A method and device for predicting water flow behavior in deep-water high-pile cap construction

Abstract

The present invention relates to the technical field of water flow behavior prediction, and in particular to a method and device for predicting water flow behavior in deep-water high-pile foundation construction. The steps are as follows: collect relevant hydrological data, meteorological data and topographic data of the construction site, and then pre-process the collected data; construct a dynamic graph according to the pre-processed data, and then initialize the data of the dynamic graph; construct AG-RGCN, input the initialized dynamic graph into AG-RGCN to obtain a graph embedding sequence; input the graph embedding sequence at each moment into the sequence model TCN, and the sequence model TCN predicts the water flow behavior in the construction of deep-water high-pile foundation by learning the time dependency in the graph embedding sequence; the decoder predicts the future graph structure of the graph embedding at each moment obtained by the sequence model TCN, and then predicts the hydrological state at each time point in the future. The present invention can improve the safety and operational efficiency of engineering projects, and solve the limitations of traditional methods in dealing with complex hydrological conditions.

Description

Technical Field

[0001] The present invention relates to the technical field of water flow behavior prediction, and in particular to a method and equipment for predicting water flow behavior in deep water high pile cap construction. Background Art

[0002] During the implementation of the deepwater high pile foundation construction project, the engineering team faced a core technical challenge: how to accurately predict and analyze the dynamic changes of water flow behavior, including multi-dimensional hydrological and meteorological conditions such as flow rate, flow direction, surge, tide, etc. These changes are directly related to construction safety, the rationality of structural design and the stability of the project, and have a decisive influence on ensuring the safety of the construction process and improving project efficiency. In view of this, the development of a set of technical solutions that can accurately predict water flow dynamics is of great significance to ensuring construction safety and improving project efficiency.

[0003] At present, traditional water flow prediction technology mainly relies on static hydrological data analysis, for example, predicting possible changes in future water flow by analyzing historical hydrological data. However, such methods face many limitations: on the one hand, they are difficult to capture the dynamic characteristics of water flow changes in real time, especially under extreme meteorological conditions or sudden hydrological events; on the other hand, these methods often ignore the complex interactions between water flow behaviors and their correlation with surrounding environmental factors (such as wind speed, rainfall, etc.), resulting in limited prediction accuracy; in recent years, although some advanced models have tried to improve prediction accuracy through dynamic system theory and computer simulation, these models have high requirements for computing resources, and the requirements for data quality and real-time performance are also difficult to meet in actual construction scenarios, especially in remote areas or construction sites with poor data collection conditions.

[0004] In addition, the shortcomings of current technology in data fusion cannot be ignored. In complex engineering construction environments, it is necessary to process and analyze information from multiple data sources such as satellite remote sensing, on-site monitoring, and historical meteorological records. However, traditional prediction models often find it difficult to effectively integrate these data from different sources and dimensions, making it impossible to form a comprehensive and unified water flow behavior prediction. Based on the above analysis, there is an urgent need for an innovative water flow behavior prediction technology that should have the ability to integrate multi-source data, reflect the dynamic changes of water flow in real time, and accurately predict the impact of these changes on deep-water high-pile foundation construction.

[0005] Therefore, the present invention provides a method and equipment for predicting water flow behavior in deepwater high pile cap construction. The present invention can effectively overcome the limitations of the existing technology, significantly improve the accuracy and practicality of the prediction, and provide a solid technical guarantee for the safe and efficient implementation of deepwater high pile cap construction projects. Summary of the invention

[0006] In view of the deficiencies in the prior art, the present invention develops a method and device for predicting water flow behavior in deep-water high-pile cap construction, which can improve the safety and operational efficiency of engineering projects and solve the limitations of traditional methods in dealing with complex hydrological conditions.

[0007] The technical solution to the technical problem of the present invention is as follows: On the one hand, the present invention provides a method for predicting water flow behavior in deep-water high-pile cap construction, comprising the following steps:

[0008] S1. Data collection and preprocessing: Collect relevant hydrological data, meteorological data and topographic data of the construction site, and then preprocess the collected data;

[0009] S2. Build dynamic graph: Build dynamic graph based on preprocessed data , and then for the dynamic graph Initialize data;

[0010] S3. Construct AG-RGCN and initialize the dynamic graph Input to AG-RGCN to get graph embedding sequence ;

[0011] S4. Predicting water flow status: embedding the graph at each moment into the sequence Input to the sequence model TCN to obtain the future time embedding graph sequence ;

[0012] S5. Generate future graphs to predict changes in water flow behavior: Use the decoder to embed future time graph sequences The future graph structure is predicted, and then the hydrological status at various time points in the future is predicted.

[0013] S1 is as follows:

[0014] S1.1. Data collection:

[0015] Hydrological data including water levels of relevant water bodies , flow rate and flow direction , mainly obtained from hydrological monitoring sites;

[0016] Weather data including precipitation , wind speed ,wind direction and temperature , obtained from the public data interface of the meteorological department and the meteorological station;

[0017] Topographic data including topographic maps of the construction area and its upstream and downstream areas , elevation data and soil type , obtained through GIS database and MSI remote sensing data;

[0018] S1.2. Data preprocessing: including data cleaning and data standardization:

[0019] S1.2.1. Clean the collected data to remove or correct the noise, missing values ​​and outliers in the data;

[0020] (1) Missing value processing: use the mean, median or interpolation method of the overall data to fill in the missing values;

[0021] (2) Noise processing: noise reduction of data is performed through a moving average filter;

[0022] (3) Outlier detection and processing: Detect outliers in the data using the triple standard deviation method, and choose to correct or delete them based on the actual situation. Then, smooth the outliers and use the average value of the neighboring points to replace the corrected or deleted outliers.

[0023] S1.2.2. Data standardization: After data cleaning, the data set is further standardized. The data of different dimensions are converted to the same scale through the z-score standardization method to finally obtain the preprocessed data.

[0024] Step S2 creates a dynamic graph based on the preprocessed data , as follows:

[0025] S2.1. Set dynamic image Updated time window: For short-term streamflow forecasts, select a time window with a time unit of days or hours, and for long-term streamflow forecasts, select a time window with a time unit of months or years;

[0026] S2.2. Divide the continuous hydrological monitoring data into a series of time windows by the timestamp of the real-time monitoring data, that is, divide the pre-processed data into time windows to obtain a set of static graph snapshots of the time series, that is, the dynamic graph , , Indicates the number of timestamps for time window division. Represents a dynamic graph Middle A static image, is a directed graph, each static graph Include node collection , edge set and relationship set , expressed as , Indicates A collection of nodes of a static graph, Indicates The edge set of a static graph, Indicates A static graph relationship set, ;

[0027] S2.2.1. Determine the content of each node set based on the hydrological measurement points nodes, represented by , Indicates The first static image nodes, each of which stores the pre-processed hydrological data, meteorological data, and topographic data of the corresponding hydrological measurement point, expressed as ,in Respectively expressed as The first static image The water level, flow velocity, flow direction, precipitation, wind speed, wind direction, topography, elevation data and soil type information in each node, ;

[0028] S2.2.2. According to The nodes determine the edge set, and an edge can be formed between any two points, expressed as , Indicates The first static image Strip edge, ;

[0029] S2.2.3. Determine the set of relationships between nodes based on the edges formed by the nodes. Each edge corresponds to a corresponding relationship. The types of relationships include water level change relationships. , velocity change relationship Relationship with flow direction change , expressed as , Indicates The first static image The relationship on the edge;

[0030] S2.3. Order the dynamic graph in chronological order Arrange the static images in the time window and select the static image of the first time window Initialize it as a node set Assign initial values ​​and set the node Establish edge set in middle node and relationship set , and then update the data of the subsequent time window static graph, initialize the data of the subsequent static graph in turn, and finally realize the whole dynamic graph Initialization of internal data.

[0031] Step S3 is as follows:

[0032] S3.1. Construct an attention-based gated relational graph convolutional network, AG-RGCN for short. AG-RGCN includes input layer, embedding layer, convolution layer, aggregation layer and output layer.

[0033] S3.2. Initialize the dynamic graph Input to AG-RGCN:

[0034] S3.2.1. First, the initialized dynamic graph Input to the input layer of AG-RGCN, and the initialized dynamic graph is It is further transmitted to the embedding layer of AG-RGCN, and the dynamic graph is Each node in each static graph contains original features, and then converts the extracted original features into vector representation. The eigenvector of ,in Representation Node The original features extracted are water level original features, flow velocity original features, flow direction original features, precipitation original features, wind speed original features, wind direction original features, topography original features, elevation data original features and soil type original features. , ;

[0035] Then the nodes are transformed linearly The feature vector is mapped to the target dimension of the embedding vector , and obtain the feature vector , the calculation formula is as follows:

[0036] ,

[0037] in, represents the bias matrix, represents the embedding matrix;

[0038] At the same time and Initialize and set Initialize to 0 vector, Initialize, The initialization is expressed as:

[0039] ,

[0040] in, represents the uniform distribution function, Indicates that from the interval Uniform sampling means that each value has the same probability of appearing in the interval. Represents the input feature vector Dimension.

[0041] S3.2.2, the convolutional layer extracts and updates features through a multi-layer gating mechanism and an attention mechanism, and the number of convolutional layers is set to , the feature vector of the node is updated once through a convolutional layer, except that the input of the first convolutional layer is Except for , the remaining layers use the output of the previous layer as the input of the next layer. The calculation process is as follows:

[0042] ,

[0043] , ,

[0044] ,

[0045] ,

[0046] ,

[0047] in, Representation Node In the The embedding vector of the layer, Representation Node In the The embedding vector of the layer, Representation Node No. The embedding vector of the layer, represents the conversion gate, represents a nonlinear transformation operation, represents the weight matrix of the conversion gate, represents the bias vector, Represents the activation function Operation, Represents the activation function Operations; Representation Node In relationship ∈ The neighbor set below, Representation Node relation The neighbor nodes below , represents the number of layers in the neural network, , ; Indicates that the current layer is layer, Indicates that the previous layer is layer, Indicated in Neighbor nodes in the layer The characteristic vector of relation The weight matrix of express The real matrix of represents the dimension of the embedding vector, represents a global weight matrix that is used to transform the node's own features and does not depend on any specific relationship, Indicates Layer Node and neighbor nodes The learnable normalized attention weights between Indicates Layer Node and neighbor nodes The attention coefficient;

[0048] Finally, each static image is obtained Each node The eigenvector of The corresponding updated node embedding vector , , , Represents the last layer of the convolutional layer The output of the layer The updated static image The embedding vector of each node.

[0049] S3.2.3, Aggregate information: Static images middle The embedding vector of each node Perform splicing to generate graph embedding , the calculation process is as follows:

[0050] ,

[0051] ∈ ,

[0052] in, Represents a splicing operation, represents the graph embedding matrix, Indicates the number of nodes, Indicates the dimension of the node;

[0053] Integrate the graph embeddings of each static graph to get the graph embedding sequence , represents the graph embedding sequence matrix, and T represents the number of timestamps for time window division.

[0054] Embed the graph at each moment into a sequence Input to the sequence model TCN to embed the graph into the sequence Move in the past time sequence direction, through the convolution filter in the TCN model, the output of the filter and the graph are embedded in the sequence After the activation function, we finally get the future time embedding graph sequence , the specific calculation is as follows:

[0055] ,

[0056] ,

[0057] in, represents the output of the filter, The filter convolution kernels, represents the number of convolution kernels of the filter, represents the operation of the filter, represents the expansion factor, Indicates the past tense direction. Representing Graph Embeddings Graph embedding into the past temporal direction, represents the activation function, Indicates A future time embedding graph, , Indicates the number of timestamps that are moved according to the filter.

[0058] S5 is as follows:

[0059] Embedding a graph at a future time Select the starting node and the target node , the relationship between the two points is set to , embedding graph sequences from future time In Extract the starting node and the target node The corresponding starting embedding vector and the target embedding vector , according to the starting node and the target node The relationship between Find the diagonal matrix , the starting embedding vector , target embedding vector and the diagonal matrix Input to the decoder, the decoder uses the DistMult factorization method to decompose each potential edge The scoring calculation process is as follows:

[0060] ,

[0061] in, represents transpose, Indicates the starting node and the target node Through the relationship Likelihood score of connection, , , , Indicates Future time embedding graph The middle node is combined, Indicates Future time embedding graph China-US relations integration;

[0062] Will The results and threshold setting For comparison, When the future time point exists, and the water flow behavior that will occur between the two nodes is the relationship .

[0063] On the other hand, the present invention provides a device for predicting water flow behavior in deep-water high-pile cap construction, which is composed of the following:

[0064] (1) Data acquisition unit:

[0065] Sensors: Various hydrological sensors deployed at key monitoring points to collect water level, flow velocity, and flow direction in real time;

[0066] Meteorological instruments: used to collect data on meteorological conditions that affect water flow, such as wind speed, direction, rainfall, and temperature;

[0067] Topographic scanning equipment: using radar or laser scanning technology to collect topographic maps and elevation data of the construction area and its surrounding areas;

[0068] (2) Data processing unit:

[0069] Central Processing Unit (CPU): responsible for processing the large amount of data collected:

[0070] Graphics Processing Unit (GPU): used to accelerate the computation of machine learning and deep learning algorithms, especially the training and application of dynamic graph convolutional neural networks:

[0071] Storage devices: used to store historical and real-time data, as well as model training and prediction results;

[0072] (3) Data communication unit:

[0073] Wireless transmission module: supports Wi-Fi and / or cellular networks to ensure that the device can receive and send data and communicate with the central monitoring system in real time;

[0074] Interface unit: mainly includes USB and Ethernet ports, providing the device with the ability to connect to other systems or devices;

[0075] (4) Display and interaction unit:

[0076] Display screen: used to display water flow prediction results and alarm information in real time:

[0077] Operator Interface: User-friendly interface allows operators to easily configure device settings, view data and reports.

[0078] The effects provided in the summary of the invention are only the effects of the embodiments, rather than all the effects of the invention. The above technical solution has the following advantages or beneficial effects:

[0079] The method of the present invention combines the latest machine learning and deep learning algorithms to achieve real-time and accurate monitoring of flow velocity, flow direction, water level and other key hydrological parameters in complex underwater environments. The technology of the present invention integrates real-time hydrological and meteorological data, topographic information and water flow characteristic data of specific construction areas, which not only greatly improves the safety and engineering efficiency of deep-water construction, but also effectively solves the limitations of existing prediction technologies in dealing with changing hydrological conditions, and provides a highly reliable risk assessment and management solution.

[0080] The core advantages of the present invention are its excellent real-time data processing and analysis capabilities, dynamic adaptive learning mechanism and high-precision prediction of complex hydrological dynamics. These characteristics ensure that the model can provide stable and accurate prediction of water flow behavior in a constantly changing underwater construction environment, thereby ensuring the safety of engineering structures and optimizing construction progress. In addition, the data processing technology of the present invention is not limited to deep-water high-pile foundation construction, but is also applicable to water conservancy projects, flood management, environmental protection and other fields, showing its broad application prospects and high practicality.

[0081] The functions of a water flow behavior prediction device in deep-water high-pile foundation construction in the present invention are: it can monitor data in real time, and the device continuously monitors hydrological and meteorological conditions through its sensors and scanning equipment to provide necessary input data for water flow behavior prediction; it can perform data processing and prediction, and the central processing unit and the graphics processing unit process the collected data, run the dynamic graph model to analyze and predict water flow behavior, and output prediction results including flow velocity, flow direction and water level; when potential dangerous water flow behavior is predicted, the device will automatically trigger an alarm and send alarm information to relevant personnel through the display screen and the communication unit; it can also perform user interaction and configuration, and the operator can adjust the monitoring parameters, set the alarm threshold or update the model through the interactive unit to adapt to different engineering needs and environmental changes. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0083] Figure 1 It is a structural schematic diagram of the method in the present invention.

[0084] Figure 2 It is a structural schematic diagram of the device in the present invention. DETAILED DESCRIPTION

[0085] In order to clearly illustrate the technical features of the present invention, the present invention is described in detail below through specific implementation methods and in conjunction with the accompanying drawings. The following disclosure provides many different embodiments or examples for implementing different structures of the present invention. In order to simplify the disclosure of the present invention, the components and settings of specific examples are described below.

[0086] Example 1

[0087] like Figure 1 As shown in FIG. 1 , a method for predicting water flow behavior in deep-water high pile cap construction is provided, and the steps are as follows:

[0088] S1. Data collection and preprocessing: Collect relevant hydrological data, meteorological data and topographic data of the construction site, and then preprocess the collected data;

[0089] S2. Build dynamic graph: Build dynamic graph based on preprocessed data , and then for the dynamic graph Initialize data;

[0090] S3. Construct AG-RGCN and initialize the dynamic graph Input to AG-RGCN to get graph embedding sequence ;

[0091] S4. Predicting water flow status: embedding the graph at each moment into the sequence The input is sent to the sequence model TCN, which predicts the water flow behavior in deep-water high pile cap construction by learning the time dependency relationship in the graph embedding sequence;

[0092] S5. Generate future graphs to predict changes in water flow behavior: The decoder predicts the future graph structure of the graph embedding at each moment obtained by the sequence model TCN, and then predicts the hydrological state at each time point in the future to provide decision support for engineering management;

[0093] The AG-RGCN in steps S3, S4 and S5, predicting water flow status and generating future graphs to predict water flow behavior changes constitute the water flow behavior prediction model AGD.

[0094] S1 is as follows:

[0095] S1.1. Data collection:

[0096] Hydrological data including water levels of relevant water bodies , flow rate and flow direction , mainly obtained from hydrological monitoring sites;

[0097] Weather data including precipitation , wind speed ,wind direction and temperature , obtained from the public data interface of the meteorological department and the meteorological station;

[0098] Topographic data including topographic maps of the construction area and its upstream and downstream areas , elevation data and soil type , obtained through GIS database and MSI remote sensing data;

[0099] S1.2. Data preprocessing: including data cleaning and data standardization:

[0100] S1.2.1. Clean the collected data to remove or correct the noise, missing values ​​and outliers in the data;

[0101] (1) Missing value processing: use the mean, median or interpolation method of the overall data to fill in the missing values;

[0102] (2) Noise processing: noise reduction of data is performed through a moving average filter;

[0103] (3) Outlier detection and processing: Detect outliers in the data using the triple standard deviation method, and choose to correct or delete them based on the actual situation. Then, smooth the outliers and use the average value of the neighboring points to replace the corrected or deleted outliers.

[0104] S1.2.2. Data standardization: After data cleaning, the data set is further standardized. The data of different dimensions are converted to the same scale through the z-score standardization method to finally obtain the preprocessed data.

[0105] Step S2 creates a dynamic graph based on the preprocessed data , as follows:

[0106] S2.1. Set dynamic image Updated time window: For short-term streamflow forecasts, select a time window with a time unit of days or hours, and for long-term streamflow forecasts, select a time window with a time unit of months or years;

[0107] S2.2. Divide the continuous hydrological monitoring data into a series of time windows by the timestamp of the real-time monitoring data, that is, divide the pre-processed data into time windows to obtain a set of static graph snapshots of the time series, that is, the dynamic graph , , Indicates the number of timestamps for time window division. Represents a dynamic graph Middle A static image, is a directed graph, each static graph Include node collection , edge set and relationship set , expressed as , Indicates A collection of nodes of a static graph, Indicates The edge set of a static graph, Indicates A static graph relationship set, ;

[0108] S2.2.1. Determine the content of each node set based on the hydrological measurement points nodes, represented by , Indicates The first static image nodes, each of which stores the pre-processed hydrological data, meteorological data, and topographic data of the corresponding hydrological measurement point, expressed as ,in Respectively expressed as The first static image The water level, flow velocity, flow direction, precipitation, wind speed, wind direction, topography, elevation data and soil type information in each node, ;

[0109] S2.2.2. According to The nodes determine the edge set, and an edge can be formed between any two points, expressed as , Indicates The first static image Strip edge, ;

[0110] S2.2.3. Determine the set of relationships between nodes based on the edges formed by the nodes. Each edge corresponds to a corresponding relationship. The types of relationships include water level change relationships. , velocity change relationship Relationship with flow direction change , expressed as , Indicates The first static image The relationship on the edge;

[0111] S2.3. Order the dynamic graph in chronological order Arrange the static images in the time window and select the static image of the first time window Initialize it as a node set Assign initial values ​​and set the node Establish edge set in middle node and relationship set , and then update the data of the subsequent time window static graph, initialize the data of the subsequent static graph in turn, and finally realize the whole dynamic graph Initialization of internal data.

[0112] Step S3 is as follows:

[0113] S3.1. Construct an attention-based gated relational graph convolutional network, AG-RGCN for short. AG-RGCN includes an input layer, an embedding layer, a convolutional layer, an aggregation layer, and an output layer. It aims to simulate the dependencies between nodes in a dynamic graph and generate potential feature representations of nodes and their relationships at each time point. According to the definition of the present invention, in a dynamic graph In any static image The information accumulated on is based on the relation type (defined by the flow event type) and the direction of the edge, and only edges with the same event type and direction are considered to belong to the same relation type.

[0114] S3.2. Initialize the dynamic graph Input to AG-RGCN:

[0115] S3.2.1. First, the initialized dynamic graph Input to the input layer of AG-RGCN, and the initialized dynamic graph is It is further transmitted to the embedding layer of AG-RGCN, and the dynamic graph is Each node in each static graph contains original features, and then converts the extracted original features into vector representation. The eigenvector of ,in Representation Node The original features extracted are water level original features, flow velocity original features, flow direction original features, precipitation original features, wind speed original features, wind direction original features, topography original features, elevation data original features and soil type original features. , ;

[0116] Then the nodes are transformed linearly The feature vector is mapped to the target dimension of the embedding vector , and obtain the feature vector , the calculation formula is as follows:

[0117] ,

[0118] in, represents the bias matrix, represents the embedding matrix;

[0119] At the same time and Initialize and set Initialize to 0 vector, Initialize, The initialization is expressed as:

[0120] ,

[0121] in, represents the uniform distribution function, Indicates that from the interval Uniform sampling means that each value has the same probability of appearing in the interval. Represents the input feature vector Dimension.

[0122] S3.2.2, the convolutional layer extracts and updates features through a multi-layer gating mechanism and an attention mechanism, and the number of convolutional layers is set to , the feature vector of the node is updated once through a convolutional layer, except that the input of the first convolutional layer is Except for , the remaining layers use the output of the previous layer as the input of the next layer. The calculation process is as follows:

[0123] ,

[0124] , ,

[0125] ,

[0126] ,

[0127] ,

[0128] in, Representation Node In the The embedding vector of the layer, Representation Node In the The embedding vector of the layer, Representation Node No. The embedding vector of the layer, represents the conversion gate, represents a nonlinear transformation operation, represents the weight matrix of the conversion gate, represents the bias vector, Represents the activation function Operation, Represents the activation function Operations; Representation Node In relationship ∈ The neighbor set below, Representation Node relation The neighbor nodes below, , represents the number of layers in the neural network, , ; Indicates that the current layer is layer, Indicates that the previous layer is layer, Indicated in Neighbor nodes in the layer The characteristic vector of relation The weight matrix of express The real matrix of represents the dimension of the embedding vector, represents a global weight matrix that is used to transform the node's own features and does not depend on any specific relationship, Indicates Layer Node and neighbor nodes The learnable normalized attention weights between Indicates Layer Node and neighbor nodes The attention coefficient;

[0129] Finally, each static image is obtained Each node The eigenvector of The corresponding updated node embedding vector , , , Represents the last layer of the convolutional layer The output of the layer The updated static image The embedding vector of each node.

[0130] S3.2.3, Aggregate information: Static images middle The embedding vector of each node Perform splicing to generate graph embedding , the calculation process is as follows:

[0131] ,

[0132] ∈ ,

[0133] in, Represents a splicing operation, represents the graph embedding matrix, Indicates the number of nodes, Indicates the dimension of the node;

[0134] Integrate the graph embeddings of each static graph to get the graph embedding sequence , represents the graph embedding sequence matrix, and T represents the number of timestamps for time window division.

[0135] Embed the graph at each moment into a sequence Input to the sequence model TCN to embed the graph into the sequence Move in the past time sequence direction, through the convolution filter in the TCN model, the output of the filter and the graph are embedded in the sequence After the activation function, we finally get the future time embedding graph sequence , the specific calculation is as follows:

[0136] ,

[0137] ,

[0138] in, represents the output of the filter, The filter convolution kernels, represents the number of convolution kernels of the filter, represents the operation of the filter, represents the expansion factor, Indicates the past tense direction. Representing Graph Embeddings Graph embedding into the past temporal direction, represents the activation function, Indicates A future time embedding graph, , Indicates the number of timestamps that are moved according to the filter.

[0139] S5 is as follows:

[0140] Embedding a graph at a future time Select the starting node and the target node , the relationship between the two points is set to , embedding graph sequences from future time In Extract the starting node and the target node The corresponding starting embedding vector and the target embedding vector , according to the starting node and the target node The relationship between Find the diagonal matrix , the starting embedding vector , target embedding vector and the diagonal matrix Input to the decoder, the decoder uses the DistMult factorization method to decompose each potential edge The scoring calculation process is as follows:

[0141] ,

[0142] in, represents transpose, Indicates the starting node and the target node Through the relationship Likelihood score of connection, , , , Indicates Future time embedding graph The middle node is combined, Indicates Future time embedding graph China-US relations integration;

[0143] Will The results and threshold setting For comparison, When the future time point exists, and the water flow behavior that will occur between the two nodes is the relationship .

[0144] Example 2

[0145] like Figure 2 As shown, a water flow behavior prediction device in deep water high pile cap construction is composed as follows:

[0146] (1) Data acquisition unit:

[0147] Sensors: Various hydrological sensors deployed at key monitoring points to collect water level, flow velocity, and flow direction in real time;

[0148] Meteorological instruments: used to collect data on meteorological conditions that affect water flow, such as wind speed, direction, rainfall, and temperature;

[0149] Topographic scanning equipment: using radar or laser scanning technology to collect topographic maps and elevation data of the construction area and its surrounding areas;

[0150] (2) Data processing unit:

[0151] Central Processing Unit (CPU): responsible for processing the large amount of data collected:

[0152] Graphics Processing Unit (GPU): used to accelerate the computation of machine learning and deep learning algorithms, especially the training and application of dynamic graph convolutional neural networks:

[0153] Storage devices: used to store historical and real-time data, as well as model training and prediction results;

[0154] (3) Data communication unit:

[0155] Wireless transmission module: supports Wi-Fi and / or cellular networks to ensure that the device can receive and send data and communicate with the central monitoring system in real time;

[0156] Interface unit: mainly includes USB and Ethernet ports, providing the device with the ability to connect to other systems or devices;

[0157] (4) Display and interaction unit:

[0158] Display screen: used to display water flow prediction results and alarm information in real time:

[0159] Operator Interface: User-friendly interface allows operators to easily configure device settings, view data and reports.

[0160] Example 3

[0161] In order to demonstrate the beneficial effects of the present invention, the application effects of the AquaDynaGraph water flow behavior prediction model (ADG model for short) of the present invention and five existing methods in deepwater high pile cap construction were compared. The experimental scenario selected an actual deepwater high pile cap construction project area to monitor the water flow behavior in the area, including key hydrological parameters such as flow velocity, flow direction and water level; the data sources include real-time hydrological data provided by hydrological monitoring stations, meteorological data provided by meteorological departments, and topographic data;

[0162] The comparison methods include: (1) Linear Regression (LR): Linear regression analysis based on historical hydrological data; (2) ARIMA model (Autoregressive Integrated Moving Average, ARIMA): Time series analysis method based on autoregressive integrated moving average model; (3) LSTM network (Long Short-Term Memory, LSTM): Deep learning method based on long short-term memory network; (4) Support Vector Regression (SVR): Regression analysis method based on support vector machine; (5) Random Forest Regression (RF): Machine learning method based on decision tree ensemble;

[0163] The experimental steps include data collection, model training, model testing and result evaluation. In the data collection stage, historical hydrological data (including water level, flow velocity, flow direction, etc.), meteorological data (including rainfall, wind speed, wind direction, temperature, etc.), and topographic data in the monitoring area are collected, and the data are divided into a training set and a test set; in the model training stage, the training set data is used to train five existing method models and the ADG model of the present invention respectively, wherein the traditional method performs linear regression analysis based on static hydrological data, and the ARIMA model and LSTM network perform modeling and prediction based on time series data respectively. , SVR and random forest regression methods are modeled based on machine learning technology, while the ADG model extracts spatiotemporal features through a dynamic graph convolutional neural network and uses a temporal convolutional network to capture temporal dependencies; in the model testing phase, six methods are applied to the test set to predict water flow behavior, and the predicted flow velocity, flow direction and water level are recorded; in the result evaluation phase, the mean square error (MSE) and mean absolute error (MAE) are used to evaluate the prediction accuracy of the six methods, and the predicted values ​​are compared with the actual values ​​to evaluate the performance of the model in different time periods and extreme hydrological events. The experimental results are shown in Table 1.

[0164] Table 1 Comparison results between ADG in the present invention and existing methods

[0165]

[0166] It can be seen from the results in Table 1 that the ADG model of the present invention is significantly better than the five existing methods in terms of flow velocity, flow direction and water level prediction. Therefore, it can be proved that the ADG model of the present invention can show higher prediction accuracy and stability when dealing with special hydrological events (such as heavy rain). This is because the ADG model can integrate multi-source data in real time, capture the dynamic changes of water flow, and has stronger dynamic adaptability.

[0167] The experimental results fully demonstrate the beneficial effect of the AquaDynaGraph (ADG) water flow behavior prediction model in deep-water high-pile cap construction. Compared with the five existing methods, the ADG model significantly improves the accuracy and real-time performance of water flow prediction, providing a strong guarantee for construction safety and efficiency. By integrating multi-source data and dynamic adaptation mechanisms, the method of the present invention demonstrates its excellent performance and broad application prospects under complex hydrological conditions.

[0168] Although the above describes the specific implementation mode of the invention in conjunction with the drawings, it is not intended to limit the scope of protection of the invention. Based on the technical solution of the present invention, various modifications or variations that can be made by those skilled in the art without creative work are still within the scope of protection of the present invention.

Claims

1. A method for predicting water flow behavior in deep-water high-pile cap construction, characterized in that: The following steps are involved: S1. Data collection and preprocessing: Collect relevant hydrological data, meteorological data and topographic data of the construction site, and then preprocess the collected data; S2. Build dynamic graph: Build dynamic graph based on preprocessed data , and then for the dynamic graph Initialize data; Step S2 creates a dynamic graph based on the preprocessed data , as follows: S2.

1. Set dynamic image Updated time window: For short-term streamflow forecasts, select a time window with a time unit of days or hours, and for long-term streamflow forecasts, select a time window with a time unit of months or years; S2.

2. Divide the continuous hydrological monitoring data into a series of time windows by the timestamp of the real-time monitoring data, that is, divide the pre-processed data into time windows to obtain a set of static graph snapshots of the time series, that is, the dynamic graph , , Indicates the number of timestamps for time window division. Represents a dynamic graph Middle A static image, is a directed graph, each static graph Include node collection , edge set and relationship set , expressed as , Indicates A collection of nodes of a static graph, Indicates The edge set of a static graph, Indicates A static graph relationship set, ; S2.2.

1. Determine the content of each node set based on the hydrological measurement points nodes, represented by , Indicates The first static image nodes, each of which stores the pre-processed hydrological data, meteorological data, and topographic data of the corresponding hydrological measurement point, expressed as ,in Respectively expressed as The first static image The water level, flow velocity, flow direction, precipitation, wind speed, wind direction, temperature, topography, elevation data and soil type information in each node. ; S2.2.

2. According to The nodes determine the edge set, and an edge can be formed between any two points, expressed as , Indicates The first static image Strip edge, ; S2.2.

3. Determine the set of relationships between nodes based on the edges formed by the nodes. Each edge corresponds to a corresponding relationship. The types of relationships include water level change relationships. , velocity change relationship Relationship with flow direction change , expressed as , Indicates The first static image The relationship on the edge; S2.

3. Order the dynamic graph in chronological order Arrange the static images in the time window and select the static image of the first time window Initialize it as a node set Assign initial values ​​and set Establish edge set in middle node and relationship set , and then update the data of the subsequent time window static graph, initialize the data of the subsequent static graph in turn, and finally realize the whole dynamic graph Initialization of internal data; S3. Construct AG-RGCN and initialize the dynamic graph Input to AG-RGCN to get graph embedding sequence ; Step S3 is as follows: S3.

1. Construct an attention-based gated relational graph convolutional network, AG-RGCN for short. AG-RGCN includes input layer, embedding layer, convolution layer, aggregation layer and output layer. S3.

2. Initialize the dynamic graph Input to AG-RGCN: S3.2.

1. First, the initialized dynamic graph Input to the input layer of AG-RGCN, and the initialized dynamic graph is It is further transmitted to the embedding layer of AG-RGCN, and the dynamic graph is Each node in each static graph contains original features, and then converts the extracted original features into vector representation. The eigenvector of is represented as follows: ， in Representation Node The original features extracted are water level original features, flow velocity original features, flow direction original features, precipitation original features, wind speed original features, wind direction original features, temperature original features, topography original features, elevation data original features and soil type original features. , ; Then the nodes are transformed linearly The feature vector is mapped to the target dimension of the embedding vector , and obtain the feature vector , the calculation formula is as follows: ， in, represents the bias matrix, represents the embedding matrix; At the same time and Initialize and set Initialize to 0 vector, Initialize, The initialization is expressed as: ， in, represents the uniform distribution function, Indicates that from the interval Uniform sampling means that each value has the same probability of appearing in the interval. Represents the input feature vector Dimensions; S3.2.2, the convolutional layer extracts and updates features through a multi-layer gating mechanism and an attention mechanism, and the number of convolutional layers is set to , the feature vector of the node is updated once through a convolutional layer, except that the input of the first convolutional layer is Except for , the remaining layers use the output of the previous layer as the input of the next layer. The calculation process is as follows: ， ， ， ， ， ， in, Representation Node In the The embedding vector of the layer, Representation Node In the The embedding vector of the layer, Representation Node No. The embedding vector of the layer, represents the conversion gate, represents a nonlinear transformation operation, represents the weight matrix of the conversion gate, represents the bias vector, Represents the activation function Operation, Represents the activation function Operations; Representation Node In relationship ∈ The neighbor set below, Representation Node relation The neighbor nodes below , represents the number of layers in the neural network, , ; Indicates that the current layer is layer, Indicates that the previous layer is layer, Indicated in Neighbor nodes in the layer The characteristic vector of relation The weight matrix of express The real matrix of represents the dimension of the embedding vector, represents a global weight matrix that is used to transform the node's own features and does not depend on any specific relationship, Indicates Layer Node and neighbor nodes The learnable normalized attention weights between Indicates Layer Node and neighbor nodes The attention coefficient; Finally, each static image is obtained Each node The eigenvector of The corresponding updated node embedding vector , , , Represents the last layer of the convolutional layer The output of the layer The updated static image The embedding vector of each node; S3.2.3, Aggregate information: Static images middle The embedding vector of each node Perform splicing to generate graph embedding , the calculation process is as follows: ， ∈ ， in, Represents a splicing operation, represents the graph embedding matrix, Indicates the number of nodes, Indicates the dimension of the node; Integrate the graph embeddings of each static graph to get the graph embedding sequence , represents the graph embedding sequence matrix, T represents the number of timestamps for time window division; S4. Predicting water flow status: embedding the graph at each moment into the sequence Input to the sequence model TCN to obtain the future time embedding graph sequence ; S5. Generate future graphs to predict changes in water flow behavior: Use the decoder to embed future time graph sequences The future graph structure is predicted, and then the hydrological status at various time points in the future is predicted.

2. A method for predicting water flow behavior in deepwater high pile cap construction according to claim 1, characterized in that: S1 is as follows: S1.

1. Data collection: Hydrological data including water levels of relevant water bodies , flow rate and flow direction , obtained from hydrological monitoring sites; Weather data including precipitation , wind speed ,wind direction and temperature , obtained from the public data interface of the meteorological department and the meteorological station; Topographic data including topographic maps of the construction area and its upstream and downstream areas , elevation data and soil type , obtained through GIS database and MSI remote sensing data; S1.

2. Data preprocessing: including data cleaning and data standardization: S1.2.

1. Clean the collected data to remove or correct the noise, missing values ​​and outliers in the data; (1) Missing value processing: use the mean, median or interpolation method of the overall data to fill in the missing values; (2) Noise processing: noise reduction of data is performed through a moving average filter; (3) Outlier detection and processing: Detect outliers in the data using the triple standard deviation method, and choose to correct or delete them based on the actual situation. Then, smooth the outliers and use the average value of the neighboring points to replace the corrected or deleted outliers. S1.2.

2. Data standardization: After data cleaning, the data set is further standardized. The data of different dimensions are converted to the same scale through the z-score standardization method to finally obtain the preprocessed data.

3. A method for predicting water flow behavior in deepwater high pile cap construction according to claim 2, characterized in that: S4 is as follows: Embed the graph at each moment into a sequence Input to the sequence model TCN to embed the graph into the sequence Move in the past time sequence direction, through the convolution filter in the TCN model, the output of the filter and the graph are embedded in the sequence After the activation function, we finally get the future time embedding graph sequence , the specific calculation is as follows: ， ， in, represents the output of the filter, The filter convolution kernels, represents the number of convolution kernels of the filter, represents the operation of the filter, represents the expansion factor, Indicates the past tense direction. Representing Graph Embeddings Graph embedding into the past temporal direction, represents the activation function, Indicates A future time embedding graph, , Indicates the number of timestamps that are moved according to the filter.

4. A method for predicting water flow behavior in deepwater high pile cap construction according to claim 3, characterized in that: S5 is as follows: Embedding a graph at a future time Select the starting node and the target node , the relationship between the two points is set to , embedding graph sequences from future time In Extract the starting node and the target node The corresponding starting embedding vector and the target embedding vector , according to the starting node and the target node The relationship between Find the diagonal matrix , the starting embedding vector , target embedding vector and the diagonal matrix Input to the decoder, the decoder uses the DistMult factorization method to decompose each potential edge The scoring calculation process is as follows: ， in, represents transpose, Indicates the starting node and the target node Through the relationship Likelihood score of connection, , , , Indicates Future time embedding graph The middle node is combined, Indicates Future time embedding graph China-US relations integration; Will The results and threshold setting For comparison, When the future time point exists, and the water flow behavior that will occur between the two nodes is the relationship .

5. A device for predicting water flow behavior in deepwater high pile cap construction, which executes a method for predicting water flow behavior in deepwater high pile cap construction as claimed in any one of claims 1 to 4, characterized in that: The composition is as follows: (1) Data acquisition unit: Sensors: Various hydrological sensors deployed at key monitoring points to collect water level, flow velocity, and flow direction in real time; Meteorological instruments: used to collect data on meteorological conditions that affect water flow, such as wind speed, wind direction, rainfall, and temperature; Topographic scanning equipment: using radar or laser scanning technology to collect topographic maps and elevation data of the construction area and its surrounding areas; (2) Data processing unit: Central Processing Unit (CPU): responsible for processing the large amount of data collected: Graphics Processing Unit (GPU): used to accelerate the computation of machine learning and deep learning algorithms, specifically for the training and application of dynamic graph convolutional neural networks: Storage devices: used to store historical and real-time data, as well as model training and prediction results; (3) Data communication unit: Wireless transmission module: supports Wi-Fi and / or cellular networks to ensure that the device can receive and send data and communicate with the central monitoring system in real time; Interface unit: includes USB and Ethernet ports, providing the device with the ability to connect to other systems or devices; (4) Display and interaction unit: Display screen: used to display water flow prediction results and alarm information in real time: Operator Interface: User-friendly interface allows operators to easily configure device settings, view data and reports.

Images