are used as a training sample;

[0184] Step 43: Initialize the maximum number of training cycles to process, and initialize the training sample update frequency of each training cycle to epoch;

[0185] Step 44: Set the number of training cycles p = 1;

[0186] Step 45: randomly select a monitoring node combination from the monitoring node combination library S;

[0187] Step 46: Initialize e=1;

[0188] Step 47: randomly obtain one training sample from the training sample set corresponding to the monitoring node combination selected in step 45, and use the selected current sample to train the spatiotemporal graph convolutional neural network;

[0189] The loss function calculation method during training is as follows: using the spatiotemporal graph convolutional neural network to output the predicted results of the liquid level data of the monitoring nodes that are set to 0 in the current sample, and calculating the loss function based on the predicted results and the actual liquid level data;

[0190] It should be noted that the spatiotemporal graph convolutional neural network can reconstruct the liquid level values ​​corresponding to all 0 elements in the output liquid level data matrix, but what needs to be used in the training process is the reconstruction results corresponding to the monitoring node data set to 0 and the actual data corresponding to the monitoring nodes set to 0;

[0191] Step 48: Determine whether e=epoch is satisfied;

[0192] If satisfied, execute step 49;

[0193] If not satisfied, set e=e+1 and return to step 47;

[0194] Step 49: determine whether p=process is satisfied;

[0195] If satisfied, the entire training process ends and a trained spatiotemporal graph convolutional neural network is obtained;

[0196] If not satisfied, set p=p+1 and return to step 4 or 5.

[0197] The other steps and parameters are the same as those in Specific Embodiments 1 to 8.

[0198] Specific implementation method ten: This implementation method is different from any one of specific implementation methods one to nine in that the feature extraction unit includes a GNN and a fully connected layer, and the specific working process is as follows:

[0199] Take X′ as the input of the fully connected layer, and then take the output f(X′) of X′ after the fully connected layer as the input of GNN:

[0200] X′ D =GNN((W d *D+b d )⊙f(X′)),A)

[0201] X′ L =GNN((W l *L+b l )⊙f(X′)),A)

[0202] X′ H =GNN((W h *H+b h )⊙f(X′)),G)

[0203] Among them, W d , W l and W h is the weight of GNN, b d , b l and b h is the bias of GNN, * represents element-wise multiplication (Hadamard product), ⊙ represents matrix product, X′ D , X′ L and X′ H is the output of the feature extraction unit.

[0204] The other steps and parameters are the same as those in Specific Embodiments 1 to 9.

[0205] The feature extraction unit uses the pipeline elevation information matrix, the pipe diameter information matrix and the pipe length information matrix to extract feature information that satisfies the hydraulic change pattern from the liquid level information data of the limited monitoring nodes.

[0206] Example

[0207] Take a drainage network system in City M as an example. The drainage network system consists of 534 nodes (including 532 inspection wells, 1 pump station forebay, and 1 river outlet) and 539 pipe sections. The drainage capacity of the terminal rainwater pump station is 10.72m 3 / s. The system covers a total area of ​​172ha, with a pipe length of 12514m and a pipe diameter range of DN300-DN2400. There are 7 inspection wells where sensors are installed.

[0208] like Figure 1 As shown, it specifically relates to a method for reconstructing global unmonitored node liquid level data based on limited monitoring node liquid level monitoring data; comprising the following steps:

[0209] Step 1: Collect the network topology data (such as GIS data) of the drainage network system and the historical liquid level data of each monitoring node throughout 2020;

[0210] The pipe network topology data includes the connection relationship between the inspection well, the pump station forebay and the river outlet and the pipeline, the pipeline diameter, the pipeline length and the pipeline elevation.

[0211] Step 2: Obtain the node-pipeline relationship matrix, pipeline elevation information matrix, node relationship adjacency matrix, pipe diameter information matrix, and pipe length information matrix based on the pipe network topology data; obtain the liquid level data matrix of all nodes based on the historical liquid level data of each monitoring node;

[0212] The node-pipeline relationship matrix is:

[0213] Take each inspection well, pump station forebay and river outlet as a node and number each node (number each node 1, 2, ..., n in sequence, random numbering is OK):

[0214]

[0215] Among them, G is the node-pipeline relationship matrix, n represents the number of nodes in the drainage network system, and m represents the number of pipes in the drainage network system;

[0216]

[0217] Among them, G ij It is the element in the i-th row and j-th column of the node-pipeline relationship matrix G.

[0218] The pipeline elevation information matrix is:

[0219]

[0220] Among them, H is the pipeline elevation information matrix;

[0221]

[0222] Among them, h j,us_invert represents the upstream invert elevation of pipe j, h j,ds_invert represents the downstream invert elevation of pipe j, H ij is the element in the i-th row and j-th column of the matrix H.

[0223] The node relationship adjacency matrix is:

[0224]

[0225] Among them, A is the node relationship adjacency matrix;

[0226]

[0227] Among them, A ij is the element in the i-th row and j-th column of the node relationship adjacency matrix A.

[0228] The pipe diameter information matrix is:

[0229] Step 1: Set the search depth δ = 15, search all nodes within the search depth range with node i as the root node, and record the set of searched nodes as node data set N δ , computing node data set N δ The standard deviation d of the diameter of all pipes formed by the internal node connection std :

[0230]

[0231] Among them, d k is the node dataset N δ The diameter of the kth pipeline formed by the internal node connection, d is the node data set N δ The average diameter of all pipes formed by the internal node connection, N is the node data set N δ The total number of pipelines formed by connecting internal nodes;

[0232] Step 2: Node i to node data set N δ The pipe diameter weight f of internal node j w (d ij )for:

[0233] (1) There is only one path from node i to node j, then

[0234]

[0235] Wherein, α is a weight distribution parameter (in the present invention, the value is α=0.5); d qj is the diameter of the pipe connecting node q and node j; q represents the node one depth level above node j on the path between node i and node j (q may be node i itself); e is the base of the natural logarithm; η qj is the depth level of the pipeline connecting node q to node j;

[0236] (2) If there are at least two paths from node i to node j, then

[0237]

[0238] Where q represents a node one depth level above node j on any path between node i and node j, and d qj represents the diameter of the pipe connecting node q and node j, η qj is the depth level of the pipeline connecting node j to node q, f w (d qj ) is the pipe diameter weight from node q to node j;

[0239]

[0240] Where Q is the set of nodes located one depth level above node j on all paths from node i to node j, and |Q| is the number of nodes in set Q;

[0241] Step 3: D ij As the element of row i and column j in the pipe diameter information matrix D;

[0242] Step 4: After processing each node in the drainage pipe network system respectively using the methods of steps 1 to 3 (i.e., taking each node as the root node), obtain the pipe diameter information matrix D:

[0243]

[0244] The pipe length information matrix is:

[0245] Step 1: Set the search depth δ = 15, search all nodes within the search depth range with node i as the root node, and record the set of searched nodes as node data set N δ , computing node data set N δ The standard deviation of the length of all pipes formed by the internal node connection is l std :

[0246]

[0247] Among them, l k is the node dataset N δ The length of the jth pipeline formed by the connection of the internal nodes, is the node dataset N δ The average length of all pipes formed by the internal node connection, N node data set N δ The total number of pipelines formed by connecting internal nodes;

[0248] Step 2: Node i to node data set N δ The tube length weight f of internal node j w (l ij )for:

[0249] (1) If there is only one path from node i to node j, then

[0250]

[0251] Wherein, α is a weight distribution parameter (the value in the present invention is α=0.5); l qj is the length of the pipe connecting node q and node j; q represents the node one depth level above node j on the path from node i to node j; e is the base of the natural logarithm; η qj is the depth level of the pipeline connecting node q and node j (in this case η qj =1);

[0252] (2) If there are at least two paths from node i to node j, then

[0253]

[0254] Where q represents a node one depth level above node j on any path between node i and node j, and d qj represents the length of the pipe connecting node q and node j, η qj is the depth level of the pipe connecting node j to node q;

[0255] Where Q is the set of nodes located one depth level above node j on all paths from node i to node j, and |Q| is the number of nodes in set Q;

[0256] Step 3: L ij As the element of row i and column j in the pipe length information matrix L;

[0257] Step 4: After processing each node in the drainage pipe network system respectively using the methods of steps 1 to 3 (i.e., taking each node as the root node), the pipe length information matrix L is obtained:

[0258]

[0259] Sort the historical liquid level data monitored by the i-th monitoring node in time series to form X p vector:

[0260] X i =[X i1 ,X i2 …,X iT’ ]

[0261] Among them, X it is the vector X i The tth monitoring value in , t=1,2,…,T', T' is the total number of monitoring values;

[0262] Then the data matrix of all node liquid levels is X:

[0263]

[0264] In the matrix X, the elements of the rows corresponding to the nodes that are not monitored are set to 0.

[0265] Step 3: Build a spatiotemporal graph convolutional neural network. The working process of the spatiotemporal graph convolutional neural network is as follows:

[0266] In the spatiotemporal graph convolutional neural network, the input data passes through the input layer, feature extraction unit, information aggregation transmission unit (GNN), feature fusion layer (Conv1D), time series recurrent unit (GRU), dense layer (Dense) and output layer in sequence; that is,

[0267] The input data is first input to the feature extraction unit through the input layer, and then the output of the feature extraction unit is used as the input of the information aggregation transmission unit;

[0268] The output of the information aggregation transfer unit is used as the input of the feature fusion layer, and the output of the feature fusion layer is used as the input of the time series recursive unit;

[0269] The output of the time series recurrent unit is used as the input of the dense layer, and the output of the dense layer is output through the output layer;

[0270] Among them, the feature extraction unit includes GNN and fully connected layer, and the specific working process is:

[0271] Take X′ as the input of the fully connected layer, and then take the output f(X′) of X′ after the fully connected layer as the input of GNN:

[0272] X′ D =GNN((W d *D+b d)⊙f(X′)),A)

[0273] X′ L =GNN((W l *L+b l )⊙f(X′)),A)

[0274] X′ H =GNN((W h *H+b h )⊙f(X′)),G)

[0275] Among them, W d , W l and W h is the weight of GNN, b d , b l and b h is the bias of GNN, * represents element-wise multiplication (Hadamard product), ⊙ represents matrix product, X′ D , X′ L and X′ H is the output of the feature extraction unit.

[0276] Step 4: Use the node-pipeline relationship matrix, pipeline elevation information matrix, node relationship matrix, pipe diameter information matrix, pipe length information matrix and all node liquid level data matrices obtained in step 2 to train the spatiotemporal graph convolutional neural network;

[0277] In step 4, the spatiotemporal graph convolutional network PyGNN is trained through the blunt masking training strategy. The specific training process is as follows:

[0278] Step 4: According to the number of monitoring nodes M, input the number of monitoring nodes M in : Number of monitoring nodes for loss calculation M loss =1:1 ratio, generate monitoring node combination library S (where M in ∪M loss =M,M in ∩M loss =0);

[0279] That is, for each monitoring node combination in the monitoring node combination library S, the monitoring node is divided into two parts: input monitoring node and loss calculation monitoring node, and the combination library S includes all possible monitoring node combinations;

[0280] For each monitoring node combination in the monitoring node combination library S, a training sample is generated correspondingly, specifically:

[0281] For any monitoring node combination in the monitoring node combination library S, only the liquid level data corresponding to the input monitoring node selected by this monitoring node combination is retained in the new liquid level data matrix X′, and the liquid level data of other monitoring nodes are set to 0. A training sample is formed using the node-pipeline relationship matrix, pipeline elevation information matrix, node relationship adjacency matrix, pipe diameter information matrix, pipe length information matrix and the new liquid level data matrix X′;

[0282] Since sensors are deployed at 7 monitoring nodes, the ratio of the number of selected monitoring nodes to the number of unselected monitoring nodes cannot be evenly distributed, so the monitoring nodes are distributed in a ratio of 4:3, that is, 3 / 7 nodes are selected from the monitoring nodes, and then the elements of the corresponding rows of the selected monitoring nodes in the matrix X are set to 0;

[0283] In the present invention, since 3 monitoring nodes need to be randomly selected from 7 monitoring nodes each time, there are 35 possible selections in total, that is, the number of samples included in the training sample library S is 35.

[0284] Step 42: Slidingly segment each training sample generated in step 41 according to a time window of a specified size to obtain a training sample library;

[0285] Step 43: Initialize the maximum number of training cycles to process, and initialize the training sample update frequency of each training cycle to epoch;

[0286] Step 44: Set the number of training cycles p = 1;

[0287] Step 45: randomly select a monitoring node combination from the monitoring node combination library S;

[0288] Step 46: Initialize e=1;

[0289] Step 47: randomly obtain one training sample from the training sample set corresponding to the monitoring node combination selected in step 45, and use the selected current sample to train the spatiotemporal graph convolutional neural network;

[0290] The loss function calculation method during training is as follows: using the spatiotemporal graph convolutional neural network to output the predicted results of the liquid level data of the monitoring nodes that are set to 0 in the current sample, and calculating the loss function based on the predicted results and the actual liquid level data;

[0291] It should be noted that the spatiotemporal graph convolutional neural network can reconstruct the liquid level values ​​corresponding to all 0 elements in the output liquid level data matrix, but what needs to be used in the training process is the reconstruction results corresponding to the monitoring node data set to 0 and the actual data corresponding to the monitoring nodes set to 0;

[0292] Step 48: Determine whether e=epoch is satisfied;

[0293] If satisfied, execute step 49;

[0294] If not satisfied, set e=e+1 and return to step 47;

[0295] Step 49: determine whether p=process is satisfied;

[0296] If satisfied, the entire training process ends and a trained spatiotemporal graph convolutional neural network is obtained;

[0297] If not satisfied, set p=p+1 and return to step 4 or 5.

[0298] Step 5. Collect the real-time liquid level monitoring data of each monitoring node in the drainage pipe network system, obtain the real-time liquid level data matrix based on the real-time liquid level monitoring data, input the real-time liquid level data matrix and the node-pipe relationship matrix, pipe elevation information matrix, node relationship matrix, pipe diameter information matrix and pipe length information matrix of the drainage pipe network system into the trained spatiotemporal graph convolutional network, and output the liquid level data reconstruction results of the unmonitored nodes through the trained spatiotemporal graph convolutional network.

[0299] The above calculation examples of the present invention are only used to explain the calculation model and calculation process of the present invention in detail, and are not intended to limit the implementation methods of the present invention. For ordinary technicians in the relevant field, other different forms of changes or modifications can be made based on the above description. It is impossible to list all the implementation methods here. All obvious changes or modifications derived from the technical solution of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for reconstructing liquid level data of unmonitored nodes based on liquid level data of monitored nodes, characterized in that: The method specifically comprises the following steps: Step 1: Collect the pipe network topology data of the drainage pipe network system and the historical liquid level data of each monitoring node; Step 2: Obtain the node-pipeline relationship matrix, pipeline elevation information matrix, node relationship adjacency matrix, pipe diameter information matrix, and pipe length information matrix based on the pipe network topology data; obtain the liquid level data matrix of all nodes based on the historical liquid level data of each monitoring node; Step 3: Build a spatiotemporal graph convolutional neural network. The working process of the spatiotemporal graph convolutional neural network is as follows: In the spatiotemporal graph convolutional neural network, the input data passes through the input layer, feature extraction unit, information aggregation transmission unit, feature fusion layer, time series recursive unit, dense layer and output layer in sequence; that is, The input data is first input to the feature extraction unit through the input layer, and then the output of the feature extraction unit is used as the input of the information aggregation transmission unit; The output of the information aggregation transfer unit is used as the input of the feature fusion layer, and the output of the feature fusion layer is used as the input of the time series recursive unit; The output of the time series recurrent unit is used as the input of the dense layer, and the output of the dense layer is output through the output layer; Step 4: Use the node-pipeline relationship matrix, pipeline elevation information matrix, node relationship adjacency matrix, pipe diameter information matrix, pipe length information matrix and all node liquid level data matrices obtained in step 2 to train the spatiotemporal graph convolutional neural network; Step 5. Collect the real-time liquid level monitoring data of each monitoring node in the drainage pipe network system, obtain the real-time liquid level data matrix based on the real-time liquid level monitoring data, input the real-time liquid level data matrix and the node-pipe relationship matrix, pipe elevation information matrix, node relationship adjacency matrix, pipe diameter information matrix and pipe length information matrix of the drainage pipe network system into the trained spatiotemporal graph convolutional network, and output the liquid level data reconstruction results of the unmonitored nodes through the trained spatiotemporal graph convolutional network.

2. A method for reconstructing unmonitored node liquid level data based on monitored node liquid level data according to claim 1, characterized in that: The pipe network topology data includes the connection relationship between the inspection well, the pump station forebay and the river outlet and the pipeline, the pipeline diameter, the pipeline length and the pipeline elevation.

3. The method of reconstructing the liquid level data of unmonitored nodes based on the liquid level data of monitored nodes according to claim 2, characterized in that: The node-pipeline relationship matrix is: Take each inspection well, pump station forebay and river outlet as a node and number each node: Among them, G is the node-pipeline relationship matrix, n represents the number of nodes in the drainage network system, and m represents the number of pipes in the drainage network system; Among them, G ij It is the element in the i-th row and j-th column of the node-pipeline relationship matrix G.

4. The method for reconstructing the liquid level data of unmonitored nodes based on the liquid level data of monitored nodes according to claim 3, characterized in that: The pipeline elevation information matrix is: Among them, H is the pipeline elevation information matrix; Among them, h j,us_invert represents the upstream invert elevation of pipe j, h j,ds_invert represents the downstream invert elevation of pipe j, H ij is the element in the i-th row and j-th column of the matrix H.

5. The method for reconstructing the liquid level data of unmonitored nodes based on the liquid level data of monitored nodes according to claim 4, characterized in that: The node relationship adjacency matrix is: Among them, A is the node relationship adjacency matrix; Among them, A ij is the element in the i-th row and j-th column of the node relationship adjacency matrix A.

6. A method for reconstructing unmonitored node liquid level data based on monitored node liquid level data according to claim 5, characterized in that: The pipe diameter information matrix is: Step 1: Set the search depth δ, search all nodes within the search depth range with node i as the root node, and record the set of searched nodes as node data set N δ , computing node data set N δ The standard deviation d of the diameter of all pipes formed by the internal node connection std : Among them, d k is the node dataset N δ The diameter of the kth pipeline formed by the connection of the internal nodes, is the node dataset N δ The average diameter of all pipes formed by the internal node connection, N is the node data set N δ The total number of pipelines formed by connecting internal nodes; Step 2: Node i to node data set N δ The pipe diameter weight f of internal node j w (d ij )for: (1) If there is only one path from node i to node j, then Among them, α is the weight distribution parameter; d qj is the diameter of the pipe connecting node q and node j; q represents the node one depth level above node j on the path from node i to node j; e is the base of the natural logarithm; η qj is the depth level of the pipeline connecting node q to node j; (2) If there are at least two paths from node i to node j, then Where q represents a node one depth level above node j on any path between node i and node j; d qj represents the diameter of the pipe connecting node q and node j; qj is the depth level of the pipeline connecting node j to node q; f w (d qj ) is the pipe diameter weight from node q to node j; Where Q is the set of nodes located one depth level above node j on all paths from node i to node j; |Q| is the number of nodes in set Q; Step 3: D ij As the element of row i and column j in the pipe diameter information matrix D; Step 4: After processing each node in the drainage pipe network system using the methods from step 1 to step 3, the pipe diameter information matrix D is obtained:

7. The method of reconstructing the liquid level data of unmonitored nodes based on the liquid level data of monitored nodes according to claim 5, characterized in that: The pipe length information matrix is: Step 1: Set the search depth δ, search all nodes within the search depth range with node i as the root node, and record the set of searched nodes as node data set N δ , computing node data set N δ The standard deviation of the length of all pipes formed by the internal node connection is l std : Among them, l k is the node dataset N δ The length of the kth pipeline formed by the connection of the internal nodes, is the node dataset N δ The average length of all pipes formed by the internal node connection, N node data set N δ The total number of pipelines formed by connecting internal nodes; Step 2: Node i to node data set N δ The tube length weight f of internal node j w (l ij )for: (1) If there is only one path from node i to node j, then Among them, α is the weight distribution parameter; l qj is the length of the pipe connecting node q and node j; q represents the node one depth level above node j on the path from node i to node j; e is the base of the natural logarithm; η qj is the depth level of the pipeline connecting node q to node j; (2) If there are at least two paths from node i to node j, then Where q represents a node one depth level above node j on any path between node i and node j; d qj represents the length of the pipe connecting node q and node j; qj is the depth level of the pipe connecting node j to node q; Where Q is the set of nodes located one depth level above node j on all paths from node i to node j; |Q| is the number of nodes in set Q; Step 3: L ij As the element of row i and column j in the pipe length information matrix L; Step 4: After processing each node in the drainage pipe network system using the methods of steps 1 to 3, the pipe length information matrix L is obtained:

8. A method for reconstructing liquid level data of unmonitored nodes based on liquid level data of monitored nodes according to claim 6 or 7, characterized in that: The liquid level data matrix of all nodes is obtained according to the historical liquid level data of each monitoring node, specifically: Sort the historical liquid level data monitored by the i-th monitoring node in time series to form X p vector: X i =[X i1 ,X i2 …,X iT’ ] Among them, X it is the vector X i The t-th monitoring value in , that is, the t-th monitoring value monitored by the ith monitoring node, T' is the total number of monitoring values, t = 1, 2, ..., T'; Then the data matrix of all node liquid levels is X: Among them, the superscript T represents the transpose of the matrix; In the matrix X, the elements of the rows corresponding to the nodes that are not monitored are set to 0.

9. The method for reconstructing the liquid level data of unmonitored nodes based on the liquid level data of monitored nodes according to claim 8, characterized in that: The specific process of step 4 is as follows: Step 41: According to the number of monitoring nodes M, input the number of monitoring nodes M in : Number of monitoring nodes for loss calculation M loss =1:1 ratio, generate monitoring node combination library S; That is, for each monitoring node combination in the monitoring node combination library S, the monitoring node is divided into two parts: input monitoring node and loss calculation monitoring node; For each monitoring node combination in the monitoring node combination library S, a training sample is generated correspondingly, specifically: For any monitoring node combination in the monitoring node combination library S, only the liquid level data corresponding to the input monitoring node selected by this monitoring node combination is retained in the new liquid level data matrix X′, and the liquid level data of other monitoring nodes are set to 0. A training sample is formed using the node-pipeline relationship matrix, pipeline elevation information matrix, node relationship adjacency matrix, pipe diameter information matrix, pipe length information matrix and the new liquid level data matrix X′; Step 42: Slidingly segment each training sample generated in step 41 according to a time window of a specified size to obtain a training sample library; Step 43: Initialize the maximum number of training cycles to process, and initialize the training sample update frequency of each training cycle to epoch; Step 44: Set the number of training cycles p = 1; Step 45: randomly select a monitoring node combination from the monitoring node combination library S; Step 46: Initialize e=1; Step 47: randomly obtain one training sample from the training sample set corresponding to the monitoring node combination selected in step 45, and use the selected current sample to train the spatiotemporal graph convolutional neural network; The loss function calculation method during training is as follows: using the spatiotemporal graph convolutional neural network to output the predicted results of the liquid level data of the monitoring nodes that are set to 0 in the current sample, and calculating the loss function based on the predicted results and the actual liquid level data; Step 48: Determine whether e=epoch is satisfied; If satisfied, execute step 49; If not satisfied, set e=e+1 and return to step 47; Step 49: determine whether p=process is satisfied; If satisfied, the entire training process ends and a trained spatiotemporal graph convolutional neural network is obtained; If not satisfied, set p=p+1 and return to step 4 or 5.

10. The method for reconstructing the liquid level data of unmonitored nodes based on the liquid level data of monitored nodes according to claim 9, characterized in that: The feature extraction unit includes GNN and a fully connected layer, and the specific working process is as follows: Take X′ as the input of the fully connected layer, and then take the output f(X′) of X′ after the fully connected layer as the input of GNN: X′ D =GNN((W d *D+b d )⊙f(X′)),A) X′ L =GNN((W l *L+b l )⊙f(X′)),A) X′ H =GNN((W h *H+b h )⊙f(X′)),G) Among them, W d , W l and W h is the weight of GNN, b d , b l and b h is the bias of GNN, * represents element-wise multiplication, ⊙ represents matrix product, X′ D , X′ L and X′ H is the output of the feature extraction unit.

Images