Open channel gate flow prediction method based on multivariate data fusion driving

Abstract

The invention relates to the technical field of open channel gate flow prediction, and discloses an open channel gate flow prediction method based on multivariate data fusion driving, and the method comprises the following steps: carrying out the hour-level processing and preprocessing of actually measured water regimen data, and generating a basic data set; simulating gate water regimen data based on the hydrodynamic model, and constructing a supplementary data set; the basic data set and the supplementary data set are fused, and a BiLSTM multi-factor flow prediction model is constructed; inputting the fused data set into the model, and fitting a traffic prediction model through iterative training; and carrying out traffic prediction by using the trained model, and carrying out reverse normalization to output a final result. According to the method, the actually measured water regimen data and the hydrodynamic model simulation data are subjected to multivariate fusion, so that the problems of precision and robustness of a single data source in gate flow prediction are solved. The simulation data of the hydrodynamic model is used as an effective supplement for actually measured data, and comprehensive gate water regimen information is provided, so that the accuracy and generalization ability of the prediction model are improved.

Description

Technical Field

[0001] This invention relates to the field of open channel gate flow prediction technology, specifically to an open channel gate flow prediction method driven by multi-source data fusion. Background Technology

[0002] Flow prediction for open channel gates is a key task in water conservancy project management. Its accuracy is crucial for water allocation, flood control, and irrigation management in large-scale water transfer projects such as the South-to-North Water Diversion Project. Traditional prediction methods have been used for many years, providing basic support for water conservancy management. However, when faced with the complex situation of numerous hydraulic control conditions and variable hydraulic states in the main canal of the South-to-North Water Diversion Project, the limitations of existing technologies are gradually becoming apparent.

[0003] Currently, two main technical approaches are used for predicting the flow rate of open channel gates. One is a statistical analysis method based on historical measured data. This method collects and analyzes a large amount of historical gate operation data and uses statistical models such as regression and time series analysis for prediction. This method provides reliable prediction results when operating conditions are relatively stable and is easy to implement. The other approach is a physical model method based on hydrodynamics. For example, it uses fundamental physical equations such as the Saint-Venant equations to predict flow rate through numerical simulation of water flow. These models describe the water flow process from a physical perspective, providing a solid theoretical foundation for prediction. When specific boundary conditions are clearly defined, accurate physical simulations can be performed, and these models are widely used in the design and research of large-scale water diversion projects.

[0004] However, existing open channel gate flow prediction technologies rely excessively on measured data, and their prediction accuracy is highly limited by the completeness and representativeness of the data. When projects face large-flow water transfers or abnormal hydrological events, historical data samples are insufficient, leading to a significant decrease in the model's generalization ability and prediction accuracy. Furthermore, in attempts to use machine learning for prediction, traditional unidirectional time series models, such as unidirectional LSTM, can only process sequence information along a single time direction. Unidirectional models struggle to effectively capture this bidirectional dependency, further limiting prediction accuracy. Therefore, this invention provides an open channel gate flow prediction method based on multi-source data fusion to address the shortcomings of existing technologies. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method for predicting open channel gate flow based on multi-source data fusion. This method solves the problems in open channel gate flow prediction technology, such as the limitation of prediction accuracy and robustness caused by a single data source, the limitations of the applicability of empirical formulas and parameter calibration, and the difficulty of traditional time series models in capturing global contextual information.

[0006] To achieve the above objectives, the present invention provides the following technical solution: The first aspect of this invention provides a method for predicting open channel gate flow based on multi-source data fusion, comprising the following steps: S1. Historical Measured Data Collection and Preprocessing: Hourly-level processing and preprocessing of measured hydrological data for specific water conservancy projects to generate a basic dataset. Specifically, hourly-level processing can be performed on long-term historical hydrological data of large-scale water conservancy projects. Control gates in different scheduling zones can be selected, and their historical operating data can be collected as the basic dataset. Outlier filtering and data segmentation can then be performed on the data.

[0007] S2. Gate Hydrological Data Simulation Based on Hydrodynamic Model: Based on the hydrodynamic model, by determining boundary conditions, discretizing and solving the Saint-Venant equations, and calibrating the channel roughness, gate hydrological data is simulated to construct a supplementary dataset. This step further includes: using the Saint-Venant equations (including the continuity equation and the momentum equation) as the basic governing equations for one-dimensional flow motion in the river channel, and determining the boundary conditions of the hydraulic engineering; using the Preissmann four-point weighted implicit difference scheme to discretize and solve the Saint-Venant equations; and obtaining the water level and flow processes of each control gate under different operating conditions by calibrating the channel roughness and adaptively adjusting the spatiotemporal step size, thereby constructing a supplementary dataset. The formulas of the Saint-Venant equations are shown below: ; ; In the formula, The width of the cross-sectional area of ​​the water passage. For water level, For time, For traffic, The longitudinal distance of the channel along the main flow direction. The lateral inflow is measured per unit river length. This is the momentum correction factor. For the water flow area, It is the acceleration due to gravity. This results in a decrease in friction ratio.

[0008] S3. Construct a multi-factor traffic prediction model based on a BiLSTM neural network: Merge the basic dataset and the supplementary dataset to construct a multi-factor traffic prediction model based on a BiLSTM neural network. This model includes three parts: data preprocessing, the BiLSTM neural network, and model training. The data preprocessing step normalizes the fused complete dataset and, according to the input requirements of the BiLSTM neural network, constructs a multi-dimensional data structure with time-series characteristics. The normalization formula is as follows: ; In the formula, The data has been normalized. This is the original data. and These represent the maximum and minimum values ​​of the original data, respectively.

[0009] S4. Training the Neural Network Model Using Multivariate Data: The fused complete dataset is used as input to the multi-factor traffic prediction model, and a loss function is set. The model is then iteratively trained to fit the multi-factor traffic prediction model. This step evaluates the error between the model's prediction results and the complete dataset by setting the mean squared error (MSE) loss function, and uses the Adam optimizer for iterative training to minimize the loss function. The formula for calculating MSE is as follows: ; In the formula, For the sample size, For the first The actual flow value of each sample The model predicts the first The flow value of each sample.

[0010] S5. Prediction of flow through open channel gates: The trained multi-factor flow prediction model is used to predict the flow through the gates, and the prediction results are inversely normalized to output the final prediction result of the flow through the gates.

[0011] A second aspect of the present invention provides a channel gate flow prediction system based on multi-source data fusion, the system being used to implement the above method, comprising: The data processing module is used to preprocess the measured hydrological data to build a basic dataset; The data simulation module is used to simulate gate water condition data based on hydrodynamic models in order to build a supplementary dataset; The data fusion module is used to merge the basic dataset with the supplementary dataset to construct a complete dataset; The model training module is used to take the complete dataset as input to build a multi-factor traffic prediction model based on BiLSTM neural network, and obtain the trained multi-factor traffic prediction model through iterative training. The prediction output module is used to predict the gate throughput using the trained multi-factor flow prediction model, and to perform inverse normalization on the prediction results to output the final gate throughput prediction result.

[0012] This invention provides a method for predicting open channel gate flow based on multi-source data fusion. It has the following advantages: 1. This invention solves the problem of limited accuracy and robustness in gate flow prediction by fusing measured hydrological data with hydrodynamic model simulation data. The hydrodynamic model simulation data, as an effective supplement to the measured data, provides more comprehensive gate hydrological information, enabling the constructed BiLSTM neural network model to capture more complex physical laws and nonlinear characteristics, thereby improving the accuracy and generalization ability of the prediction model.

[0013] 2. This invention employs a BiLSTM neural network model for flow prediction, fully leveraging its advantages in processing time-series data. BiLSTM can perform bidirectional time-series feature extraction on historical gate water level data, capturing the dependencies between data points in the time series and avoiding the information loss problem that may exist in traditional unidirectional LSTM. This bidirectional feature extraction mechanism makes the model's interpretation of the gate's operating status more complete and the prediction results more accurate.

[0014] 3. This invention employs a modular approach to construct an open channel gate flow prediction system. Developed using Python, this system boasts high scalability and automation. The system only requires hydrological data and boundary condition information as input to automatically complete the calibration of the hydrodynamic model, the construction of a multivariate dataset, the training of the BiLSTM neural network model, and the final flow prediction. This modular design and automated execution process simplify the deployment of complex models and reduce the need for manual intervention. Attached Figure Description

[0015] Figure 1 This is a flowchart of a method for predicting open channel gate flow based on multi-source data fusion according to the present invention. Figure 2 This is a diagram of the BiLSTM model framework of the present invention; Figure 3 This is a schematic diagram illustrating the upstream and downstream modeling scope of the channel in this invention; Figure 4 This is an architecture diagram of an open channel gate flow prediction system based on multi-source data fusion driven by the present invention. Figure 5 This is a comparison chart of the prediction results of the BiLSTM model for the No. 1 control gate of the present invention; Figure 6 This is a comparison chart of the prediction results of the BiLSTM model for the No. 2 control gate of the present invention; Figure 7 This is a comparison chart of the prediction results of the BiLSTM model for the No. 3 control gate of the present invention. Detailed Implementation

[0016] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] See attached document Figure 1 Appendix Figure 2 and attached Figure 3 , Figure 1 This invention provides a flowchart of a method for predicting open channel gate flow based on multi-source data fusion, as an embodiment of the present invention. The invention provides a method for predicting open channel gate flow based on multi-source data fusion, comprising the following steps: S1. Historical measured data collection and preprocessing: Hourly processing of long-sequence historical measured water condition data for specific water conservancy projects, and outlier screening and data segmentation to generate a basic dataset.

[0018] S2. Simulation of gate water level data based on hydrodynamic model: Establish a one-dimensional hydrodynamic model of the river and canal, solve the Saint-Venant equations by discretization, calibrate the channel roughness, and simulate the water level and flow process of the control gate to construct a supplementary dataset.

[0019] S3. Construct a flow prediction model for open channel gates based on BiLSTM neural network: Based on the bidirectional long short-term memory (BiLSTM) neural network, construct a multi-factor flow prediction model that includes three parts: data preprocessing, BiLSTM network structure and model training.

[0020] S4. Training the neural network model using multivariate data: The obtained basic dataset and supplementary dataset are merged to form a complete dataset. This complete dataset is then used as input, and the constructed multifactor traffic prediction model is iteratively trained by setting a loss function and an optimizer.

[0021] S5. Prediction of flow through open channel gates: The trained multi-factor flow prediction model is used for actual prediction. The input data is of the same type as when the model was trained, and the output results of the model are inversely normalized to obtain the final gate flow prediction results.

[0022] This embodiment directly reflects the actual operating status of the gate at a specific point in time through measured hydrological data, forming the basis for prediction. The hydrological data simulated by the hydrodynamic model is calculated based on the laws of fluid physics (Saint-Venant's equations). It can not only verify and supplement the measured data in terms of physical mechanism, but also provide some flow data that is difficult to measure directly or is missing under certain operating conditions, thus forming a more comprehensive and information-rich driving dataset.

[0023] Furthermore, a BiLSTM neural network is adopted as the core of the prediction model. This network structure can perform bidirectional learning on the input time series data, both forward and backward, thereby capturing the temporal dependencies of the data. This bidirectional feature extraction mechanism enables the model to more completely understand the dynamic process of gate flow changing over time, and thus establish a precise nonlinear mapping relationship between the input hydrological data (including water levels before and after the gate, gate opening, simulated flow, etc.) and the final flow through the gate.

[0024] This invention also provides a channel gate flow prediction system based on multi-source data fusion, which is used to execute the above-described method. It includes: The data processing module is used to preprocess the measured hydrological data to build a basic dataset; The data simulation module is used to simulate gate water condition data based on hydrodynamic models in order to build a supplementary dataset; The data fusion module is used to merge the basic dataset with the supplementary dataset to construct a complete dataset.

[0025] The model training module receives the complete dataset output by the data fusion module to build and train a multi-factor traffic prediction model based on a BiLSTM neural network. The prediction output module uses the multi-factor flow prediction model trained by the model training module to predict and calculate the flow rate through the gate, and performs inverse normalization on the predicted values ​​to output the final prediction results.

[0026] For S1, hourly processing was performed on the historical long-sequence measured hydraulic engineering data of the Central Route Water Conservancy Project. This processing selected three control gates in different scheduling zones of the Central Route Project as the research objects, and collected the historical operation data of these three control gates as the basic dataset.

[0027] After data collection, the base dataset undergoes preprocessing. This preprocessing includes outlier filtering to exclude data containing outliers, such as null values. Simultaneously, the base dataset is divided, with one year's data used for model training and the other year's data used to validate the model's predictive performance. This data division ensures that the training and validation processes are independent, thereby objectively evaluating the model's performance.

[0028] For S2, a hydrodynamic model is established to simulate the gate's water level data, thereby constructing a supplementary dataset. Specifically, this involves collecting and organizing the basic engineering parameters within the modeling area and defining the modeling region. The external boundary conditions of the model utilize upstream flow and downstream water level processes. The internal boundaries of the model, such as pumping stations, control gates, inverted siphons, and transition sections, require special treatment. Based on the flow characteristics of these structures, two of the following are selected as governing equations: water level-flow relationship, momentum equation, energy equation, or continuity equation.

[0029] The Saint-Venant equations are used as the fundamental governing equations for one-dimensional water flow in rivers and canals. These equations include the continuity equation and the momentum equation, as shown below: ; ; In the formula, The width of the cross-sectional area of ​​the water passage. For water level, For time, For traffic, The longitudinal distance of the channel along the main flow direction. The lateral inflow is measured per unit river length. This is the momentum correction factor. For the water flow area, It is the acceleration due to gravity. This results in a decrease in friction ratio.

[0030] The Saint-Venant equations are discretized using the Preissmann four-point weighted implicit difference scheme. Under this scheme, the difference forms of the dependent variable and its spatial and temporal derivatives are as follows: ; ; ; In the formula, These are weighting coefficients, 0 ≤ ≤1; Represents the function value; and They represent the first and the One cross-section; and They represent the first and time; It is the time from the walk, in seconds. It is the distance from the walk in space, in meters (m). These are discrete points.

[0031] Based on the formulas for calculating the Courant number and the Frod number, a coordination relationship between the time step and the spatial step is established to achieve adaptive adjustment of the two step sizes. The specific formulas are as follows: ; ; ; In the formula, It is the Courant number; For Freud numbers; It is the time elapsed since the walk, measured in seconds (s). It is the distance from the walking distance in space, in meters; The speed of water flow in the channel; The velocity of the water flow; It is the acceleration due to gravity; The water depth is determined. Furthermore, for enclosed water conveyance structures such as inverted siphons and tunnels, energy equations are used for equivalent substitution to accelerate the model solution process. Through this process, the water level and flow processes of each control gate under different operating conditions can be solved, thereby constructing a supplementary dataset.

[0032] For S3, the aim is to build a multi-factor traffic prediction model based on a BiLSTM neural network. The model consists of three parts: data preprocessing, BiLSTM neural network structure, and model training.

[0033] Data preprocessing involves splitting the training set into input and output data. Input data contains the fundamental information used by the model for learning and prediction, including water level, gate opening data, and the flow process simulated by the hydrodynamic model. Output data is the target value for prediction, i.e., the gate throughput. After splitting, the data is normalized to adjust the data range to [0, 1] to unify the units and improve model convergence speed. The normalization formula is: ; In the formula, The data has been normalized. This is the original data. and These represent the maximum and minimum values ​​of the original data, respectively.

[0034] This embodiment performs single-point regression prediction, meaning it uses input data at a specific time point to predict the output data at that time point. Therefore, the time series length of the input data is set to 1. Furthermore, the batch size (batch_size), i.e., the number of data points processed in a single iteration, is set to 32. The number of features in the input data is determined by the data itself. Through these settings, the input data is processed into a three-dimensional tensor form to meet the input requirements of the deep learning model.

[0035] Model construction: The BiLSTM neural network used in this embodiment consists of two main parts: BiLSTM layers and output layers.

[0036] The BiLSTM layer primarily uses the LSTM module from the PyTorch library. The main parameters of the BiLSTM layer include input_size, hidden_neurons, and num_layers, corresponding to the number of input features, the number of neurons, and the number of neural network layers, respectively. This method sets these parameters to 5, 128, and 3, respectively, where the number of input features depends on the specific input data. The LSTM network internally contains two channels and three gating units (forgetting unit, input unit, and output unit). The lower channel is the short-term memory channel, used to store short-term information; the upper channel is the long-term memory channel, used to retain long-term memory information. The addition, modification, and discarding of information are accomplished through the three gating units. Data transfer uses the following formula: ; ; ; ; ; ; ; In the formula, This represents the Sigmoid activation function; This is the weight matrix. Here is the weight matrix for the forget gate; Here is the weight matrix of the input gate; This is the weight matrix for candidate memory units; This is the weight matrix of the output gate; This is the input vector at the current moment; These are the bias terms for the forgetting unit, input unit, cell state, and output gate, respectively. This is the hidden layer state; This is the final output. The output from the previous time step. By combining the current moment's long-term memory state and the information output by the output gate This yields the output of the hidden layer at the current time step. The BiLSTM structure can perform bidirectional temporal feature extraction on the input data, thereby improving the globality and completeness of temporal feature extraction.

[0037] The output layer primarily uses the fully connected layer module from the PyTorch library. The entire output layer consists of three fully connected layers, with the LeakyReLU activation function used between each pair of fully connected layers. The design of this structure aims to progressively reduce the number of features, ultimately outputting a single predicted value.

[0038] For model training, this embodiment defines a training function for training the deep learning model. This function receives the number of iterations, training data, a loss function, and a model parameter optimizer as input parameters. During training, the model is set to training mode. Then, by iteratively training the data, the input and target data are moved to a specified computing device (such as a GPU), and the dimensions of the input data are adjusted to match the model's input requirements. Next, forward propagation is performed to compute the model output, and the loss value is calculated using the loss function. The accumulated loss values ​​are then used to calculate the average loss. Afterward, the optimizer's gradients are cleared, backpropagation is performed to compute the gradients, and the model parameters are updated.

[0039] The model uses the mean squared error loss function (MSE) as the loss function, employs the Adam optimizer in the PyTorch framework to optimize the model parameters, and undergoes 100 iterations of training. The formula for the mean squared error loss function (MSE) is as follows: ; In the formula, For the sample size, and These are the actual value and the predicted value, respectively.

[0040] For S4, the aim is to train the constructed BiLSTM neural network model using the obtained base dataset and supplementary dataset.

[0041] By fusing historical measured hydraulic engineering data with hydrodynamic model simulation data, a complete dataset containing both measured and simulated data is constructed. This complete dataset is used as the training data for the model.

[0042] The complete dataset is input into the constructed BiLSTM neural network model. In this embodiment, the model loss function is set to mean squared error (MSE), the optimizer is set to Adam, and the number of training iterations is set to 100.

[0043] The training process is implemented through a defined training function. In each iteration of the training function, the model performs forward propagation and calculates the loss value. Then, the gradient is calculated through backpropagation, and the Adam optimizer updates the model parameters based on the gradient. This process is repeated 100 times until the model parameters converge.

[0044] Through the above training process, the model can learn the nonlinear mapping relationship between the gate flow rate and multiple input variables such as upstream water level, downstream water level, and gate opening from a complete dataset that integrates measured and simulated data. Finally, a trained BiLSTM neural network model for flow prediction is obtained.

[0045] S5 aims to predict the flow rate through open channel gates using a trained BiLSTM neural network model.

[0046] Real-time data, consistent with the data type used during model training, is input into the trained BiLSTM neural network model. This real-time data includes upstream and downstream water levels, gate openings, etc. After processing the input data, the model outputs normalized prediction results.

[0047] Subsequently, the model's output is denormalized. This denormalization is the inverse operation of normalization, used to restore the predicted values ​​from the normalized range of [0, 1] back to the actual physical dimensions of the original data, thus obtaining the final predicted flow rate through the open channel gate. The denormalization formula is: ; In the formula, This is the final predicted value; The normalized predicted value output by the model; and These represent the maximum and minimum values ​​of the original data, respectively.

[0048] Example 2 See attached document Figure 4 , Figure 4 This is an architecture diagram of an open channel gate flow prediction system based on multi-source data fusion driven by an embodiment of the present invention. This embodiment provides an open channel gate flow prediction system based on multi-source data fusion driven by the above-mentioned method, including a data processing module, a data simulation module, a data fusion module, a model training module, and a prediction output module.

[0049] Data Processing Module: This module is used to preprocess measured hydrological data to construct a basic dataset. Specifically, its functions include hourly processing of historical long-sequence measured hydrological data for large-scale water conservancy projects, selecting historical operation data of control gates in specific scheduling zones as the basic dataset, and performing outlier filtering and data partitioning on the basic dataset.

[0050] Data Simulation Module: This module is used to simulate gate water level data based on a hydrodynamic model to construct a supplementary dataset. Its functions include using the Saint-Venant equations as the fundamental governing equations for one-dimensional flow motion in rivers and canals, and determining the boundary conditions for hydraulic engineering projects; discretizing and solving the Saint-Venant equations using the Preissmann four-point weighted implicit difference scheme; and obtaining the water level and flow processes of each control gate under different operating conditions by calibrating the channel roughness and adaptively adjusting the spatiotemporal step size, thereby constructing a supplementary dataset.

[0051] Data Fusion Module: This module is used to merge the basic dataset output by the data processing module with the supplementary dataset output by the data simulation module to construct a complete dataset.

[0052] Model Training Module: This module takes the complete dataset output by the data fusion module as input to construct a multi-factor traffic prediction model based on a BiLSTM neural network, and obtains the trained multi-factor traffic prediction model through iterative training. Specifically, this module includes a data preprocessing section for normalizing the complete dataset and constructing a multi-dimensional data structure with time-series characteristics according to the input requirements of the BiLSTM neural network; it also includes a model training section for adjusting the neural network parameters through iterative training to fit the complete dataset. This module evaluates the error between the model's prediction results and the complete dataset by setting a mean squared error loss function, and uses the Adam optimizer for iterative training to minimize the loss function.

[0053] Prediction Output Module: This module is used to predict gate throughput using a trained multi-factor flow prediction model and performs inverse normalization on the prediction results to output the final gate throughput prediction result. Specifically, this module uses the trained multi-factor flow prediction model as the prediction tool, receives test set data for the period to be predicted as input, obtains the predicted value of gate throughput based on the model, and performs inverse normalization on the predicted value to output the prediction result converted into the actual physical value of the flow.

[0054] Example 3 See attached document Figure 5 Appendix Figure 6 and attached Figure 7 , Figure 5 , Figure 6 and Figure 7This is a model prediction effect diagram of an embodiment of the present invention.

[0055] To better illustrate the execution process of the method of this invention, this example takes three control gates of the South-to-North Water Diversion Project and two adjacent upstream and downstream canals as examples to construct and evaluate a one-dimensional hydrodynamic model and a BiLSTM neural network model.

[0056] In the one-dimensional hydrodynamic simulation, the modeling scope of the three gates and two channels is refined into two parts for separate calculation. Based on the upstream water level of the target gate and the flow rate and downstream water level of the upstream gate, the upstream flow rate of the target gate is calculated. Conversely, based on the downstream water level of the target gate and the flow rate and upstream water level of the downstream gate, the downstream flow rate of the target gate is calculated. This achieves multi-dimensional simulation of the gate's hydrological elements, striving to reproduce the true hydrological state of the gates. The hydrological data of the target gate includes the upstream and downstream water levels and gate opening data, along with the upstream and downstream flow rates from the one-dimensional hydrodynamic simulation, all serving as input data for the neural network model. Through iterative model training, a black-box relationship is constructed between the input hydrological data and the gate flow rate, jointly explaining the driving process of the gate flow rate.

[0057] Within the modeling scope of three selected control gates, the mean square error (MSE) of the measured flow rate through the target control gate versus the predicted flow rate is used as a benchmark for iterative training and parameter adjustment. A coefficient of determination (R²) is selected. 2 The mean square error (MSE), root mean square error (RMSE), and mean absolute error (MAE) were used as evaluation indicators to verify the accuracy of the model prediction results. The relevant evaluation indicator results for the three control gates are shown in Table 1.

[0058] Table 1: BiLSTM Model Prediction Evaluation Metrics

[0059] Analyzing the prediction results, we can find that the determination coefficient (R) of the three control gates... 2 A value between 0.93 and 0.99 indicates a high degree of model fit to the data and a good ability to explain data variability. Specifically, the R-value of control gate No. 1... 2 The highest value, reaching 0.99, indicates that its model has the most ideal fit to the data and can explain almost all the variability in the data; while the R-value of the No. 3 control gate is... 2 The value is 0.93, which is slightly lower than that of control gates No. 1 and No. 2, but still indicates that the model has a high degree of fit and can reflect the regularity of the data well.

[0060] Regarding the mean squared error (MSE), the MSE values ​​for the three control gates ranged from 5.87 to 22.95. A smaller MSE value indicates a smaller difference between the predicted and actual values, and thus higher prediction accuracy. Control gate No. 3 had the lowest MSE value at 5.87, indicating the smallest difference between its predicted and actual values, and the highest prediction accuracy. The root mean square error (RMSE), the square root of the MSE, measures the standard deviation between the predicted and actual values. The RMSE values ​​for the three control gates ranged from 2.42 to 4.79.

[0061] Finally, the mean absolute error (MAE) is used to measure the average difference between the predicted and actual values. The MAE values ​​for the three control gates range from 1.55 to 3.75. The smaller the MAE value, the higher the prediction accuracy. Control gate No. 3 has the lowest MAE value at 1.55, indicating that its average difference between the predicted and actual values ​​is the smallest, and its prediction accuracy is the highest. Control gate No. 2 has the highest MAE value at 3.75, the largest among the three control gates, but it is still at a relatively low level, demonstrating a good fit of the model.

[0062] In summary, the prediction models for the three control gates have a high overall fit. The open channel gate flow prediction method based on multivariate data fusion has demonstrated good results in practice and has the ability to make predictions in combination with actual water conditions and engineering conditions, thus possessing practical value.

[0063] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for predicting open channel gate flow based on multi-source data fusion, characterized in that, Includes the following steps: S1. Perform hourly-level processing and preprocessing on measured hydrological data for specific water conservancy projects to generate a basic dataset; S2. Based on the hydrodynamic model, by determining boundary conditions, discretizing and solving the Saint-Venant equations, and calibrating the channel roughness, simulate gate water condition data and construct a supplementary dataset. S3. Merge the basic dataset with the supplementary dataset to construct a multi-factor traffic prediction model based on BiLSTM neural network; S4. Use the fused complete dataset as input to the multi-factor traffic prediction model, set the loss function, and iteratively train to fit the multi-factor traffic prediction model. S5. The trained multi-factor flow prediction model is used to predict the gate flow rate, and the prediction results are inversely normalized to output the final gate flow rate prediction result.

2. The method for predicting open channel gate flow based on multi-source data fusion as described in claim 1, characterized in that, In step S1, the step of performing hourly processing and preprocessing of the measured hydrological data for a specific water conservancy project further includes: Hourly processing of historical hydraulic engineering data from large-scale water conservancy projects over long periods; Three control gates from different scheduling zones of a large-scale water conservancy project were selected, and historical operation data were collected as the basic dataset. Outlier filtering and data segmentation were then performed on the basic dataset.

3. The method for predicting open channel gate flow based on multi-source data fusion as described in claim 1, characterized in that, In step S2, the step of constructing a supplementary dataset by determining boundary conditions, discretely solving the Saint-Venant equations, calibrating the channel roughness, simulating gate hydrological data, and building a supplementary dataset further includes: The Saint-Venant equations are used as the basic governing equations for one-dimensional water flow in rivers and canals, and the boundary conditions for water conservancy projects are determined. The Saint-Venant equations are discretized and solved using the Preissmann four-point weighted implicit difference scheme. By calibrating the channel roughness and the adaptive adjustment of the spatiotemporal step, the water level and flow process of each control gate under different operating conditions are obtained, thereby constructing a supplementary dataset.

4. The method for predicting open channel gate flow based on multi-source data fusion as described in claim 1, characterized in that, In step S3, the step of constructing a multi-factor traffic prediction model based on a BiLSTM neural network further includes: The BiLSTM neural network model consists of three parts: data preprocessing, BiLSTM neural network, and model training. The data preprocessing normalizes the complete dataset obtained by fusing the basic dataset and the supplementary dataset, and forms a multidimensional data structure with time series characteristics according to the input requirements of the BiLSTM neural network. The model training is used to adjust the parameters of the BiLSTM neural network through iterative training to fit the complete dataset.

5. The method for predicting open channel gate flow based on multi-source data fusion as described in claim 1, characterized in that, In step S4, the step of using the fused complete dataset as input to the multi-factor traffic prediction model and setting a loss function, and iteratively training and fitting the multi-factor traffic prediction model, further includes: By setting a loss function, the error between the prediction results of the multi-factor traffic prediction model and the complete dataset is evaluated. By employing the Adam optimizer for iterative training to minimize the loss function, a multi-factor traffic prediction model that best fits the complete dataset after training is obtained.

6. The method for predicting open channel gate flow based on multi-source data fusion as described in claim 1, characterized in that, In step S5, the step of using the trained multi-factor flow prediction model for gate flow prediction and performing inverse normalization on the prediction results further includes: The trained multi-factor flow prediction model is used as a prediction tool. The test set data of the period to be predicted is received as input, and the predicted value of the gate flow is obtained based on the trained multi-factor flow prediction model. The predicted gate flow rate is inversely normalized to obtain the actual physical flow rate, and then the converted gate flow rate physical value is output.

7. The method for predicting open channel gate flow based on multi-source data fusion as described in claim 4, characterized in that, The data preprocessing also includes dividing the dataset into a training dataset and a validation dataset. The training dataset is used for training the multi-factor traffic prediction model, and the validation dataset is used for validating the multi-factor traffic prediction model.

8. The method for predicting open channel gate flow based on multi-source data fusion as described in claim 4, characterized in that, The BiLSTM neural network consists of multiple BiLSTM layers and an output layer composed of multiple fully connected layers, with the LeakyReLU activation function used between every two fully connected layers.

9. The method for predicting open channel gate flow based on multi-source data fusion as described in claim 5, characterized in that, The error between the prediction results of the multi-factor traffic prediction model and the complete dataset is evaluated using the mean squared error loss function, which is calculated using the following formula: ； In the formula, For the sample size, For the first The actual flow value of each sample The model predicts the first The flow value of each sample.

10. A multi-source data fusion-driven open channel gate flow prediction system, applied to the multi-source data fusion-driven open channel gate flow prediction method according to any one of claims 1-9, characterized in that, include: The data processing module is used to preprocess the measured hydrological data to build a basic dataset; The data simulation module is used to simulate gate water condition data based on hydrodynamic models in order to build a supplementary dataset; The data fusion module is used to merge the basic dataset with the supplementary dataset to construct a complete dataset; The model training module is used to take the complete dataset as input to build a multi-factor traffic prediction model based on BiLSTM neural network, and obtain the trained multi-factor traffic prediction model through iterative training. The prediction output module is used to predict the gate throughput using the trained multi-factor flow prediction model, and to perform inverse normalization on the prediction results to output the final gate throughput prediction result.

Images