Pump station group regulation and control method based on drainage system global water volume prediction

Abstract

The invention belongs to the field of drainage system intelligent management and control, and particularly relates to a pump station group regulation and control method based on drainage system global water volume prediction. The method comprises the following steps: acquiring original data sets of a pump station, a monitoring node and a sewage treatment plant; data cleaning and vacancy filling are carried out on the original data set, normalization processing is carried out on the original data set, and meanwhile structured storage is carried out; establishing a system topological relation according to the actual structure of the drainage system; constructing a model architecture based on a graph neural network, determining an extraction scheme of time sequence and spatial topology information, and training based on a historical data set; the trained model is evaluated, and model parameters with errors meeting expectations are stored; and finally, establishing a pump station group regulation and control strategy through a drainage system global water volume prediction model.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent control of drainage systems, and specifically relates to a pump station group control method based on global water volume prediction of the drainage system. Background Art

[0002] With the acceleration of urbanization, the stable operation of drainage systems is crucial to the development of cities and the quality of life of residents. The operation of drainage systems, especially the joint control of pumping station groups, relies on accurate water volume prediction in order to rationally allocate resources, prevent urban flooding disasters, and ensure the balance of water supply and drainage. However, traditional drainage system water volume prediction methods often rely on simplified physical models or statistical methods, which have certain limitations when dealing with complex and dynamic drainage systems, especially the inability to consider the impact of topological structures on water volume convergence, which has a greater impact.

[0003] In recent years, the development of artificial intelligence technology has provided new ideas for drainage system water volume prediction. In particular, the emergence of Graph Neural Networks (GNNs) has provided a powerful tool for processing complex network structure data. Graph neural networks can effectively learn the relationship between nodes, and perform information transmission and feature extraction in the graph structure. They have significant advantages in processing drainage system data with spatial dependence and topological structure. Therefore, a drainage system water volume prediction method based on graph neural networks is urgently needed. Summary of the invention

[0004] This method proposes a pump station group control method based on the global water volume prediction of the drainage system. By constructing the graph structure of the drainage system, the drainage area is divided into nodes and the drainage pipes are divided into edges, forming a graph model that can reflect the spatial dependence and topological structure of the drainage system. On this basis, the graph neural network is used to learn the relationship between nodes, extract the spatiotemporal characteristics of the system, and realize the global water volume prediction through the training model. Finally, the pump station group control strategy is formed based on a reliable water volume prediction model.

[0005] Specifically, the method comprises the following steps:

[0006] Step S1: Obtain the original data set of pumping stations, monitoring nodes, and sewage treatment plants;

[0007] Step S2: Clean and fill in the gaps of the original data set, normalize it, and save it in a structured manner;

[0008] Step S3: Establishing the system topology relationship according to the actual structure of the drainage system;

[0009] Step S4: Build a model architecture based on graph neural network, determine the extraction scheme of time series and spatial topology information, and train it based on historical data sets;

[0010] Step S5: Evaluate the trained model and save the model parameters whose errors meet the expectations;

[0011] Step S6: Establish a pump station group control strategy through the global water volume prediction model of the drainage system, and ultimately achieve the goals of safe operation and energy saving and consumption reduction.

[0012] Furthermore, the original data set includes a time series data set of nodes and a spatial topology data set of nodes;

[0013] Time series data set of nodes: pump station, monitoring node, name of sewage treatment plant, operating liquid level or operating flow, facility load rate, rainfall data; data sampling frequency is not less than 5 minutes;

[0014] Spatial topological data set of nodes: geographic location information of pumping stations, monitoring nodes, and sewage treatment plants, pipeline connection information (including pipe diameter, slope, and flow rate), and rainwater and sewage mixing data are also required in the diversion area.

[0015] Furthermore, the original data set is saved in a structured manner, including:

[0016] Input data: After the original data is sorted and divided, it is saved in a three-dimensional matrix format. 1 , x 2 , x 3 …, x t ), X is an n×f×t dimensional matrix, where t is the length of the time slice. Each x i It is an n×f dimensional matrix, where n represents the total number of key nodes such as pumping stations, monitoring nodes, and sewage treatment plants, and f represents the number of input parameters in each node. Figure 2 .

[0017] The present invention intends to make predictions at two levels: daily level and weekly level, so the input data is finally organized as follows:

[0018] 1) Designed with a sampling frequency of 5 min, the number of samples per day is 288;

[0019] 2) Designed with a sampling frequency of 5 min, the number of samples in a week is 2016;

[0020] Target data: After sorting and segmenting the original data, extract the flow data of each key node as the target data and save it in a three-dimensional matrix format, Y = (y 1 ,y 2,y 3 …, y t ), Y is an n×1×t dimensional matrix, where t is the length of the time slice. Each y i It is an n×1 dimensional matrix, where n represents the total number of key nodes such as pumping stations, monitoring nodes, and sewage treatment plants, and 1 represents the number of target data in each node. The present invention only predicts water volume data, so 1 is taken.

[0021] The target data uses the water volume values ​​of each node 2 hours after the input data, so Y T =(y T+1 ,y T+2 ,y T+3 …, y T+23 )=(x T+1 [flow],x T+2 [flow],x T+3 [flow]…, x T+23 [flow])

[0022] Furthermore, according to the actual structure of the drainage system, the system topology relationship is established, including:

[0023] The system topology relationship is established as follows: G = (V, E, A), where V represents the node, obtained in step S2, and E represents the directed edge information, which is an adjacency matrix used to represent the connection relationship between nodes;

[0024] E uses a two-dimensional matrix:

[0025]

[0026] in i In the middle, it represents the upstream node, b i Pointing to the downstream node, c i Represents the topological weight between upstream and downstream;

[0027] where c i is a 4×1 dimensional matrix:

[0028]

[0029] Record the pipeline specifications, length, slope, and degree of regional mixing between key nodes respectively.

[0030] Adjacency matrix element a ij ∈A, defined as

[0031] Furthermore, we built a model architecture based on graph neural network, determined the extraction scheme of time series and spatial topology information, and trained it based on historical data sets, including:

[0032] Build a model architecture based on graph neural network. The model architecture consists of two parallel computing links. The computing process modules in the computing links are consistent. The only difference is that the input data is daily level data and weekly level data. The specific computing process modules include attention mechanism module, spatial domain graph convolution GCN, and temporal domain convolution TConv. After the calculation is completed, the output vectors of the two modules are summarized through a fully connected neural network layer to produce the prediction result:

[0033] 1) Attention mechanism module

[0034] Using the self-attention mechanism, for the input vector X = (x 1 , x 2 , x 3 …, x t )(n×f×t dimensional matrix), calculate its query (Query), key (Key) and value (Vlue) respectively:

[0035] Q=W Q × i ;

[0036] K=W k × i ;

[0037] V=W v × i ;

[0038] Perform a dot product operation on Query and Key, then divide by a normalization factor, and finally normalize through the softmx function:

[0039]

[0040] 2) Spatial Graph Convolution GCN

[0041] Perform the classic spectral domain graph convolution operation on the spatial domain, and perform the input vector X = (x 1 , x 2 , x 3 …, x t )(n×f×t dimensional matrix), perform graph Fourier transform and obtain spatial eigenvalues:

[0042]

[0043] Where x is input data;

[0044] *G — graph convolution operation;

[0045] L—Laplace moment;

[0046] K—Use K-order Chebyshev polynomial approximation;

[0047] T k — Chebyshev polynomials:

[0048] β k — Chebyshev polynomials can learn parameters.

[0049] 3) Time Domain Convolution TConv

[0050] Perform time domain convolution TConv operation on the time series data of each node in the spatial domain graph convolution result, extract the time series features and output them through a fully connected neural network to provide evidence for the traffic of each node in the next 2 hours;

[0051] The time domain convolution TConv consists of two identical residual network modules connected in series; each residual network consists of a dilated convolution layer and a pooling layer;

[0052]

[0053] ψ* — convolution kernel of the dilated convolution layer;

[0054] βdown() — pooling layer operation;

[0055] b′, b″ — trainable bias terms for dilated convolutional and pooling layers;

[0056] The residual network module is followed by a fully connected neural network. The residual network settlement result is expanded in one dimension and then operated through a multi-layer neural network. The output is:

[0057]

[0058] It is an n×1×t-dimensional moment, where n represents the total number of key nodes such as pumping stations, monitoring nodes, and sewage treatment plants, 1 represents the flow data predicted by each node, t represents the prediction time step, and the present invention adopts 2 hours, t=24; d represents the daily level prediction result, and l represents l batches of training samples.

[0059] 4) The fully connected neural network:

[0060] Perform Hadamard product calculation on the daily and weekly forecast results to obtain the final forecast results.

[0061]

[0062] W d , W w —Learning parameters, trained through the learning process;

[0063] ⊙—Hadamard product operation;

[0064] —Daily level forecast results and weekly level forecast results;

[0065] —Final prediction results.

[0066] Furthermore, the trained model is evaluated and the model parameters whose errors meet the expectations are saved, including:

[0067] The model horizontal comparison method is used to compare and evaluate the RMSE (root mean square error) and MAE (mean absolute error) of the four predictions of the method of the present invention, the HA model, the ARIMA model and the VAR model. If the prediction results of this method have no obvious advantages, it is necessary to adjust the selection of input variables, optimize the selection of the upstream and downstream weight matrices of the directed edges in the topological relationship, and then retrain the model.

[0068] Furthermore, a pump station group control strategy is established through the global water volume prediction model of the drainage system, including:

[0069] After encapsulating the global water volume prediction model of the drainage system, it is embedded in the smart water system to form a pump station group control strategy. Through the collection of on-site data, the water volume of all nodes in the system within the next 2 hours is predicted. At the same time, according to the prediction results, the target lifting volume of the pump group of each pump station is preset to achieve safe operation of the sewage system and energy saving and consumption reduction.

[0070] The method for predicting the global water volume of a drainage system of the present invention has the following beneficial effects:

[0071] This method can process complex drainage system data, fully consider spatial dependence and topological structure, and improve prediction accuracy. The model based on graph neural network has strong generalization ability and can be applied to drainage systems of different scales and characteristics. By updating data and models in real time, it can realize dynamic water volume prediction and provide strong support for the operation and management of drainage systems. Compared with traditional physical models and statistical methods, this method has higher prediction efficiency and accuracy, which helps to improve the operation efficiency of drainage systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 It is a workflow diagram of a specific embodiment of the present invention;

[0073] Figure 2 is a conceptual diagram of data structure preprocessing of a specific embodiment of the present invention;

[0074] Figure 3 is an architectural diagram of a specific embodiment of the present invention;

[0075] Figure 4 It is a time domain convolution TConv module architecture diagram in a specific embodiment of the present invention;

[0076] Figure 5 It is a generalized drainage system diagram of the present invention. DETAILED DESCRIPTION

[0077] The present application will be further described in detail below in conjunction with the embodiments. It is to be understood that the specific embodiments described herein are only used to explain the relevant invention, rather than to limit the invention. It is also necessary to explain that, for ease of description, only the parts related to the relevant invention are shown in the drawings.

[0078] In the field of intelligent management and control of drainage systems, predicting the water volume of key nodes in the system through historical data is a prerequisite for making group control decisions of system pump stations. The present invention aims to solve the problem that traditional single or multi-variable models are difficult to simultaneously capture the time series and spatial topological correlation inherent in the basic data, and cannot effectively capture the above-mentioned characteristic information. It can perform efficient and accurate analysis on massive historical data and make highly feasible predictions for the future, thereby forming a method for regulating and controlling drainage system pump stations.

[0079] The originality of the present invention lies in creating the edge structure of the graph neural network through multi-dimensional spatial information; at the same time, the spatial and temporal features are extracted respectively through spatial domain graph convolution GCN and temporal domain convolution TConv.

[0080] This example takes a generalized model of a drainage system as an example. The generalized drainage system is composed of a regional sewage treatment plant (A3) and two pumping stations (A1 and A2) directly connected to the sewage treatment plant. The three-month measured data of the above drainage system is used for prediction.

[0081] Step S1, obtaining the original data set of the pump station, monitoring node, and sewage treatment plant;

[0082] This example collects the generalized system and the operating data of the three nodes for three months from February 16, 2024 to May 16, 2024, with a sampling frequency of 1 minute, totaling 129,600×3 data items.

[0083] Table 1 Original data sets of pumping stations, monitoring nodes, and sewage treatment plants

[0084]

[0085]

[0086] Step S2, cleaning and filling in gaps in the original data set, normalizing it, and saving it in a structured manner;

[0087] First, the original data is cleaned and filled to ensure that the data is complete and continuous, and then the frequency is reduced to 5 points per group, resulting in 25920×3 data.

[0088] The input data is organized into two levels: daily and weekly.

[0089] 1) The daily input data obtained 23,800 samples, each sample It is a 3×4×288 dimensional matrix

[0090] 2) Weekly input data yields 23,800 samples, each of which is a 3×4×2016 dimensional matrix

[0091] The target data has 23,800 samples, each sample Y T It is a 3×1×24 dimensional matrix.

[0092] 80% of the above samples are taken as training samples, and the remaining 20% ​​are taken as test samples.

[0093] Step S3, establishing a system topology relationship according to the actual structure of the drainage system;

[0094] Construct graph structure data G=(V, E, A), where V represents a node obtained in step S2, E represents directed edge information, and A is an adjacency matrix used to represent the connection relationship between nodes.

[0095] The edge vector E is:

[0096]

[0097] In this case, for ease of demonstration, the topology weights in the third row are simply represented by pipeline-related information.

[0098] The adjacency matrix A is:

[0099]

[0100] Step S4, constructing a model architecture based on a graph neural network, determining a scheme for extracting time series and spatial topology information, and performing training based on the data set;

[0101] 1) Attention mechanism module

[0102] Day level input sample It is a 3×4×288-dimensional matrix. After calculation, we get a 3×4×288-dimensional matrix of attention.

[0103] Weekly level input sample It is a 3×4×2016 dimensional matrix. After calculation, we get the 3×4×2016 dimensional matrix of full attention.

[0104] 2) Spatial Graph Convolution GCN

[0105] After the daily-level input samples and the weekly-level input samples are processed by the spatial graph convolution GCN, the aggregation results of the spatial features are obtained, and the output vector dimension does not change.

[0106] 3) Time Domain Convolution TConv

[0107] The convolution kernel specification of the expanded convolution layer is 64@3×3, the stride is 3, and the ReLU activation function is used; the pooling layer adopts maximum pooling; the pooling size is 2×2, and the stride is 2; a residual transcendence channel is set for the two layers.

[0108] The residual network module is followed by a fully connected neural network. The residual network settlement result is expanded in one dimension and then output through a multi-layer neural network:

[0109]

[0110] It is a 3×1×24 dimensional matrix, 3 represents 2 pumping stations and sewage treatment plants and 1 sewage treatment plant, 1 represents the flow data predicted by each node, and 24 represents the prediction time step, which is 2 hours in the present invention.

[0111] 4) Fully connected neural network

[0112] Perform Hadamard product calculation on the daily and weekly forecast results to obtain the final forecast results. Wd, Ww——Learning parameters, both are 3×1×24 dimensional matrices available for learning. It is a 3×1×24 dimensional matrix, representing the predicted traffic of the three key nodes in the next 2 hours.

[0113] Step S5, evaluate the trained model and save the model parameters whose errors meet the expectations.

[0114] The model horizontal comparison method was adopted to compare and evaluate the RMSE (root mean square error) and MAE (mean absolute error) of the four predictions of the method of the present invention, HA model, ARIMA model and VAR model.

[0115] Table 2 Model evaluation results

[0116] RMSE MAE HA Model 77.46 41.53 ARIMA Model 84.31 46.12 VAR Model 79.63 39.42 This method 57.89 27.56

[0117] Step S6: After encapsulating the global water volume prediction model of the drainage system, embed it into the smart water system to form a pump station group control strategy. Through the collection of on-site data, the water volume of all nodes in the system within the next 2 hours is predicted, and the target lifting volume of the pump group of each pump station is preset according to the prediction results.

[0118] The above embodiments are merely illustrative of the principles and effects of the present invention, and are not intended to limit the present invention. Anyone familiar with the art may modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by a person of ordinary skill in the art without departing from the spirit and technical concept disclosed by the present invention shall still be covered by the claims of the present invention.

Claims

1. A method for controlling a group of pump stations based on global water volume prediction of a drainage system, characterized in that: Step S1: Obtain the original data set of pumping stations, monitoring nodes, and sewage treatment plants; Step S2: Clean and fill in the gaps of the original data set, normalize it, and save it in a structured manner; Step S3: Establishing the system topology relationship according to the actual structure of the drainage system; Step S4: Build a model architecture based on graph neural network, determine the extraction scheme of time series and spatial topology information, and train it based on historical data sets; Step S5: Evaluate the trained model and save the model parameters whose errors meet the expectations; Step S6: Establish a pump station group control strategy through the global water volume prediction model of the drainage system, and ultimately achieve the goals of safe operation and energy saving and consumption reduction.

2. The method for controlling a group of pump stations based on global water volume prediction of a drainage system according to claim 1 is characterized in that: The original data set includes a time series data set of nodes and a spatial topology data set of nodes; Time series data sets of nodes: pumping station, monitoring node, sewage treatment plant name, operating liquid level or operating flow, facility load rate, rainfall data; Spatial topological data set of nodes: geographical location information of pumping stations, monitoring nodes, and sewage treatment plants, pipeline connection information, and rainwater and sewage mixing data in diversion areas.

3. The method for controlling a group of pump stations based on global water volume prediction of a drainage system according to claim 1 is characterized in that: The system topology relationship is established as follows: G = (V, E, A), where V represents the node, obtained in step S2, and E represents the directed edge information, which is an adjacency matrix used to represent the connection relationship between nodes; E uses a two-dimensional matrix: in i In the middle, it represents the upstream node, b i Pointing to the downstream node, c i Represents the topological weight between upstream and downstream; where c i is a 4×1 dimensional matrix: Record the pipeline specifications, length, slope, and degree of regional mixing between key nodes respectively.

4. The method for controlling a group of pump stations based on global water volume prediction of a drainage system according to claim 1 is characterized in that: Build a model architecture based on graph neural network. The model architecture consists of two parallel computing links. The computing process modules in the computing links are consistent. The only difference is that the input data is daily level data and weekly level data respectively. The specific calculation process modules include attention mechanism module, spatial domain graph convolution GCN, and temporal domain convolution TConv. Spatial domain graph convolution GCN and temporal domain convolution TConv are used to extract spatiotemporal features respectively. After the calculation is completed, the output vectors of the two modules are summarized and passed through a fully connected neural network layer to output the prediction results.

5. The method for controlling a group of pump stations based on global water volume prediction of a drainage system according to claim 4 is characterized in that: The attention mechanism module: Using the self-attention mechanism, for the input vector X = (x1, x2, x3..., x t )(n×f×t dimensional matrix), calculate its query (Query), key (Key) and value (Vlue) respectively: Q=W Q ×x i ; K=W k ×x i ; V=W v ×x i ; Perform a dot product operation on Query and Key, then divide by a normalization factor, and finally normalize through the softmx function:

6. The method for controlling a group of pump stations based on global water volume prediction of a drainage system according to claim 4 is characterized in that: The spatial domain graph convolution GCN: Perform the classic spectral domain graph convolution operation on the spatial domain, and perform the following operations on the input vector X = (x1, x2, x3…, x t )(n×f×t dimensional matrix), perform graph Fourier transform and obtain spatial eigenvalues: Where x is the input data; * G —Graph convolution operation; g θ —convolution kernel; L—Laplace moment; K—approximation using K-order Chebyshev polynomial; T k —Chebyshev polynomial; β k — Chebyshev polynomials can learn parameters.

7. The method for controlling a group of pump stations based on global water volume prediction of a drainage system according to claim 4, characterized in that: The time domain convolution TConv: Perform time domain convolution TConv operation on the time series data of each node in the spatial domain graph convolution result, extract the time series features and output them through a fully connected neural network to provide evidence for the traffic of each node in the next 2 hours; The time domain convolution TConv consists of two identical residual network modules connected in series; each residual network consists of a dilated convolution layer and a pooling layer; ψ* — convolution kernel of the dilated convolution layer; βdown() — pooling layer operation; b′, b″ — trainable bias terms for dilated convolutional and pooling layers; The residual network module is followed by a fully connected neural network. After the residual network settlement result is expanded in one dimension, it is calculated through a multi-layer neural network to output the prediction result.

8. The method for controlling a group of pump stations based on global water volume prediction of a drainage system according to claim 4 or 7, characterized in that: The fully connected neural network: Perform Hadamard product calculation on the daily and weekly forecast results to obtain the final forecast results. W d , W w —Learning parameters, trained through the learning process; ⊙—Hadamard product operation; —Daily level forecast results and weekly level forecast results; —Final prediction results.

9. The method for controlling a group of pump stations based on global water volume prediction of a drainage system according to claim 1, characterized in that: Through the drainage system's global water volume prediction model, a pump station group control strategy is established, which specifically includes: packaging the drainage system's global water volume prediction model and embedding it into the smart water system to form a pump station group control strategy; through the collection of on-site data, the water volume of all nodes in the system in the future is predicted, and at the same time, according to the prediction results, the target lift of the pump group of each pump station is preset.

Images