Air shield dam flow coefficient calculation system

Abstract

The invention discloses an air shield dam flow coefficient calculation system, and relates to the technical field of flow coefficient calculation. The system comprises a multi-source data acquisition module used for acquiring multi-source training data of the air shield dam and preprocessing the multi-source training data of the air shield dam; the feature engineering processing module is used for performing feature processing on the preprocessed multi-source training data of the air shield dam to obtain an air shield dam training feature data set; the model training module is used for inputting the air shield dam training feature data set into the constructed hybrid model and training the hybrid model; and the air shield dam flow coefficient prediction module is used for acquiring an actual characteristic data set of the air shield dam and inputting the actual characteristic data set into the trained hybrid model to obtain a predicted air shield dam flow coefficient. Therefore, the influence of the change of the air shield dam structure, the dam crest shape and the incoming flow condition on the flow coefficient is inconvenient to capture.

Description

Technical Field

[0001] The present invention relates to the technical field of flow coefficient calculation, and specifically provides a system for calculating the flow coefficient of an air shield dam. Background Art

[0002] An air shield dam is a new type of hydraulic structure that combines the advantages of a rubber dam and a steel plate dam. This dam absorbs the essence of traditional movable dam types and abandons their deficiencies. It has the characteristics of simple structure, short construction and installation periods, outstanding flood control and flood passing capabilities, safe and reliable operation, controllable water passing capacity and operation status, stronger sewage cleaning and silt discharging capabilities, short charging and discharging times, simple operation management, extremely long service life, high comprehensive benefits, very beautiful overfall at the dam crest, and good landscape effects. This new dam type has the characteristics of traditional water conservancy flow discharge and water retention, as well as the ecological water conservancy landscape and environmental protection advantages. In recent years, it has gradually been widely applied to various rivers with complex hydrological conditions and rivers for beautifying urban construction.

[0003] Chinese Patent No. CN112182701B discloses a method for calculating the flow coefficient and flow rate of an air shield dam. According to Bernoulli's equation, the relationship between the discharge per unit width and the water head of the air shield dam is obtained. Secondly, the water head and discharge per unit width values of the air shield dam without side contraction under 12 working conditions are measured. Then, using excel data analysis, the relationship between the discharge per unit width and the water head of each opening of the air shield dam is determined. And by comparing this relationship with the discharge per unit width formula of the free outfall weir flow without side contraction, the flow coefficient corresponding to each opening of the air shield dam is obtained. Finally, the discharge of the air shield dam is calculated based on the flow coefficient, and the flow rate of the air shield dam without side contraction is determined according to the relationship between the discharge per unit width and the water head.

[0004] The existing calculation methods for flow coefficients are mainly based on the relationship between the discharge per unit width and the water head of the air shield dam. The data is relatively single and difficult to process from multiple dimensions, making it inconvenient to capture the influence of changes in the air shield dam structure, dam crest shape, and incoming flow conditions on the flow coefficient. Summary of the Invention

[0005] Aiming at the deficiencies of the prior art, the present invention provides a system for calculating the flow coefficient of an air shield dam, which solves the problem that the existing calculation methods for flow coefficients have relatively single data and are difficult to process from multiple dimensions, making it inconvenient to capture the influence of changes in the air shield dam structure, dam crest shape, and incoming flow conditions on the flow coefficient.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: A gas shield dam flow coefficient calculation system includes a multi-source data acquisition module, which is used to acquire multi-source training data of the gas shield dam and preprocess the multi-source training data of the gas shield dam. Among them, the multi-source training data of the gas shield dam includes field measurement data, physical model test data, and numerical simulation data; a feature engineering processing module, which is used to perform feature processing on the preprocessed multi-source training data of the gas shield dam to obtain a gas shield dam training feature dataset; a model training module, which is used to input the gas shield dam training feature dataset into the constructed hybrid model and train the hybrid model; a gas shield dam flow coefficient prediction module, which is used to acquire the actual feature dataset of the gas shield dam and input it into the trained hybrid model to obtain the predicted gas shield dam flow coefficient.

[0007] Further, the field measurement data includes measured gas shield dam structure form parameters, measured gas shield dam crest shape parameters, and measured incoming flow condition data, where the measured incoming flow condition data includes the measured gas shield dam flow rate; the physical model test data includes physical model gas shield dam structure form parameters, physical model gas shield dam crest shape parameters, and physical model incoming flow condition data, where the physical model incoming flow condition data includes the physical model gas shield dam flow rate; the numerical simulation data includes numerical simulation gas shield dam structure form parameters, numerical simulation gas shield dam crest shape parameters, and numerical simulation incoming flow condition data, where the numerical simulation incoming flow condition data includes the numerical simulation gas shield dam flow rate.

[0008] Further, the gas shield dam training feature dataset includes multiple groups of training feature data and the corresponding training gas shield dam flow coefficients for each group of training feature data. Among them, the training feature data includes gas shield dam structure form parameters, crest shape numerical features, and incoming flow condition features. The acquisition process is as follows: Sampling the field measurement data, physical model test data, and numerical simulation data according to a set ratio to obtain an initial gas shield dam training dataset, which includes multiple groups of initial training feature data and the corresponding training gas shield dam flow coefficients for each group of initial training feature data. The initial training feature data includes initial gas shield dam structure form parameters, initial crest shape numerical features, and initial incoming flow condition features; Analyzing the influence degree of each parameter feature in the initial training feature data on the flow coefficient based on the correlation coefficient. If the influence degree of a certain parameter feature on the flow coefficient is greater than the set degree threshold, then retain the parameter feature, otherwise delete it to obtain an optimized feature dataset; Normalizing all continuous features of the optimized feature dataset to obtain the gas shield dam training feature dataset.

[0009] Further, the method for obtaining the influence degree of each parameter feature on the flow coefficient is as follows: ; Among them, is the influence degree, is the i-th value of the parameter feature x, is the average value of the parameter feature x, is the i-th value of the flow coefficient, is the average value of the flow coefficient.

[0010] Further, the hybrid model includes an RF-LSTM layer, a GNN layer, and a fusion layer, where: The RF-LSTM layer trains and learns the input training feature dataset of the air shield dam based on random forest and long short-term memory network. After training, it processes the actual feature dataset of the air shield dam and outputs the first predicted flow coefficient of the air shield dam; The GNN layer trains and learns the input training feature dataset of the air shield dam based on the graph neural network. After training, it processes the actual feature dataset of the air shield dam and outputs the second predicted flow coefficient of the air shield dam; The fusion layer performs weighted summation on the first predicted flow coefficient of the air shield dam and the second predicted flow coefficient of the air shield dam to obtain the predicted flow coefficient of the air shield dam.

[0011] Further, the RF-LSTM layer includes a random forest unit, a long short-term memory network unit, and a fusion prediction unit: Among them, the random forest unit processes the air shield dam structure form parameters and the numerical characteristics of the dam crest shape to obtain a preliminary evaluation result of the influence of the air shield dam structure form parameters and the numerical characteristics of the dam crest shape on the flow coefficient. The long short-term memory network unit processes the oncoming flow condition characteristics and outputs the hidden state vector output at the last time step.

[0012] Further, the construction process of the random forest unit is as follows: Concatenate the air shield dam structure form parameters and the numerical characteristics of the dam crest shape in each group to obtain multiple groups of air shield dam structure characteristic parameter data, and fuse them to obtain an air shield dam structure characteristic training dataset containing all groups of air shield dam structure characteristic parameter data; From the air shield dam structure characteristic training dataset, M sub-datasets with the same scale as the multi-source training data of the air shield dam are repeatedly sampled with replacement for constructing M decision trees; During the construction of each decision tree, for each internal node, randomly select some features in the sub-dataset, and determine the optimal splitting feature and splitting point by calculating the information gain, and repeat this process continuously until the preset stop condition is reached; Construct a decision tree output result synthesis layer, and synthesize the output results of the M decision trees to obtain the final output result of the decision tree; The construction process of the long short-term memory network unit is as follows: Set multiple LSTM layers, and each LSTM layer contains N memory units; Connect two fully connected layers after the last LSTM layer. The first fully connected layer contains G neurons, and the second fully connected layer contains 1 neuron.

[0013] Further, the fusion prediction unit includes two fully-connected neural network layers. The first fully-connected neural network layer takes the concatenated feature vector of the final output result of the decision tree and the hidden state vector output by the last LSTM layer as input and performs a linear transformation. The second fully-connected neural network layer performs a second linear transformation on the linear transformation result of the first fully-connected neural network layer, and the output result is used as the first predicted flow coefficient of the pneumatic shield dam.

[0014] Further, the construction process of the GNN layer is as follows: Determine the nodes, including the dam body node, the pneumatic shield node, the dam crest node, and the water flow node; Define the edges: between the dam body node and the adjacent dam body nodes, and between the dam body node and the pneumatic shield node, if there is a physical connection, create an undirected edge with the weight initialized to 1; between the dam crest node and the adjacent dam crest nodes, and between the water flow node and the adjacent water flow nodes, calculate the edge weight according to the reciprocal of the Euclidean distance d, where d > 0; Connect the water flow nodes at the same spatial position in chronological order to create a directed edge with the weight fixed to 1; Initialize the node features of each node; Build an S-layer GCN network to update the node features of each layer; Perform node feature aggregation; Define a global virtual node, and the global virtual node outputs a feature vector after being operated by the S-layer GCN network; The feature vector is input into the GNN fully-connected layer for linear transformation to predict the flow coefficient. The loss function uses the mean square error, and the weights of the GCN layer and the parameters of the fully-connected layer are updated through the Adam optimizer.

[0015] Further, the update formula is as follows: ; where, is the node feature matrix of the l-th layer GCN network, is the node feature matrix of the (l + 1)-th layer GCN network, is the degree matrix, represents the reciprocal of the square root of the degree matrix for symmetric normalization, is the adjacency matrix with self-loops, , is the adjacency matrix, is the identity matrix, is the weight matrix, is the activation function, using the ReLU non-linear function; The formula for predicting the flow coefficient is as follows: ; where, is the predicted flow coefficient output by the GNN layer. When the GNN layer inputs the actual feature dataset of the pneumatic shield dam, it is denoted as the second predicted flow coefficient of the pneumatic shield dam, is the weight matrix of the GNN fully-connected layer, is the feature vector output by the last layer GCN network, is the offset vector.

[0016] The present invention has the following beneficial effects: The gas shield dam discharge coefficient calculation system integrates field measurement, physical model test and numerical simulation data through the multi-source data acquisition module, providing a comprehensive data basis for subsequent analysis; the feature engineering processing module screens key features and normalizes continuous features to improve data quality; the hybrid model combines the advantages of different models, learns data features from multiple dimensions, synthesizes prediction results, and improves the accuracy of gas shield dam discharge coefficient prediction.

[0017] Of course, it is not necessary for any product implementing the present invention to achieve all the above-mentioned advantages simultaneously. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 is a flow chart of the gas shield dam discharge coefficient calculation system of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0019] Please refer to Figure 1 , the embodiment of the present invention provides a technical solution: a gas shield dam discharge coefficient calculation system, including a multi-source data acquisition module, which is used to acquire multi-source training data of the gas shield dam and preprocess the multi-source training data of the gas shield dam. Among them, the multi-source training data of the gas shield dam includes field measurement data, physical model test data and numerical simulation data, which contains various key factors affecting the gas shield dam discharge coefficient and lays a foundation for subsequent accurate analysis and prediction of the discharge coefficient.

[0020] The field measurement data includes measured gas shield dam structure form parameters, measured gas shield dam crest shape parameters and measured incoming flow condition data, where the measured incoming flow condition data includes the measured gas shield dam discharge; at the actual gas shield dam project site, professional measurement equipment and instruments are used to measure the structure form of the gas shield dam (such as dam body dimensions, gas shield layout methods, etc.), the crest shape (accurately measure the curve shape and related dimensions of the crest), and the incoming flow conditions (including discharge, velocity, water level, water temperature, etc.). The corresponding discharge coefficient can be initially estimated using the formula based on broad-crested weir flow, and the formula for broad-crested weir flow is , where Q is the water flow rate, m is the discharge coefficient, B is the crest width of the pneumatic shield dam, g is the acceleration due to gravity, and h is the head above the weir, that is, the difference between the water level in front of the dam and the crest elevation. The discharge coefficient can be preliminarily estimated based on the measured flow rate of the pneumatic shield dam. Similarly, such a preliminary estimation can also be performed on the discharge coefficients corresponding to the physical model test data and numerical simulation data in the following text. This provides a basis for analyzing the relationship between the structural form parameters of the pneumatic shield dam, the shape parameters of the dam crest, and the data of the incoming flow conditions and the discharge coefficient, so that the discharge coefficient can be predicted according to the actual characteristic dataset of the pneumatic shield dam, where the actual characteristic dataset of the pneumatic shield dam includes the actual characteristic parameters of the structural form of the pneumatic shield dam, the actual characteristic parameters of the shape of the dam crest of the pneumatic shield dam, and the actual characteristic data of the incoming flow conditions.

[0021] The physical model test data includes the structural form parameters of the physical model pneumatic shield dam, the shape parameters of the dam crest of the physical model pneumatic shield dam, and the data of the incoming flow conditions of the physical model, where the data of the incoming flow conditions of the physical model includes the flow rate of the physical model pneumatic shield dam; according to certain similarity criteria, a physical model of the pneumatic shield dam is established in the laboratory. By adjusting various parameters in the model, different structural forms, dam crest shapes, and incoming flow conditions are simulated, and then the corresponding discharge coefficients are measured.

[0022] The numerical simulation data includes the structural form parameters of the numerical simulation pneumatic shield dam, the shape parameters of the dam crest of the numerical simulation pneumatic shield dam, and the data of the incoming flow conditions of the numerical simulation, where the data of the incoming flow conditions of the numerical simulation includes the flow rate of the numerical simulation pneumatic shield dam. Using numerical simulation software such as computational fluid dynamics (CFD), a numerical model of the pneumatic shield dam is established. By setting different structural parameters, incoming flow boundary conditions, etc., the water flow movement of the pneumatic shield dam is simulated, and the discharge coefficient is calculated. For example, using software such as ANSYS Fluent, a three-dimensional numerical simulation of the water flow inside the pneumatic shield dam is carried out, and by adjusting parameters such as the shape of the pneumatic shield and the roughness of the dam body, the discharge coefficients under different working conditions are obtained.

[0023] The feature engineering processing module is used to perform feature processing on the preprocessed multi-source training data of the pneumatic shield dam to obtain the training feature dataset of the pneumatic shield dam; The training feature dataset of the pneumatic shield dam includes multiple groups of training feature data and the corresponding training discharge coefficients of each group of training feature data, where the training feature data includes the structural form parameters of the pneumatic shield dam, the numerical characteristics of the dam crest shape, and the characteristics of the incoming flow conditions. The acquisition process is as follows: Sample the on-site measurement data, physical model test data, and numerical simulation data according to a set ratio to obtain the initial training dataset of the air shield dam, including multiple groups of initial training feature data and the corresponding training flow coefficient of the air shield dam for each group of initial training feature data. The initial training feature data includes the initial air shield dam structure form parameters, the numerical characteristics of the initial dam crest shape, and the characteristics of the initial incoming flow conditions. Sample the on-site measurement, physical model test, and numerical simulation data according to a set ratio to obtain the initial training dataset of the air shield dam. This can not only make full use of the diversity of multi-source data, covering the data characteristics under different scenarios, but also control the data scale to a certain extent, avoid waste of computing resources and low training efficiency caused by excessive data volume, and ensure the high efficiency and feasibility of subsequent model training.

[0024] Take the structure form parameters of the air shield dam, namely the dam body length, dam height, air shield diameter, air shield quantity, air shield spacing, etc., as the input features of the model; for discrete features such as air shield quantity and air shield arrangement, use the one-hot encoding method for encoding so that it can be processed by machine learning models; extract the parameters related to the dam crest shape such as the dam crest slope, dam crest curvature, and dam crest flatness as features; convert the description of the dam crest shape into numerical features; select the incoming flow rate, flow velocity, water level, water temperature, etc. as the features of the incoming flow conditions.

[0025] Based on the correlation coefficient, analyze the influence degree of each parameter feature in the initial training feature data on the flow coefficient. If the influence degree of a certain parameter feature on the flow coefficient is greater than the set degree threshold, then retain this parameter feature; otherwise, delete it to obtain the optimized feature dataset. Retain the parameter features with an influence degree greater than the set threshold and delete the irrelevant or less influential features to obtain the optimized feature dataset. This operation removes redundant information and reduces noise interference, enabling the model training to focus on key factors and improving the accuracy and reliability of the model's prediction of the flow coefficient.

[0026] Normalize all the continuous features in the optimized feature dataset to obtain the training feature dataset of the air shield dam. The continuous features include the incoming flow rate, flow velocity, water level, water temperature, etc.

[0027] The method for obtaining the influence degree of each parameter feature on the flow coefficient is as follows: ; Among them, is the influence degree, is the i-th value of the parameter feature x, is the average value of the parameter feature x, is the i-th value of the flow coefficient, is the average value of the flow coefficient.

[0028] For example, the parameter feature x is the flow velocity, with the unit of meters per second. During calculation, dimensionality reduction processing is performed. There are 3 values in total, and the respective values and corresponding discharge coefficients are shown in Table 1 below: Table 1: Values of flow velocity and discharge coefficient

[0029] Based on the values in the table, the mean value of the flow velocity is 1.43, and the average value of the discharge coefficient is 0.39. Through the above formula calculation, the calculation result of the influence degree r is 0.984.

[0030] The model training module is used to input the gas shield dam training feature dataset into the constructed hybrid model and train the hybrid model; The hybrid model includes an RF-LSTM layer, a GNN layer, and a fusion layer, where: The RF-LSTM layer trains and learns the input gas shield dam training feature dataset based on the random forest and long short-term memory network. After training is completed, it processes the actual feature dataset of the gas shield dam and outputs the first predicted gas shield dam discharge coefficient; The RF-LSTM layer includes a random forest unit, a long short-term memory network unit, and a fusion prediction unit: Among them, the random forest unit processes the gas shield dam structure form parameters and the numerical characteristics of the dam crest shape, and obtains a preliminary evaluation result of the influence of the gas shield dam structure form parameters and the numerical characteristics of the dam crest shape on the discharge coefficient. The random forest is mainly used to process static characteristics such as the gas shield dam structure form. For the input gas shield dam structure form feature data, such as dam body length, dam height, gas shield diameter, gas shield quantity, gas shield spacing, and fractal dimension, etc., the random forest first performs multiple random samplings on the dataset to construct multiple decision trees. During the node splitting process of each decision tree, some features are randomly selected, and by calculating indicators such as information gain or Gini coefficient, the optimal splitting feature and splitting point are determined, so as to gradually divide the dataset and form a decision tree structure. Many decision trees are trained in parallel, and finally, the prediction results of each decision tree are integrated by voting or averaging, etc., to obtain a preliminary evaluation of the influence of the gas shield dam structure form on the discharge coefficient.

[0031] The long short-term memory network unit processes the incoming flow condition features and outputs the hidden state vector output at the last time step. The long short-term memory network (LSTM) is used to capture the time series information of dynamic features such as the incoming flow condition. The incoming flow condition data, such as flow rate, flow velocity, water level, water temperature, water flow turbulence characteristic parameters (turbulence intensity, Reynolds stress, etc.), water quality parameters (pH value, dissolved oxygen, etc.), and sediment content, etc., are input into the LSTM network in the form of a time series.

[0032] In the LSTM network, the memory unit selectively memorizes and updates information in the time series under the control of the input gate, forget gate, and output gate. At each time step, the input data is fused with the information in the memory unit, and after a series of linear transformations and activation function operations, the hidden state at the current time step is output. These hidden states not only contain the information at the current time step but also incorporate the long-term dependence information from previous time steps. By learning the hidden states over multiple time steps, the LSTM network can capture the patterns of the oncoming flow conditions changing over time and the impact of these changes on the discharge coefficient.

[0033] The construction process of the random forest unit is as follows: The structural form parameters of each group of air shield dams and the numerical characteristics of the dam crest shape are spliced to obtain multiple groups of air shield dam structural characteristic parameter data, and a training dataset of air shield dam structural characteristics containing all groups of air shield dam structural characteristic parameter data is fused. From the training dataset of air shield dam structural characteristics, M sub-datasets of the same scale as the multi-source training data of the air shield dam are repeatedly sampled with replacement multiple times to construct M decision trees. During the construction of each decision tree, for each internal node, a part of the features in the sub-dataset is randomly selected, and the optimal splitting feature and splitting point are determined by calculating the information gain. This process is continuously repeated until a preset stopping condition is reached. A comprehensive layer for the output results of the decision trees is constructed to synthesize the output results of the M decision trees to obtain the final output result of the decision trees. When all M decision trees are constructed, for the newly input air shield dam structural form feature data, each decision tree makes a prediction independently. If the prediction task is regression (predicting the discharge coefficient as a continuous value), the prediction results of all decision trees are averaged to obtain a preliminary evaluation value of the impact of the air shield dam structural form on the discharge coefficient by the random forest. If it is a classification task (such as classifying the range of the discharge coefficient), the category with the most votes is selected as the prediction result of the random forest through a voting method.

[0034] The construction process of the long short-term memory network unit is as follows: Multiple LSTM layers are set, and each LSTM layer contains N memory units. In this embodiment, 3 LSTM layers are set, and each LSTM layer contains 128 memory units. Two fully connected layers are connected after the last LSTM layer. The first fully connected layer contains G neurons, and the second fully connected layer contains 1 neuron. In this embodiment, the first fully connected layer contains 64 neurons, and the 1 neuron in the second fully connected layer corresponds to outputting a single discharge coefficient prediction value.

[0035] The fusion prediction unit includes two fully connected neural network layers. The first fully connected neural network layer takes the concatenated feature vector of the final output result of the decision tree and the hidden state vector output by the last LSTM layer as input for linear transformation. The second fully connected neural network layer performs a second linear transformation on the linear transformation result of the first fully connected neural network layer, and the output result is used as the first predicted flow coefficient of the pneumatic shield dam.

[0036] Finally, the evaluation result of the pneumatic shield dam structure form output by the random forest is fused with the inflow condition time series feature result output by the LSTM network. The concatenation method can be used to connect the output vectors of the two in sequence to form a new feature vector. Then, this fused feature vector is input into a fully connected neural network. Through the linear transformation of multiple layers of neurons and the action of a non-linear activation function (such as the ReLU function), the fused features are further learned and mapped, and finally the predicted value of the flow coefficient is output. In this process, the fully connected neural network continuously adjusts its own weight and bias parameters to minimize the error between the predicted value and the true flow coefficient, so as to achieve accurate learning of the flow coefficient.

[0037] The GNN layer trains and learns the input pneumatic shield dam training feature dataset based on the graph neural network. After training, it processes the actual feature dataset of the pneumatic shield dam and outputs the second predicted flow coefficient of the pneumatic shield dam; factors such as the structure form of the pneumatic shield dam, the shape of the dam crest, and the inflow conditions are constructed into a graph structure. In the graph structure, each structural component of the pneumatic shield dam (such as the dam body, pneumatic shield, etc.), different positions of the dam crest, and the inflow condition data at different time and space points can be used as nodes of the graph, and the physical connections, spatial position relationships, or time sequence relationships between them can be used as edges of the graph. For node features, the structural parameters of the pneumatic shield dam (such as the dam body length, pneumatic shield diameter, etc.), the dam crest shape parameters (such as the dam crest slope, curvature, etc.), and the inflow condition parameters (such as flow rate, flow velocity, etc.) are used as the initial features of the nodes. Through graph convolution operations, each node continuously aggregates the feature information of its adjacent nodes. After the operations of multiple layers of graph convolution networks, the node features gradually fuse the relevant information in its local area and the global graph structure. Finally, by analyzing the features of specific nodes in the graph (such as the virtual node representing the overall characteristics of the entire pneumatic shield dam), the predicted value of the flow coefficient is output. During the training process, the graph neural network adjusts the parameters of the graph convolution kernel and the update method of node features to minimize the error between the predicted flow coefficient and the true value, so as to learn the complex relationship between the various factors of the pneumatic shield dam and the flow coefficient.

[0038] The construction process of the GNN layer is as follows: Determine nodes, including dam body nodes (units subdivided by dam body volume or geometry, each node carrying a unique number, three-dimensional spatial coordinates, and a structural position identifier), air shield nodes (independent units divided by air shield physical forms (such as diameter, installation angle) and spatial positions, recording air shield types (standard type / enhanced type), inflation state parameters, and installation coordinates on the dam body), dam crest nodes (discrete points divided by dam crest geometric features (slope change segments, key curvature positions), marked with the elevation relative to the dam body base surface, horizontal coordinates, and the dam crest partition they belong to (such as overflow section, non-overflow section)), and water flow nodes (monitoring points set at fixed spatial intervals (such as every 5 meters) on the water flow path, generating new nodes at each time step, carrying a timestamp, spatial coordinates, and an incoming flow direction vector); Define edges: Between dam body nodes and adjacent dam body nodes, and between dam body nodes and air shield nodes, if there is a physical connection (such as bolt fixation, welding), create an undirected edge with the weight initially set to 1; For dam crest nodes and adjacent dam crest nodes, and water flow nodes and adjacent water flow nodes, calculate the edge weight as the reciprocal of the Euclidean distance d, where d > 0; Water flow nodes at the same spatial position are connected in chronological order to create a directed edge with the weight fixed at 1; Initialize the node features of each node, including initializing the dam body node features (standardize geometric parameters such as length, width, and height, and use material parameters directly as feature components, such as elastic modulus, Poisson's ratio), initializing the air shield node features (standardize physical parameters such as air shield diameter, inflation pressure, etc., and keep spatial parameters in their original dimensions, such as installation angle), initializing the dam crest node features (standardize morphological parameters such as slope, curvature), and initializing the water flow node features (standardize dynamic parameters such as flow rate, flow velocity, etc.) Build an S-layer GCN network to update the node features of each layer; Perform node feature aggregation; Define a global virtual node. The global virtual node outputs a feature vector after being operated on by the S-layer GCN network. The global virtual node is connected to the following key nodes through trainable edges: dam body nodes with the top 10% stress concentration coefficients in the dam body (marked as key stress-bearing nodes), dam crest nodes with a dam crest curvature change rate exceeding a threshold (such as 0.1 / m) (marked as shape key nodes), and water flow nodes with the top 15% water flow velocity gradients (marked as flow-sensitive nodes). The edge weight is initially set to 1 and is adaptively adjusted through an attention mechanism during training.

[0039] Input the feature vector into the fully connected layer of the GNN for linear transformation, perform flow coefficient prediction, use the mean squared error as the loss function, and update the weights of the GCN layer and the parameters of the fully connected layer through the Adam optimizer.

[0040] The node features of each layer are updated, and the update formula is as follows: ; where, where, is the node feature matrix of the l-th layer GCN network, is the node feature matrix of the (l + 1)-th layer GCN network, is the degree matrix, represents the reciprocal of the square root of the degree matrix, which is used for symmetric normalization, is the adjacency matrix with self-loops, , is the adjacency matrix, is the identity matrix, is the weight matrix, is the activation function, using the ReLU non-linear function, in the form of , which means that when X is greater than 0, the ReLU non-linear function outputs X, otherwise 0; In one embodiment, assume there is a simplified gas shield dam graph structure, which contains 3 nodes: 1 dam body node (B1), 1 gas shield node (S1), and 1 water flow node (W1). Initially, their feature vectors are as follows: The feature vector of the dam body node B1 , representing length, width, elastic modulus, and Poisson's ratio respectively; the feature vector of the gas shield node S1 , representing gas shield diameter, wall thickness, inflation pressure, and installation angle respectively; the feature vector of the water flow node W1 , representing flow rate, flow velocity, and water level respectively.

[0041] Adjacency matrix A represents the connection relationship between nodes. Assume that in this graph structure, the dam body node B1 is connected to the gas shield node S1, and the water flow node W1 is connected to the dam body node B1. Its adjacency matrix A is: ; Due to the addition of self-loops, ; Degree matrix The diagonal elements of are the degrees of the corresponding nodes (including self-loops), that is: ; Assume that the weight matrix of the 0-th layer is a randomly initialized matrix, which is set here as: ; The activation function uses the ReLU non-linear function, and we get: ; That is 。

[0042] The formula for predicting the flow coefficient is as follows: ; Wherein, is the predicted flow coefficient output by the GNN layer. When the actual characteristic data set of the pneumatic shield dam is input to the GNN layer, it is denoted as the second predicted flow coefficient of the pneumatic shield dam. is the weight matrix of the GNN fully connected layer. is the feature vector output by the last layer of the GCN network. is the bias vector.

[0043] The fusion layer performs a weighted sum of the first predicted flow coefficient of the pneumatic shield dam and the second predicted flow coefficient of the pneumatic shield dam to obtain the predicted flow coefficient of the pneumatic shield dam.

[0044] The pneumatic shield dam flow coefficient prediction module is used to obtain the actual characteristic data set of the pneumatic shield dam and input it into the trained hybrid model to obtain the predicted flow coefficient of the pneumatic shield dam.

[0045] An electronic device includes: a processor; and a memory in which computer program instructions are stored, and when the computer program instructions are run by the processor, the processor executes the pneumatic shield dam flow coefficient calculation system as described above.

[0046] A computer-readable storage medium is used to store a program, and when the program is executed by a processor, the pneumatic shield dam flow coefficient calculation system as described above is implemented.

[0047] Those skilled in the art should understand that the embodiments of the present invention can be provided as a method, a system, or a computer program product. Therefore, the present invention can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0048] The present invention is described with reference to the flowcharts and / or block diagrams of systems, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of flows and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate for implementing in the process Figure 1one or more processes and / or blocks Figure 1 means for the functions specified in one or more blocks

[0049] These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to operate in a particular manner, such that the instructions stored in the computer-readable memory produce a manufacture including an instruction means that implements the functions in the process Figure 1 one or more processes and / or blocks Figure 1 specified in one or more blocks

[0050] These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus, such that a series of operational steps are performed on the computer or other programmable apparatus to produce a computer-implemented process, whereby the instructions executed on the computer or other programmable apparatus provide steps for implementing the functions specified in the process Figure 1 one or more processes and / or blocks Figure 1 specified in one or more blocks

[0051] Although the preferred embodiments of the present invention have been described, additional changes and modifications can be made by those skilled in the art once they learn of the basic inventive concept. Therefore, the appended claims are intended to be construed to cover the preferred embodiments as well as all changes and modifications falling within the scope of the present invention

[0052] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalent technologies, the present invention is also intended to include these modifications and variations

Claims

1. A gas shield dam flow coefficient calculation system, characterized in that: include: A multi-source data acquisition module is used to acquire and pre-process the multi-source training data of the gas shield dam, wherein the multi-source training data of the gas shield dam includes field measurement data, physical model test data and numerical simulation data; A feature engineering processing module is used to perform feature processing on the preprocessed gas shield dam multi-source training data to obtain a gas shield dam training feature data set; A model training module is used to input the gas shield dam training feature data set into the constructed hybrid model to train the hybrid model; The gas shield dam discharge coefficient prediction module is used to obtain the actual characteristic data set of the gas shield dam and input it into the trained hybrid model to obtain the predicted gas shield dam discharge coefficient.

2. A gas shield dam flow coefficient calculation system according to claim 1, characterized in that: The field measurement data include the measured gas shield dam structural form parameters, the measured gas shield dam crest shape parameters and the measured inflow condition data, wherein the measured inflow condition data include the measured gas shield dam flow rate; The physical model test data include physical model gas shield dam structural form parameters, physical model gas shield dam crest shape parameters and physical model inflow condition data, wherein the physical model inflow condition data include physical model gas shield dam flow rate; The numerical simulation data include numerical simulation gas shield dam structural form parameters, numerical simulation gas shield dam top shape parameters and numerical simulation flow condition data, wherein the numerical simulation flow condition data include numerical simulation gas shield dam flow rate.

3. A gas shield dam flow coefficient calculation system according to claim 2, characterized in that: The gas shield dam training feature data set includes multiple sets of training feature data and the training gas shield dam discharge coefficients corresponding to each set of training feature data. The training feature data includes the gas shield dam structural form parameters, dam top shape numerical features and inflow condition features. The acquisition process is as follows: The field measurement data, physical model test data and numerical simulation data are sampled according to a set ratio to obtain an initial gas shield dam training data set, including multiple groups of initial training feature data and training gas shield dam flow coefficients corresponding to each group of initial training feature data. The initial training feature data includes initial gas shield dam structural form parameters, initial dam top shape numerical characteristics and initial inflow condition characteristics; Based on the correlation coefficient, the influence of each parameter feature in the initial training feature data on the flow coefficient is analyzed. If the influence of a certain parameter feature on the flow coefficient is greater than the set degree threshold, the parameter feature is retained, otherwise it is deleted to obtain an optimized feature data set; All continuous features of the optimized feature data set are normalized to obtain the gas shield dam training feature data set.

4. A gas shield dam flow coefficient calculation system according to claim 3, characterized in that: The method for obtaining the influence of each parameter characteristic on the flow coefficient is as follows: ； in, For the degree of impact, is the i-th value of parameter feature x, is the average value of parameter feature x, is the ith value of the flow coefficient, is the average value of the flow coefficient.

5. The gas shield dam flow coefficient calculation system according to claim 3, characterized in that: The hybrid model includes RF-LSTM layer, GNN layer and fusion layer, where: The RF-LSTM layer trains and learns the input gas shield dam training feature data set based on random forest and long short-term memory network. After the training is completed, the actual feature data set of the gas shield dam is processed and the first predicted gas shield dam discharge coefficient is output; The GNN layer trains and learns the input gas shield dam training feature data set based on the graph neural network. After the training is completed, the actual feature data set of the gas shield dam is processed and the second predicted gas shield dam discharge coefficient is output; The fusion layer performs weighted summation of the first predicted gas shield dam flow coefficient and the second predicted gas shield dam flow coefficient to obtain the predicted gas shield dam flow coefficient.

6. A gas shield dam flow coefficient calculation system according to claim 5, characterized in that: The RF-LSTM layer includes a random forest unit, a long short-term memory network unit, and a fusion prediction unit: Among them, the random forest unit processes the structural parameters of the air shield dam and the numerical characteristics of the dam crest shape to obtain the preliminary evaluation results of the influence of the structural parameters of the air shield dam and the numerical characteristics of the dam crest shape on the discharge coefficient. The long short-term memory network unit processes the characteristics of the incoming flow conditions and outputs the hidden state vector output at the last time step.

7. A gas shield dam flow coefficient calculation system according to claim 6, characterized in that: The construction process of the random forest unit is as follows: Each group of gas shield dam structural form parameters and dam top shape numerical features are spliced ​​to obtain multiple groups of gas shield dam structural characteristic parameter data, and then fused to obtain a gas shield dam structural characteristic training data set containing all groups of gas shield dam structural characteristic parameter data; From the gas shield dam structural feature training data set, M sub-data sets with the same scale as the gas shield dam multi-source training data are extracted multiple times in a sampling method with replacement, and used to construct M decision trees; In the process of building each decision tree, for each internal node, some features in the sub-dataset are randomly selected, and the optimal split features and split points are determined by calculating the information gain. This process is repeated until the preset stop condition is reached; Construct a decision tree output result synthesis layer to synthesize the output results of M decision trees to obtain the final output result of the decision tree; The construction process of the long short-term memory network unit is as follows: Set up multiple LSTM layers, each containing N memory units; Two fully connected layers are connected after the last LSTM layer. The first fully connected layer contains G neurons and the second fully connected layer contains 1 neuron.

8. A gas shield dam flow coefficient calculation system according to claim 6, characterized in that: The fusion prediction unit includes two neural network fully connected layers, in which the first neural network fully connected layer takes the concatenated feature vector of the final output result of the decision tree and the hidden state vector output by the last LSTM layer as input for linear transformation, and the second neural network fully connected layer performs a second linear transformation on the linear transformation result of the first neural network fully connected layer and outputs the result as the first predicted gas shield dam discharge coefficient.

9. The gas shield dam flow coefficient calculation system according to claim 5, characterized in that: The construction process of the GNN layer is as follows: Determine the nodes, including dam body nodes, air shield nodes, dam top nodes and water flow nodes; Define the edges: If there is a physical connection between the dam body node and the adjacent dam body node, or between the dam body node and the air shield node, an undirected edge is created, and the weight is initialized to 1; The edge weights between the dam crest node and the adjacent dam crest node, and between the water flow node and the adjacent water flow node are calculated according to the inverse of the Euclidean distance d, where d is greater than 0; The water flow nodes at the same spatial position are connected in time order to create directed edges with a fixed weight of 1; Initialize the node features of each node; Build an S-layer GCN network and update the features of nodes in each layer; Perform node feature aggregation; Define a global virtual node, which outputs a feature vector after being operated by the S-layer GCN network; The feature vector is input into the GNN fully connected layer for linear transformation to predict the flow coefficient. The loss function uses the mean square error, and the GCN layer weights and fully connected layer parameters are updated through the Adam optimizer.

10. A gas shield dam discharge coefficient calculation system according to claim 9, characterized in that: The node features of each layer are updated, and the update formula is as follows: ； in, is the node feature matrix of the l-th layer GCN network, is the node feature matrix of the l+1th layer GCN network, is the degree matrix, represents the inverse square root of the degree matrix, used for symmetric normalization, is an adjacency matrix with self-loops, , is the adjacency matrix, is the identity matrix, is the weight matrix, As the activation function, the ReLU nonlinear function is adopted; The formula for predicting the flow coefficient is as follows: ； in, is the predicted discharge coefficient output by the GNN layer. When the GNN layer inputs the actual characteristic data set of the gas shield dam, it is recorded as the second predicted gas shield dam discharge coefficient. is the weight matrix of the GNN fully connected layer, is the feature vector output by the last layer of GCN network, is the bias vector.

Images