Construction and evaluation method of drainage basin runoff forecasting model

By constructing a runoff forecast model based on GIS and hydrological meteorological data, using convolutional neural network and gated cycle unit combined with optimization algorithm, the problem of low runoff forecast accuracy in the existing technology is solved, and higher accuracy and adaptability are achieved .

CN119005413AInactive Publication Date: 2024-11-22XIAN UNIV OF TECH
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202411051931.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-01
Publication Date
2024-11-22
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

The accuracy of medium- and long-term weather forecasts and long-term rainfall forecasts in the prior art is not high, which affects the runoff forecasting accuracy. The research methods lack universality, which increases the complexity and difficulty of application.

Method used

The data set is constructed based on GIS data and hydrological meteorological data, the runoff forecast model is constructed using convolutional neural network and gated cyclic units, and the model is trained and evaluated through the TPE optimization algorithm and the Adam optimization algorithm, combining SG noise reduction and cube samples. The bar difference method performs data preprocessing.

Benefits of technology

It improves the generalization ability and adaptability of the model, can accurately capture the dynamic changes of floods and the temporal and spatial relationship, and improves the accuracy and adaptability of runoff forecasts.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119005413A_ABST
    Figure CN119005413A_ABST
Patent Text Reader

Abstract

The invention provides a construction and evaluation method of a drainage basin runoff forecasting model. The method comprises the following steps: constructing a data set based on GIS data and hydro meteorological data; preprocessing the data set; constructing a runoff forecasting model based on a convolutional neural network and a gated circulation unit; according to the preprocessed data set, training a runoff forecasting model by adopting a TPE optimization algorithm and an Asian optimization algorithm; performing model evaluation on the trained runoff forecasting model; and performing runoff simulation forecasting based on the evaluated runoff forecasting model. Dynamic changes and time-space relations of various data in the flood development process are captured through feature extraction of a convolution kernel and the time sequence processing capacity of a gating unit, model hyper-parameters are optimized through a TPE optimization algorithm, and the model hyper-parameters most suitable for a current data set are obtained; the constructed model has high adaptability and accuracy in the aspect of capturing flood dynamic changes.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of runoff prediction, and in particular to a method for constructing and evaluating a watershed runoff prediction model. Background Art

[0003] Floods are not only valuable freshwater resources, but also contain huge energy. If the runoff process can be obtained in advance and reasonably dispatched, it will produce huge benefits. First, the runoff forecast with a certain foreseeable period can provide important water information for reservoir power generation dispatching. By arranging the order of water storage and release in advance and dynamically adjusting the flood limit water level, the power generation of the power station can be increased. Secondly, timely and accurate runoff forecasts can help business personnel determine the time of flood occurrence and ensure the safety and benefits of reservoir water supply.

[0004] However, the accuracy of long-term weather forecasts is not high at present, so the accuracy of long-term rainfall forecasts is difficult to meet the requirements. The amount of runoff is directly related to the amount of rainfall, so the accuracy of long-term runoff forecasts is also affected. There are many methods for studying runoff prediction, but these methods are not universal, and no model can be applied to all situations. This leads to the need to select appropriate forecasting methods according to the characteristics of specific basins in practical applications, which increases the complexity and difficulty of application. Summary of the invention

[0005] In view of this, the present invention proposes a method for constructing and evaluating a watershed runoff forecast model to solve the problems existing in the above-mentioned prior art.

[0006] To achieve the above object, the present invention proposes a method for constructing and evaluating a watershed runoff forecast model, comprising:

[0007] Constructing a data set based on GIS data and hydrological and meteorological data; and preprocessing the data set;

[0008] Construct a runoff forecast model based on convolutional neural network and gated recurrent unit;

[0009] According to the preprocessed data set, the runoff forecast model is trained using the TPE optimization algorithm and the Adam optimization algorithm;

[0010] Conduct model evaluation on the runoff forecast model after training;

[0011] A runoff simulation forecast is performed based on the runoff forecast model after the evaluation is completed.

[0012] Optionally, the GIS data includes DEM data, watershed GIS data, river GIS data and hydrological station GIS data, and the hydrometeorological data includes hydrological station water condition data, rainfall station rainfall data and evaporation data.

[0013] Optionally, the process of preprocessing the data set includes:

[0014] The SG denoising and cubic spline difference method are used to perform smoothing and denoising processing on the hydrological and meteorological data to obtain the time process number of the hydrological and meteorological data;

[0015] Based on the GIS data, Python and GDAL are used to generate terrain, soil distribution and land use layers for convolutional neural network input;

[0016] The data set is divided into a training set and a test set according to the acquisition year.

[0017] Optionally, the process of constructing a runoff forecast model based on a convolutional neural network and a gated recurrent unit includes:

[0018] A convolutional neural network is used to extract features of rainfall station rainfall data and evaporation data in the hydrological and meteorological data to obtain data features;

[0019] Inputting the data features into the gated recurrent unit to obtain the change characteristics in the time series;

[0020] The change characteristics in the time series are combined with the data features to construct the runoff forecast model.

[0021] Optionally, in the process of extracting features from the rainfall station rainfall data and evaporation data in the hydrological and meteorological data using a convolutional neural network, a pooling layer is added after the convolution layer of the convolutional neural network.

[0022] Optionally, the convolution kernel in the convolutional neural network is calculated as follows:

[0023] l t =tanh(x t *k t )+b t

[0024] Among them, l t is the output value after convolution, tanh is the activation function, x t is the input vector, k t is the weight of the convolution kernel, b t is the bias value of the convolution kernel.

[0025] Optionally, during the training process, the gated recurrent unit retains key parameters in the data transmission process by resetting the gate and updating the gate. The calculation formula of the reset gate state is as follows:

[0026] r t =σ(Wr ·[h t-1 ,x t ])

[0027] Among them, r t To reset the gate state, W r To reset the gate weight value, h t-1 is the output variable at the previous moment, x t is the output vector at the current moment;

[0028] The calculation formula for updating the gate state is as follows:

[0029] z t =σ(W z ·[h t-1 ,x t ])

[0030] Among them, z t To update the door status.

[0031] Optionally, the process of training the runoff forecast model using the TPE optimization algorithm and the Adam optimization algorithm includes:

[0032] Based on the Optuna development tool, we build a decision tree-Bayesian function, use the Gaussian process as the probability model, use the decision tree structure to capture the interaction and importance between hyperparameters, and adaptively update the probability model of hyperparameters based on historical data;

[0033] The first-order moment estimate and the second-order moment estimate of the runoff forecast model are obtained based on the adaptive estimation of the contraction matrix, a dynamic learning rate is set for the runoff forecast model according to the first-order moment estimate and the second-order moment estimate, and the model parameters are continuously updated through each iteration.

[0034] Optionally, in the process of training the runoff forecast model using the TPE optimization algorithm and the Adam optimization algorithm, the number of convolutional layers, the number of GRU layers, the dropout rate, the time step, the number of feature maps, the convolution kernel size and the batch size are used as optimization parameters, and the loss of the test set is used as the objective function to tune the parameters of the runoff forecast model.

[0035] Optionally, the process of evaluating the runoff forecast model after training includes:

[0036] The mean absolute error, mean absolute percentage error and determination coefficient of the output results of the runoff forecast model are calculated respectively, and the learning effect and generalization ability of the runoff forecast model are evaluated based on the mean absolute error, mean absolute percentage error and determination coefficient.

[0037] Compared with the prior art, the present invention has the following beneficial effects:

[0038] The present invention uses the SG algorithm to smooth the original data, reducing the noise mixed in the original data, which is conducive to improving the generalization ability of the model. The feature extraction of the convolution kernel and the time series processing ability of the gating unit are used to capture the dynamic changes of various types of data and their temporal and spatial relationships during the development of the flood, and the TPE optimization algorithm is used to optimize the model hyperparameters to obtain the model hyperparameters that are most suitable for the current data set. The model described in the present invention has high adaptability and accuracy in capturing the dynamic changes of floods. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the detailed description of the preferred embodiments below. The accompanying drawings are only used for the purpose of illustrating the preferred embodiments and are not to be considered as limiting the present invention. In the accompanying drawings:

[0040] Figure 1 A schematic diagram of a flow chart of a method for constructing and evaluating a watershed runoff forecasting model according to the present invention;

[0041] Figure 2 A schematic diagram of a flow chart for constructing a runoff forecast model in an embodiment of the present invention;

[0042] Figure 3 Schematic diagram of the GRU model structure in an embodiment of the present invention;

[0043] Figure 4 is a training test loss graph in an embodiment of the present invention;

[0044] Figure 5 is the output accuracy of the model for the training set in the embodiment of the present invention;

[0045] Figure 6 It is the output accuracy of the model for the test set in the embodiment of the present invention. DETAILED DESCRIPTION

[0046] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided in order to enable a more thorough understanding of the present disclosure and to be able to fully convey the scope of the present disclosure to those skilled in the art. It should be noted that, in the absence of conflict, the embodiments of the present invention and the features described in the embodiments can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0047] This embodiment proposes a method for constructing and evaluating a watershed runoff forecast model. Figure 1 As shown, including:

[0048] Model Overview

[0049] The original data used are GIS data (DEM, watershed, river and hydrological station, etc.), hydrological and meteorological data from 2007 to 2012 (water data from hydrological stations, rainfall data from rainfall stations and evaporation data).

[0050] The model uses rainfall data, evaporation data, topographic maps, soil distribution maps, and land use maps as input data, and the flow rate of the hydrological station as the output of the model.

[0051] Model building:

[0052] The model of this embodiment uses part of the available data in the original data as the model input, and smoothes and reduces the noise of rainfall and evaporation data through SG noise reduction and multiple spline interpolation methods to obtain the time process data of various data. At the same time, based on the GIS data in the original data, Python and GDAL are used to generate terrain, soil distribution and land use layers that can be used for convolutional neural network input. When dividing the data set, the relevant data from 2007 to 2010 are used to train the model, and the flood data from 2011 to 2012 are used to test the generalization ability of the model.

[0053] When designing the model infrastructure, a convolutional neural network (CNN) was first used to extract features from rainfall data from 45 rain gauges and evaporation data from 6 meteorological stations to adaptively learn the potential relationship between these data. The features extracted by the CNN were then used as input to the gated recurrent unit (GRU) time series prediction model. The GRU learns the changing characteristics in the time series by updating its gated units, and then the output of the GRU layer is concatenated with the convolution results of the terrain, soil distribution, and land use layers using a fully connected layer, and finally a regression model between the input data and the target variable (i.e., flood flow) is constructed.

[0054] In order to ensure that the model has the best simulation accuracy and good generalization ability, the TPE algorithm based on decision tree and Bayesian optimization is used to optimize the model hyperparameters of the CNN-GRU hybrid network structure. 2 Model accuracy evaluation indicators such as sparse_ ...

[0055] The specific modeling process is as follows: Figure 2As shown in the figure. First, the original data is preprocessed by noise reduction, interpolation and normalization to ensure that the data meets the model input requirements. Secondly, the hyperparameter optimization technology and ADAM algorithm are used to optimize the model training process to improve the accuracy and generalization ability of the model in flood flow prediction. Then, the model is used to predict and analyze the flood event in 2013. Finally, the key influencing factors and scope of the hydrological station runoff are further determined by performing sensitivity analysis on the model input characteristics.

[0056] Convolutional Neural Networks

[0057] Convolutional neural network (CNN) is a deep learning algorithm commonly used in image recognition, computer vision, natural language processing and other fields. The main feature of CNN is that it can automatically extract features from data. Its local perception and weight sharing can greatly reduce the number of parameters, thereby improving the learning efficiency of the model. CNN mainly consists of three parts: convolution layer, pooling layer and fully connected layer. Through the alternating superposition of convolution layer and pooling layer and the use of ReLU activation function, the model can be quickly converged and redundant information can be filtered to obtain a fused feature description. Each convolution layer contains multiple convolution kernels, and the calculation formula is shown in formula (1). Due to the characteristics of the data, the dimension of the features to be extracted is very high. In order to solve this problem and reduce the cost of network training, a pooling layer is added after the convolution layer to reduce the feature dimension.

[0058] l t =tanh(x t *k t )+b t #(2-1)

[0059] Among them, l t is the output value after convolution, tanh is the activation function, x t is the input vector, k t is the weight of the convolution kernel, b t is the bias value of the convolution kernel.

[0060] GRU Gated Recurrent Unit

[0061] GRU proposed the theory of using two vectors, namely the reset gate and the update gate, to determine what kind of message needs to be sent to the output unit. It can retain information from a long time ago while removing irrelevant information. Compared with LSTM, GRU has smaller information changes, more convenient calculations, and faster training speed. The function of the control gate is to control which signals in the previous stage are input into the current candidate set h. th-tThe smaller the reset gate is, the smaller the signal input is in the previous stage. The update gate is used to control the degree to which the state information of the previous moment is introduced into the current state information. The larger the number of the update gate is, the more the current state information of the previous moment is introduced. In the process of backward transmission, the current memory content will be activated, and the activated update gate will control the amount of new information and the amount of past information saved. The new information and the past information complement each other. The GRU model structure is as follows Figure 3 shown.

[0062] GUR forward propagation public:

[0063] Reset gate formula:

[0064] r t =σ(W r ·[h t-1 ,x t ])

[0065] Update gate formula:

[0066] z = σ(W z ·[h t-1 ,x t ])

[0067] Current memory content:

[0068]

[0069] Final memory at the current time step:

[0070]

[0071] In the formula, x t 、h t-1 、h t 、r t , z t、 They are the input vector at the current moment, the output variable at the previous moment, the output variable at the current moment, the reset gate state, the update gate state, and the current candidate set state. z , W r , They are update gate, reset gate, candidate set and x respectively. t and h t-1 The weight value of the matrix product. σ and tanh represent the sigmoid function and tanh function respectively, where [] represents vector concatenation and * represents matrix product.

[0072] GRU uses various gate functions to preserve the key parameters of signal transmission, so that parameter information is not lost even during LONG-TERM transmission. In addition, compared with LSTM, GRU removes a gate function, so the training parameters are smaller than LSTM, and the overall training efficiency of GRU is higher than LSTM.

[0073] Genetic Algorithms

[0074] Genetic Algorithm (GA) is a search algorithm and optimization technique that simulates the biological evolution process. Inspired by Darwin's theory of evolution, it optimizes the solution of the problem by simulating biological mechanisms such as natural selection, inheritance, crossover and mutation. It starts with a set of candidate solutions (individuals), which form a population. Each individual has a fitness value, which represents its ability to solve the problem. Through the selection process, individuals with higher fitness have a greater chance of being retained and used to produce the next generation. Then the characteristics of the parent individuals are passed to the offspring through crossover and mutation. Crossover refers to the exchange of parts of the chromosomes of two individuals to produce new individuals with the characteristics of both parents. Mutation is another genetic operation that randomly changes one or more genes on the chromosomes of an individual, thereby introducing new genetic diversity. The fitness function is used to evaluate the quality of individuals and is a key component in genetic algorithms. It determines the probability that an individual is selected for reproduction. The population is a set of candidate solutions that evolve continuously during the iteration of the algorithm. In genetic algorithms, the solution to the problem needs to be encoded in the form of chromosomes, usually binary strings or other forms of encoding. It has inherent parallelism and uses a probabilistic search method, which makes it perform well in finding the optimal solution in the global search space.

[0075] Genetic algorithms do not guarantee that they can always find the global optimal solution, but they can usually find a solution close to the optimal one, and are particularly effective for complex problems. Therefore, they are widely used in engineering, scientific research, economics, art and other fields to solve various optimization and search problems.

[0076] TPE Optimization Algorithm

[0077] TPE (Tree-structured Parzen Estimator) is an algorithm for Bayesian optimization, which is a sequential model optimization strategy that is particularly suitable for high-cost black-box function optimization problems. TPE guides the selection of hyperparameters by building a probabilistic model, thereby improving optimization efficiency.

[0078] TPE uses Gaussian process as a probability model to approximate the distribution of the objective function, and uses a decision tree structure to capture the interaction and importance between hyperparameters. When selecting a combination of hyperparameters, it balances exploration (exploring new hyperparameter space areas) and utilization (using existing knowledge to improve the current best solution). It can adaptively update the probability model of hyperparameters based on historical data, thereby accurately guiding the selection of hyperparameters.

[0079] TPE dynamically adjusts the selection of hyperparameters by evaluating historical test results, thereby gradually approaching the optimal solution. TPE's performance is better than traditional grid search and random search methods in many cases, especially in high-dimensional hyperparameter spaces. Therefore, using the TPE algorithm to optimize the structure of the neural network to adapt it to the current data set is a more appropriate method.

[0080] ADAM Optimization

[0081] In order to avoid local minima in the algorithm and to improve the prediction speed and accuracy of the model, the algorithm of the model is optimized. The Adam algorithm is an extension of the stochastic gradient descent method. Based on the adaptive estimation of the conclusion matrix, the algorithm calculates the first-order moment estimation and second-order moment estimation of the model. According to the results, different dynamic learning rates are set for each model, and the model parameters are continuously updated through each iteration. Adam not only saves the historical gradient square exponential decay average v t , and also saves the past gradient exponential decay average m t , which is similar to the momentum generated when an object moves, is a vector, so the previous direction is retained to a certain extent during the update.

[0082] In practical applications, the Adam method works very well in most cases. Compared with other adaptive learning rate algorithms, it converges faster and has more effective learning efficiency. It can also correct problems existing in other optimization techniques, such as learning rate disappearance, slow convergence, or high variance parameter updates that lead to large fluctuations in the loss function.

[0083] The calculation formula is:

[0084] m t =β1m t-1 +(1-β1)g t

[0085]

[0086] Where: m t is the first-order momentum of the current gradient; v t is the second-order momentum of the current gradient; g tis the current gradient value; β1 is the first-order matrix exponential decay rate, the default value is 0.9; β2 is the second-order matrix exponential decay rate, the default value is 0.999.

[0087] Due to m t and v t The initial vector is zero, so the decay rate may be biased towards the zero vector at the initial time, so deviation correction is required. The formula is as follows:

[0088]

[0089] Where: is the weighted average of the corrected gradient; is the weighted biased variance of the corrected gradient.

[0090]

[0091] Where: ε is the default value of 10 -8 ; θ is the learning rate, and its initial value is 0.05.

[0092] CNN-GRU hybrid network structure

[0093] The CNN-GRU hybrid network model combines the features of convolutional neural networks (CNN) and gated recurrent units (GRU). It is a powerful deep learning architecture that is mainly used to process sequence data, especially when the input data contains both temporal and spatial features. This model performs well in video analysis, speech recognition, natural language processing, and time series prediction.

[0094] It mainly extracts the spatial features of the input data through CNN, extracts useful features from each time step, and enhances the model's feature extraction capability and model invariance (translation invariance, etc.) through multi-layer convolution and pooling. The features extracted by CNN are flattened and then the GRU layer learns the temporal dependency of its sequence data, thereby effectively capturing the long-term dependency in the input features.

[0095] The feature extraction capability of convolutional neural network (CNN) is combined with the ability of gated recurrent unit (GRU) to capture temporal features. CNN is used to learn the spatial correlation between rainfall data observed at rain gauges and evaporation data at meteorological stations. GRU is then used to learn the changing trends of various types of data. The CNN-GRU hybrid model that combines convolutional neural network and gated recurrent unit can not only adaptively capture the correlation between various input data, but also learn the potential changing trends in the data. It has the potential to extract features from various complex nonlinear rainfall and evaporation data and construct a regression model with flood flow.

[0096] TPE optimizes model hyperparameters

[0097] Hyperparameter optimization is usually an iterative and multi-dimensional search process. Reasonable model hyperparameter configuration can improve the performance of the model on the training and validation sets, and has the function of reducing the risk of overfitting and accelerating the convergence of the model. Through model hyperparameter optimization, the model structure is finally adapted to the current input data set.

[0098] In the process of hyperparameter optimization, Python was used as the development platform and Optuna as the development tool to construct a decision tree-Bayesian (TPE) function. The number of convolutional layers (n_cnn_layers), the number of GRU layers (n_gru_layers), droupout_rate, the time step of the time series (d), the number of feature maps (filters), the convolution kernel size (kernel_size) and the batch size (batchsize) were used as optimization parameters, and the loss of the test set was used as the objective function to tune the generalization ability of the model.

[0099] The final optimization results are shown in Table 6, where the three iterations with the lowest test loss are the 17th, 9th, and 48th optimizations, and the 15th optimization has the highest test loss of 0.02. The final values ​​of various hyperparameters are shown in the table, where the best value of the time step d is 67-74, and 72 is taken in the subsequent modeling process, that is, the various data of the previous 72 hours predict the flood flow at the next moment.

[0100] Table 6

[0101]

[0102] Model training strategy

[0103] During the model training process, the ADAM optimizer is used to adaptively adjust the learning rate to adapt to the gradient changes of different data, thereby improving training efficiency and model performance. At the same time, checkpoint technology and early stopping strategy are introduced to prevent overfitting. Checkpoint technology allows the state of the model to be saved during training so that training can be resumed from the most recent save point in the event of an accident. Early stopping technology observes the performance of the model on the validation set by setting a monitoring cycle (40 training cycles in this case); if the model performance does not improve within 40 consecutive cycles, the training will be terminated early to prevent the model from overfitting.

[0104] To ensure the effectiveness of model training, the data order is shuffled before each iteration by setting the 'shuffle' option of the dataset to 'True', which helps prevent the model from being overly dependent on the input order of the data and reduces the risk of data leakage. In addition, all features are standardized before the data is input into the model to ensure that all features are on the same scale, which helps optimize the model training process and prevents the model from ignoring certain key features during learning.

[0105] Model Accuracy Evaluation

[0106] It is often difficult to evaluate the fitting effect of the model by relying on a single error indicator. Therefore, this model accuracy evaluation adopts a multi-dimensional comprehensive evaluation method to provide more comprehensive guidance for model optimization and improvement. At the same time, the comprehensive evaluation can also help identify potential problems of the model under specific conditions and provide a basis for further adjustment of model parameters and structure. MAE (mean absolute error) is used to evaluate the difference between the model prediction value and the actual observation value in regression problems; MAPE provides a method to quantify the prediction error by calculating the percentage of the difference between the predicted value and the actual value to evaluate the performance of the model. Compared with MAE, MAPE gives greater weight to larger errors because it takes into account the relative size of the error; in addition, the coefficient of determination (R 2 ) to measure the linear correlation between the actual value and the predicted value. The specific calculation formula is as follows:

[0107]

[0108] In the calculation of MAE and MAPE:

[0109] N—the total number of observations in the data set;

[0110] P i —The actual value of the i-th observation;

[0111] Y i —The predicted or estimated value of the ith observation.

[0112]

[0113] in:

[0114] n—the total number of observations in the dataset;

[0115] y i —The actual value of the i-th observation;

[0116] —The predicted value or estimated value of the i-th observation, given by the regression model;

[0117] —The average of all actual observations.

[0118] This modeling uses the data of 12 floods from 2007 to 2010 as the training set, and the data of 4 floods from 2011 to 2012 as the test set. MAE, MAPE, MSE, R 2 To judge the fitting effect of the model, when the training set fitting effect is good, that is, the lower the MAE, MAPE, and MSE values, the better the R 2 The higher the value, the better the model's learning effect on the training samples. When the MAE, MAPE, and MSE of the test set are lower, the 2 When it is high, the model can also have a good fitting effect on the unlearned data set, indicating that the model has strong generalization ability.

[0119] Simulation results and evaluation

[0120] The training process loss image of the flood prediction model is as follows Figure 4 As shown in the figure, the loss (MSE) of the training set (Train) and the test set (Test) decreased rapidly in the first 10 epochs, and then the loss of the training set remained basically unchanged in 20-100 epochs, while the loss of the test set decreased slowly during this period, indicating that although the training accuracy of the model did not continue to improve during this process, the test effect of the model improved, that is, the generalization ability of the model is still improving. In the subsequent training process, the test accuracy remained basically unchanged, so the model stopped early at the 138th epoch. The overall accuracy of the training set and test set of the final model is shown in Figure 5 and Figure 6 As shown, the R 2 The R on the test set is 0.975. 2 It is 0.917, indicating that the model has good fitting effect and generalization ability.

[0121] The training set mainly includes 12 floods from 2007 to 2010, and the validation set includes four floods from 2011 to 2012. The fitting results of each flood are shown in Table 7.

[0122] In Table 7 (flood simulation results), in the training set, except for the 3rd and 8th flood simulations, the fitting R 2All are above 0.95, indicating that the model can learn the characteristics of these floods very well. When analyzing the flood simulation results, we noticed that the 8th flood event experienced many ups and downs during its flood development process, which may be too complicated, resulting in its poor accuracy. However, in the flood process simulation results, the simulation curves of the 3rd and 8th floods maintained a high consistency with the fluctuation trend of the actual observation data. Although there are certain deviations in the values ​​of the peak flow, the simulation results are basically consistent with the observed values ​​in terms of the time point of the peak flow and the overall trend of the water volume change. Based on these observations, it can be preliminarily concluded that the model adopted in this embodiment can effectively learn and reproduce the evolution characteristics of flood events on the training set, and achieve a satisfactory level of simulation accuracy, which shows that the model has high adaptability and accuracy in capturing the dynamic changes of floods.

[0123] Table 7

[0124]

[0125]

[0126] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for constructing and evaluating a watershed runoff forecast model, characterized in that: include: Construct datasets based on GIS data and hydrological and meteorological data; Preprocessing the data set; Construct a runoff forecast model based on convolutional neural network and gated recurrent unit; According to the preprocessed data set, the runoff forecast model is trained using the TPE optimization algorithm and the Adam optimization algorithm; Conduct model evaluation on the runoff forecast model after training; A runoff simulation forecast is performed based on the runoff forecast model after the evaluation is completed.

2. The method for constructing and evaluating a watershed runoff forecast model according to claim 1, characterized in that: The GIS data includes DEM data, watershed GIS data, river GIS data and hydrological station GIS data; the hydrological and meteorological data includes water condition data of hydrological stations, rainfall condition data of rainfall stations and evaporation data.

3. The method for constructing and evaluating a watershed runoff forecast model according to claim 1, characterized in that: The process of preprocessing the data set includes: The SG denoising and cubic spline difference method are used to perform smoothing and denoising processing on the hydrological and meteorological data to obtain the time process number of the hydrological and meteorological data; Based on the GIS data, Python and GDAL are used to generate terrain, soil distribution and land use layers for convolutional neural network input; The data set is divided into a training set and a test set according to the acquisition year.

4. The method for constructing and evaluating a watershed runoff forecast model according to claim 1, characterized in that: The process of building a runoff forecast model based on convolutional neural networks and gated recurrent units includes: A convolutional neural network is used to extract features of rainfall station rainfall data and evaporation data in the hydrological and meteorological data to obtain data features; Inputting the data features into the gated recurrent unit to obtain the change characteristics in the time series; The change characteristics in the time series are combined with the data features to construct the runoff forecast model.

5. The method for constructing and evaluating a watershed runoff forecast model according to claim 4, characterized in that: In the process of extracting features from the rainfall station rainfall data and evaporation data in the hydrological and meteorological data using a convolutional neural network, a pooling layer is added after the convolutional layer of the convolutional neural network.

6. The method for constructing and evaluating a watershed runoff forecast model according to claim 1, characterized in that: The calculation method of the convolution kernel in the convolutional neural network is as follows: l t =tanh(x t *k t )+b t Among them, l t is the output value after convolution, tanh is the activation function, x t is the input vector, k t is the weight of the convolution kernel, b t is the bias value of the convolution kernel.

7. The method for constructing and evaluating a watershed runoff forecast model according to claim 1, characterized in that: During the training process, the gated recurrent unit retains the key parameters in the data transmission process by resetting the gate and updating the gate. The calculation formula of the reset gate state is as follows: r t =σ(W r ·[h t-1 ,x t ]) Among them, r t To reset the gate state, W r To reset the gate weight value, h t-1 is the output variable at the previous moment, x t is the output vector at the current moment; The calculation formula for updating the gate state is as follows: z t =σ(W z ·[h t-1 ,x t ]) Among them, z t To update the door status.

8. The method for constructing and evaluating a watershed runoff forecast model according to claim 1, characterized in that: The process of training the runoff forecast model using the TPE optimization algorithm and the Adam optimization algorithm includes: Based on the Optuna development tool, we build a decision tree-Bayesian function, use the Gaussian process as the probability model, use the decision tree structure to capture the interaction and importance between hyperparameters, and adaptively update the probability model of hyperparameters based on historical data; The first-order moment estimate and the second-order moment estimate of the runoff forecast model are obtained based on the adaptive estimation of the contraction matrix, a dynamic learning rate is set for the runoff forecast model according to the first-order moment estimate and the second-order moment estimate, and the model parameters are continuously updated through each iteration.

9. The method for constructing and evaluating a watershed runoff forecast model according to claim 1, characterized in that: In the process of training the runoff forecast model using the TPE optimization algorithm and the Adam optimization algorithm, the number of convolutional layers, the number of GRU layers, the dropout rate, the time step, the number of feature maps, the convolution kernel size and the batch size are used as optimization parameters, and the loss of the test set is used as the objective function to tune the parameters of the runoff forecast model.

10. The method for constructing and evaluating a watershed runoff forecast model according to claim 1, characterized in that: The process of evaluating the runoff forecast model after training includes: The mean absolute error, mean absolute percentage error and determination coefficient of the output results of the runoff forecast model are calculated respectively, and the learning effect and generalization ability of the runoff forecast model are evaluated based on the mean absolute error, mean absolute percentage error and determination coefficient.

Citation Information

Cited By

  • Runoff prediction method fusing sparse neural network and quantum kernel function

    CN119249282A