River flow evolution method and system of physical mechanism collaborative deep learning algorithm

Abstract

The invention discloses a channel flow evolution method and system based on a physical mechanism collaborative deep learning algorithm, and relates to the technical field of channel flow evolution, and the method comprises the steps: carrying out the data collection and preprocessing; carrying out feature extraction and constructing an evolution model; deducing a hydrological physical model through simultaneous operation of a Muskinggen tank storage curve equation and a river reach water balance equation; the section flow of each river reach is predicted through the hydrological physical model and the evolution model; and through the prediction result of the index evaluation model in each river reach, carrying out collaborative optimization intelligent evolution of each river reach through the evaluation result. The method guides the establishment of the constrained deep learning neural network model through a hydrological physical mechanism, organically fuses the hydrological physical model and the deep learning algorithm model, comprehensively applies the physical model and the mathematical model to better solve the hydrological problem of a complex system, gives full play to the respective advantages, and improves the reliability of the system. The method is suitable for flow simulation prediction of various river channels, operation is easy and convenient, training is fast, and results are reliable.

Description

Technical Field

[0001] The present invention relates to the field of river flow evolution technology, and more specifically to a river flow evolution method and system that combines a physical mechanism with a deep learning algorithm. Background Art

[0002] River flow refers to the volume of water passing through a certain cross-section of a river per unit time. It is an important indicator of the hydrological cycle and a basic component of water balance. The evolution of river flow is the coordination point of the water inflow process, water diversion process, water withdrawal process, and reservoir scheduling process. The evolution of flow is the basis and core of runoff prediction and flood forecasting. The evolution process is affected by many factors such as interval inflow, interval outflow, river morphology, riverbed properties, groundwater level, hydrometeorology, etc. The basic hydrophysical principles of river flow evolution are as follows: Figure 1 shown.

[0003] At present, the methods used for river flow evolution are mainly hydrology and hydrodynamics based on physical mechanisms. The two methods use the water balance principle and the momentum kinetic energy theorem respectively. The physical model is established based on physical theory and experimental data. By modeling the physical structure and characteristics of the system, it intuitively reflects the properties and characteristics of the real system. It has the advantages of strong interpretability and clear structural logic, and its application is relatively mature and widespread. However, traditional physical models always have some limitations:

[0004] (1) The physical model has a complex structure. Model establishment usually requires a large number of parameters. The determination of these parameters depends on a large amount of measured data, including hydrological data, meteorological data, topographic data, river network data, river channel characteristic data, underground aquifer data, etc. It is often impossible to establish a complete and perfect model due to lack of data, inaccurate data or data mismatch. The modeling workload is large, the difficulty is high and the cycle is long.

[0005] (2) Physical models are limited by the physical structure and characteristics of the actual system and may not be able to accurately describe complex systems. Even if a complete physical model is constructed on the surface, the simulation results may be quite different from the actual physical process. The uncertainty of the parameters cannot guarantee a good fitting effect, and the model accuracy needs to be improved.

[0006] With the rapid development of human society and economy, the evolution mechanism of climate and environment has become more and more complex. Traditional hydrological physical models can no longer meet the needs of refined management and precise scheduling. The limitations of physical models themselves have also become a constraint on their service to social management. At the same time, deep learning is becoming an important driving force for a new round of scientific and technological revolution and industrial transformation. In recent years, deep learning has been widely used in fields such as data mining, speech recognition, machine translation, and robotics, demonstrating its powerful ability to process sequence data. Considering physical constraints in modeling is a current research hotspot and difficulty. With its powerful feature representation and learning capabilities, deep learning algorithm models have increasingly become a technical support for solving physical equations, exploring physical laws, and deducing the physical world.

[0007] Therefore, how to propose a river flow evolution method and system that coordinates physical mechanisms with deep learning algorithms, guide the establishment of deep learning neural network models through hydrological physical mechanisms, organically integrate hydrological physical models with deep learning algorithm models, and comprehensively apply physical models and mathematical models to better solve hydrological problems of complex systems are issues that technical personnel in this field urgently need to solve. Summary of the invention

[0008] In view of this, the present invention provides a river flow evolution method and system that cooperates with a physical mechanism and a deep learning algorithm. The hydrological physical mechanism guides the establishment of a constraint deep learning neural network model, organically integrates the hydrological physical model with the deep learning algorithm model, and comprehensively applies the physical model and the mathematical model to better solve the hydrological problems of complex systems. In order to achieve the above purpose, the present invention adopts the following technical solutions:

[0009] A river flow evolution method based on physical mechanism and deep learning algorithm, including:

[0010] Collect hydrological data of river sections, build basic data sets, and perform data preprocessing;

[0011] The preprocessed data is subjected to feature extraction through convolutional neural networks, and the extracted features are used to construct an evolution model based on long short-term memory neural networks;

[0012] The hydrological physical model is derived by combining the Muskingum trough storage curve equation and the river section water balance equation.

[0013] The cross-sectional flow of each river section is predicted by hydrophysical model and evolution model respectively;

[0014] The prediction results of the model in each river section are evaluated through indicators, and the collaborative optimization and intelligent evolution of each river section are carried out based on the evaluation results.

[0015] Optionally, collecting hydrological data of a river section and constructing a basic data set includes:

[0016] Obtain natural geography, hydrology, meteorology, and socio-economic data and information through surveys, and transform survey data into various readable and writable basic data through information integration and extraction;

[0017] Through literature research, on-site investigation, statistical analysis and comprehensive evaluation, the collected data forms are identified, the required information is extracted, the obtained data are checked and unified, and data are updated and maintained regularly to build a basic data set.

[0018] Optionally, the data preprocessing includes: performing data cleaning, checking data consistency and compliance with data logical relationships, processing missing values, invalid values, and outlier data through interpolation, envelope constraints, and mean value replacement methods, and generating data that meets data training quality requirements.

[0019] Optionally, the derivation of the hydrological physical model by jointly deriving the Muskingum Trough storage curve equation and the river section water balance equation includes:

[0020] Muskingum root groove storage curve equation: S = KQ' = K(xI+(1-x)Q);

[0021] Among them, x is the weight factor of the inflow and outflow on the tank storage, S is the tank storage, K is the tank storage coefficient, Q′ is the indicated storage flow, I is the upper section flow, and Q is the lower section flow;

[0022] Water balance equation for a river section:

[0023] Among them, S1 and S2 are the tank storage at the upper and lower moments respectively, Δt represents the time step, I1 and I2 are the flow rates at the upper and lower moments of the upper section respectively, and Q1 and Q2 are the flow rates at the upper and lower moments of the lower section respectively;

[0024] By combining the above two equations, the weight factor x is taken as 0.5, the influence of water flow propagation time is considered, the channel storage coefficient K is taken as the propagation time τ, and the interval water inflow, interval water withdrawal, interval water diversion, interval water loss and interval unbalanced water volume parameters are considered to obtain the flow evolution equation based on the hydrological and physical mechanism.

[0025] Optionally, the flow evolution equation is a flow calculation formula for a certain section in the river:

[0026]

[0027] Among them, I jt represents the flow rate at section j at time t, I j-1，t is the flow rate of section j-1 at time t, I j-1，t-1 is the flow rate of section j-1 at time t-1, τ jis the propagation time from section j-1 to section j, Δt is the time step, Q 区jt represents the interval inflow between the j-1th node and the jth node at time t, Q 退jt represents the amount of water withdrawal between the j-1th node and the jth node at time t, Q 引jt represents the water diversion volume between the j-1th node and the jth node at time t, Q 损jt represents the loss between the j-1th node and the jth node at time t, Q 平j,t Represents the unbalanced water volume between the j-1th node and the jth node at time t.

[0028] Optionally, the evolution model predicts the cross-sectional flow of each river section including: dividing each river section into two types, namely, a river section without a reservoir and a river section including a reservoir.

[0029] Optionally, it also includes: selecting root mean square error and mean relative error indicators to evaluate the prediction results of the evolution model;

[0030]

[0031] Among them, x i is the measured value, y i is the predicted value, N is the number of measured values; RMSE is the root mean square error;

[0032]

[0033] Among them, x i is the measured value, y i is the predicted value, N is the number of measured values; MRE is the mean relative error.

[0034] Optionally, it also includes screening the parameters of the evolution model according to the evaluation indicators:

[0035] The number of training times for each river section is set to m times. After the training, the RMSE and MRE evaluation indicators of the m samples are calculated. First, the top n RMSE evaluation indicators among the m samples are selected as candidate samples. Secondly, the samples with the best MRE performance are selected from the candidate samples and stored as the final evolution model parameter information file of the river section.

[0036] Optionally, the evolution model predicts the cross-sectional flow of each river section including:

[0037] Through the evolution model, the river section models are connected in series in the order of upstream and downstream to simulate the evolution process of river flow;

[0038] Each river section is a separate model unit, and the corresponding model parameter files are stored. The river section models are simulated in series in the order of upstream and downstream.

[0039] Before simulation and prediction, the corresponding data are input into the evolution model according to the input item list of the river section type. First, the data of the first river section is input, and the model parameter file of the first river section is called through the Python language. The flow process of the lower section of the first river section is simulated and used as the flow process of the upper section of the second river section, and so on, until the flow process of each section downstream is obtained.

[0040] Optionally, a river flow evolution system using a physical mechanism and a deep learning algorithm, including:

[0041] Collection and processing module: used to collect hydrological data of river sections, build basic data sets, and perform data preprocessing;

[0042] Model building module: used to extract features from preprocessed data through convolutional neural networks, and use the extracted features to build an evolutionary model based on long short-term memory neural networks;

[0043] Model derivation module: used to derive the hydrological physical model by combining the Muskingum trough storage curve equation and the river section water balance equation;

[0044] Prediction module: used to predict the cross-sectional flow of each river section through hydrological physical model and evolution model respectively;

[0045] Evolution module: used to evaluate the prediction results of the model in each river section through indicators, and to carry out intelligent evolution of collaborative optimization of each river section based on the evaluation results.

[0046] It can be seen from the above technical solutions that, compared with the prior art, the present invention discloses a river flow evolution method and system that uses a physical mechanism to collaborate with a deep learning algorithm, which has the following beneficial effects:

[0047] The present invention proposes a river flow evolution method of physical mechanism and deep learning algorithm, including: collecting hydrological data of river sections, constructing basic data sets, and preprocessing data; extracting features of preprocessed data through convolutional neural network, and constructing evolution model based on long short-term memory neural network by using extracted features; deriving hydrological physical model by Muskingum root channel storage curve equation and river section water balance equation; predicting cross-sectional flow of each river section through hydrological physical model and evolution model respectively; evaluating the prediction results of each river section through index evaluation model, and performing collaborative optimization intelligent evolution of each river section through evaluation results. The present invention guides the establishment of deep learning neural network model through hydrological physical mechanism, organically integrates hydrological physical model with deep learning algorithm model, and proposes a high-performance and high-precision river flow evolution method. The comprehensive application of physical model and mathematical model can better solve the hydrological problems of complex systems, give full play to their respective advantages, and the designed neural network model structure is suitable for flow simulation and prediction of various types of rivers as long as the data volume meets the requirements. It is easy to operate, fast to train, and reliable in results, which promotes the development and application of hydrological mechanics simulation technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0049] Figure 1 The basic hydrophysical principle diagram of river flow evolution provided by the present invention.

[0050] Figure 2 This is a schematic diagram of the physical equation for river flow evolution provided by the present invention.

[0051] Figure 3 The present invention provides a data collection principle.

[0052] Figure 4 This is a schematic diagram of the data cleaning principle provided by the present invention.

[0053] Figure 5 This is a schematic diagram of the model structure provided by the present invention.

[0054] Figure 6 The present invention provides a schematic diagram of the model parameter file storage principle.

[0055] Figure 7 This is a schematic diagram of the simulation prediction principle provided by the present invention.

[0056] Figure 8This is the generalized diagram of the case river node provided by the present invention.

[0057] Fig. 9 This is a comparison bar chart of MRE during the test period provided by the present invention.

[0058] Fig.10 This is the intelligent evolution result of the gantry section provided by the present invention.

[0059] Fig.11 A schematic flow chart of a river flow evolution method that combines a physical mechanism with a deep learning algorithm provided by the present invention.

[0060] Fig.12 A structural framework diagram of a river flow evolution system that combines a physical mechanism with a deep learning algorithm provided by the present invention. DETAILED DESCRIPTION

[0061] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0062] The embodiment of the present invention discloses a river flow evolution method using a physical mechanism and a deep learning algorithm, such as Fig.11 As shown, including:

[0063] Collect hydrological data of river sections, build basic data sets, and perform data preprocessing;

[0064] The preprocessed data is subjected to feature extraction through convolutional neural networks, and the extracted features are used to construct an evolution model based on long short-term memory neural networks;

[0065] The hydrological physical model is derived by combining the Muskingum trough storage curve equation and the river section water balance equation.

[0066] The cross-sectional flow of each river section is predicted by hydrophysical model and evolution model respectively;

[0067] The prediction results of the model in each river section are evaluated through indicators, and the collaborative optimization and intelligent evolution of each river section are carried out based on the evaluation results.

[0068] Furthermore, the collection of hydrological data of the river section and the construction of the basic data set include:

[0069] Obtain natural geography, hydrology, meteorology, and socio-economic data and information through surveys, and transform survey data into various readable and writable basic data through information integration and extraction;

[0070] Through literature research, on-site investigation, statistical analysis and comprehensive evaluation, the collected data forms are identified, the required information is extracted, the obtained data are checked and unified, and data are updated and maintained regularly to build a basic data set.

[0071] Furthermore, the data preprocessing includes: performing data cleaning, checking data consistency and compliance with data logical relationships, processing missing values, invalid values, and abnormal value data through interpolation, envelope constraints, and average value replacement methods, and generating data that meets data training quality requirements.

[0072] Furthermore, the hydrological physical model is derived by combining the Muskingum trough storage curve equation and the river section water balance equation, including:

[0073] Muskingum root groove storage curve equation: S = KQ' = K(xI+(1-x)Q);

[0074] Among them, x is the weight factor of the inflow and outflow on the tank storage, S is the tank storage capacity, K is the tank storage coefficient, Q′ is the indicated storage flow, I is the upper section flow, and Q is the lower section flow;

[0075] Water balance equation for a river section:

[0076] Among them, S1 and S2 are the tank storage at the upper and lower moments respectively, Δt represents the time step, I1 and I2 are the flow rates at the upper and lower moments of the upper section respectively, and Q1 and Q2 are the flow rates at the upper and lower moments of the lower section respectively;

[0077] By combining the above two equations, the weight factor x is taken as 0.5, the influence of water flow propagation time is considered, the channel storage coefficient K is taken as the propagation time τ, and the interval water inflow, interval water withdrawal, interval water diversion, interval water loss and interval unbalanced water volume parameters are considered to obtain the flow evolution equation based on the hydrological and physical mechanism.

[0078] Furthermore, the flow evolution equation includes:

[0079] The flow calculation formula for a certain section in the river is as follows:

[0080]

[0081] Among them, I jt represents the flow rate at section j at time t, I j-1，t is the flow rate of section j-1 at time t, I j-1，t-1 is the flow rate of section j-1 at time t-1, τ j is the propagation time from section j-1 to section j, Δt is the time step, Q 区jt represents the interval inflow between the j-1th node and the jth node at time t, Q退jt represents the amount of water withdrawal between the j-1th node and the jth node at time t, Q 引jt represents the water diversion volume between the j-1th node and the jth node at time t, Q 损jt represents the loss between the j-1th node and the jth node at time t, Q 平j,t It represents the unbalanced water volume between the j-1th node and the jth node at time t. The physical equation of river flow evolution is as follows: Figure 2 shown.

[0082] Furthermore, the evolution model predicts the cross-sectional flow of each river section by dividing each river section into two types, namely, a river section without a reservoir and a river section with a reservoir.

[0083] Furthermore, it also includes: selecting root mean square error and mean relative error indicators to evaluate the prediction results of the evolution model;

[0084]

[0085] Among them, x i is the measured value, y i is the predicted value, N is the number of measured values; RMSE is the root mean square error;

[0086]

[0087] Among them, x i is the measured value, y i is the predicted value, N is the number of measured values; MRE is the mean relative error.

[0088] Furthermore, it also includes screening the parameters of the evolution model according to the evaluation indicators:

[0089] The number of training times for each river section is set to m times. After the training, the RMSE and MRE evaluation indicators of the m samples are calculated. First, the top n RMSE evaluation indicators among the m samples are selected as candidate samples. Secondly, the samples with the best MRE performance are selected from the candidate samples and stored as the final evolution model parameter information file of the river section.

[0090] Furthermore, the evolution model predicts the cross-sectional flow of each river section including:

[0091] Through the evolution model, the river section models are connected in series in the order of upstream and downstream to simulate the evolution process of river flow;

[0092] Each river section is a separate model unit, and the corresponding model parameter files are stored. The river section models are simulated in series in the order of upstream and downstream.

[0093] Before simulation and prediction, the corresponding data are input into the evolution model according to the input item list of the river section type. First, the data of the first river section is input, and the model parameter file of the first river section is called through the Python language. The flow process of the lower section of the first river section is simulated and used as the flow process of the upper section of the second river section, and so on, until the flow process of each section downstream is obtained.

[0094] In a specific implementation, the specific steps of extracting features from the preprocessed data by a convolutional neural network and constructing an evolution model based on a long short-term memory neural network using the extracted features are:

[0095] The modeling process of the river flow evolution model based on the deep learning algorithm is as follows: data collection-data cleaning-learning and training-model storage-simulation prediction. Data collection and data cleaning are achieved through manual recognition combined with mathematical statistics; learning and training are achieved through bidirectional convolutional neural network long short-term memory (CNN-LSTM); model storage is achieved by saving the model parameter information file through Python language; simulation prediction is achieved by serially calling the model parameter file through Python language.

[0096] (1) Data collection

[0097] The collection and organization of data is the basis and prerequisite for the establishment of deep learning algorithm models. After mining the inherent laws in historical data, the future evolution process can be predicted. The longer the historical data series, the better the model captures the data features and the higher the model accuracy. Data collection relies on surveys to obtain a large amount of natural geography, hydrology, meteorology, socio-economic data and other information, and transforms the survey data into various types of readable and writable basic data through information integration and extraction. The main collection methods used are: literature research method, field survey method, statistical analysis method, comprehensive evaluation method, etc., to identify the various forms of collected data, extract the required information, check and unify the obtained data, and regularly update and maintain the data to ensure the timeliness, accuracy and completeness of the data, thereby forming a basic data set. The principles of data collection are as follows: Figure 3 shown.

[0098] (2) Data cleaning

[0099] Data cleaning refers to the last step of discovering and correcting identifiable errors in data files, including checking data consistency and data logical relationships, processing missing values, invalid values, outliers and other data through methods such as manual identification, interpolation, envelope constraints, and average value substitution, to generate data that meets the data training quality requirements. The principles of data cleaning are as follows: Figure 4 shown.

[0100] (3) Learning and training

[0101] Learning and training are implemented in two steps: feature extraction and deep learning. Feature extraction uses a convolutional neural network model (hereinafter referred to as "CNN"), and deep learning uses a long short-term memory artificial neural network model (hereinafter referred to as "LSTM"). CNN is good at extracting local features of network structure data, while LSTM is good at capturing long-term dependencies in sequence data. By combining the advantages of both, the model can take into account both spatial and temporal information of the data, reduce parameters and reduce the risk of overfitting, thereby providing more accurate predictions, better performance and higher training efficiency.

[0102] Feature extraction refers to the process of obtaining raw data and extracting data features. During the extraction process, "noise" features will be deleted to help the model better summarize the data rules and save the most important features, thereby assisting in building a more interpretable model. The CNN algorithm is a type of neural network with a deep structure that includes convolution calculations. It is one of the representative algorithms for deep learning and is particularly suitable for deep learning problems that process grid structure data such as image recognition. The core idea of ​​CNN is to build a network through convolution, pooling, and fully connected layers to extract the features and information of grid data. The convolution layer filters the grid data through the convolution kernel to extract features from the data; the pooling layer reduces the dimension by downsampling to reduce the number of parameters and computational complexity; the fully connected layer maps the input to the output through linear operations for tasks such as classification and detection. By using multiple filters, CNN is able to identify different features of the data and generate feature curves.

[0103] The construction of the river flow evolution model relies on a large amount of hydrological data, and the uncertainty of the hydrological data itself affects the accuracy of the deep learning algorithm model. Using CNN for preprocessing before data training is conducive to capturing the typical characteristics of the data sequence.

[0104] Specifically, the river sections are divided into two types: river sections without reservoirs and river sections with reservoirs. The model structures of the two types of river sections are as follows: Figure 5 The list of model input items is shown in Table 1.

[0105] Table 1

[0106]

[0107] After using the CNN model to extract and integrate the data features of each input item, the results are input into the long short-term memory artificial neural network (hereinafter referred to as "LSTM") for learning and training, and deep mining of big data. LSTM is a special recursive neural network. By introducing a unique gating mechanism, it can ensure the transmission and forgetting of effective information, better capture and utilize long-term dependencies, avoid the gradient vanishing problem, and perform well in processing time series data. It is extremely suitable for hydrological work such as runoff prediction. The basic structure of LSTM includes three gating mechanisms and a memory cell. The four main components are: input gate, forget gate, output gate and memory cell. The forget gate determines which old information is forgotten, the input gate controls the acceptance of new information, and the output gate determines how much information in the cell state will be used to generate the hidden state.

[0108] (4) Model storage

[0109] After the data is input into the CNN-LSTM model training according to the river section type, the model parameter information file of the corresponding river section is generated. Since the model parameter information generated by each training is inconsistent, in order to obtain relatively excellent model parameters for the current river section, the root mean square error (RMSE) and mean relative error (MRE) are selected to evaluate the quality of the training results.

[0110]

[0111] Among them, x i is the measured value, y i is the predicted value, N is the number of measured values; the characteristic of RMSE is that it has a greater penalty for larger error values ​​because it performs a square operation on the difference value. This indicator requires that the flow process has better simulation effects at peak and valley values.

[0112]

[0113] Among them, x i is the measured value, y i is the predicted value, and N is the number of measured values. The characteristic of MRE is that large errors and small errors contribute equally to the overall error. Considering the relative value of the error, it is more robust and requires the flow process to have a better simulation effect on the overall trend.

[0114] The more times you train, the more model parameter samples you get. However, after a certain number of trainings, the frequency of updating the optimal sample drops dramatically. After testing and considering the training efficiency, the number of trainings for each river section is set to 30. After the training, the RMSE and MRE of the 30 samples are calculated. First, select the top 10 RMSE samples from the 30 samples as candidate samples. Then select the sample with the best MRE performance from the candidate samples and store it as the final model parameter information file for the river section. The storage principle of the model parameter information file for a river section is as follows: Figure 6 .

[0115] (5) Simulation prediction

[0116] Each river section is a separate model unit, and the corresponding model parameter file is stored in each section. By simulating the models of each section in series in the order of upstream and downstream, the evolution of river flow can be simulated in a linked manner. Before simulation and prediction, the corresponding data needs to be input into the river section prediction model according to the input item list in Table 1. First, the data of river section 1 is input, and the model parameter file of river section 1 is called through the Python language. The flow process of the lower section of river section 1 is simulated, and this is used as the flow process of the upper section of river section 2, and so on, until the flow process of each section downstream is obtained. The simulation and prediction principle is as follows: Figure 7 .

[0117] In a specific implementation method, a river flow evolution method of a physical mechanism and a deep learning algorithm is applied to the mainstream of the Yellow River for a case demonstration, namely, the Tangnaihai section to the Lijin section into the sea, with a total of 26 sections and 25 river sections, including 5 river sections with reservoirs and 21 river sections without reservoirs. The generalized diagram of the case river node is as follows Figure 8 shown.

[0118] (1) Hydrological and physical model: After inputting the parameters required by the hydrological and physical model, such as the Tangnaihai inflow and the inflow of each tributary, the withdrawal and discharge, and the imbalance, the flow process of each section downstream is deduced.

[0119] (2) Deep learning algorithm model. After the historical data is collected and cleaned, the historical data of each river section is input according to the list shown in Table 1, and the CNN-LSTM model is used for learning and training, and the model parameter information file after optimization is saved; the test period data is input according to the list shown in Table 1, and the model parameter information file is saved according to the list shown in Table 1. Figure 7 The process principle shown in the figure calls the parameter information of each river section and simulates the downstream to the Lijin section where it enters the sea. The mean relative error (MRE) index is used to evaluate the advantages and disadvantages of the two models, as shown in the figure. Fig. 9 shown.

[0120] (3) Co-evolution. Through the comparison of MRE of the two models, it can be seen that among the 25 river sections, the deep learning algorithm model is better than the hydrological and physical model in 14 river sections and worse than the hydrological and physical model in 11 river sections, and the average MRE of the whole river is also better than the latter. Therefore, overall, the accuracy of the deep learning model is better than the traditional hydrological and physical model. In practical applications, the algorithm with better performance is selected and applied to the river section. That is, the 14 river sections with better performance of the deep learning algorithm model are simulated and evolved using this method, and the other 11 river sections are simulated and evolved using the hydrological and physical model. In this way, collaborative optimization and intelligent evolution can be achieved, so that the simulation accuracy of the whole river is optimized. Taking the Longmen section as an example, the intelligent evolution results after collaborative optimization are shown, as shown below: Fig.10 shown.

[0121] In a specific implementation, a river flow evolution system using a physical mechanism and a deep learning algorithm, such as Fig.12 As shown, including:

[0122] Collection and processing module: used to collect hydrological data of river sections, build basic data sets, and perform data preprocessing;

[0123] Model building module: used to extract features from preprocessed data through convolutional neural networks, and use the extracted features to build an evolutionary model based on long short-term memory neural networks;

[0124] Model derivation module: used to derive the hydrological physical model by combining the Muskingum trough storage curve equation and the river section water balance equation;

[0125] Prediction module: used to predict the cross-sectional flow of each river section through hydrological physical model and evolution model respectively;

[0126] Evolution module: used to evaluate the prediction results of the model in each river section through indicators, and to carry out intelligent evolution of collaborative optimization of each river section based on the evaluation results.

[0127] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0128] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A river flow evolution method based on physical mechanism and deep learning algorithm, characterized in that: include: Collect hydrological data of river sections, build basic data sets, and perform data preprocessing; The preprocessed data is subjected to feature extraction through convolutional neural networks, and the extracted features are used to construct an evolution model based on long short-term memory neural networks; The hydrological physical model is derived by combining the Muskingum trough storage curve equation and the river section water balance equation. The cross-sectional flow of each river section is predicted by hydrophysical model and evolution model respectively; The prediction results of the model in each river section are evaluated through indicators, and the collaborative optimization and intelligent evolution of each river section are carried out based on the evaluation results.

2. According to claim 1, a river flow evolution method based on physical mechanism and deep learning algorithm is characterized in that: The hydrological data collection of the river section and the construction of the basic data set include: Obtain natural geography, hydrology, meteorology, and socio-economic data and information through surveys, and transform survey data into various readable and writable basic data through information integration and extraction; Through literature research, on-site investigation, statistical analysis and comprehensive evaluation, the collected data forms are identified, the required information is extracted, the obtained data are checked and unified, and data are updated and maintained regularly to build a basic data set.

3. According to claim 1, a method for river flow evolution using a physical mechanism and a deep learning algorithm is characterized in that: The data preprocessing includes: performing data cleaning, checking data consistency and data logical relationship, processing missing values, invalid values, and abnormal value data through interpolation, envelope constraint and average value replacement methods, and generating data that meets the data training quality requirements.

4. According to claim 1, a method for river flow evolution using a physical mechanism and a deep learning algorithm is characterized in that: The hydrological physical model derived by combining the Muskingum trough storage curve equation and the river section water balance equation includes: Muskingum root groove storage curve equation: S = KQ' = K(xI+(1-x)Q); Among them, x is the weight factor of the inflow and outflow on the tank storage, S is the tank storage capacity, K is the tank storage coefficient, Q′ is the indicated storage flow, I is the upper section flow, and Q is the lower section flow; Water balance equation for a river section: Among them, S1 and S2 are the tank storage at the upper and lower moments respectively, Δt represents the time step, I1 and I2 are the flow rates at the upper and lower moments of the upper section respectively, and Q1 and Q2 are the flow rates at the upper and lower moments of the lower section respectively; By combining the above two equations, the weight factor x is taken as 0.5, the influence of water flow propagation time is considered, the channel storage coefficient K is taken as the propagation time τ, and the interval water inflow, interval water withdrawal, interval water diversion, interval water loss and interval unbalanced water volume parameters are considered to obtain the flow evolution equation based on the hydrological and physical mechanism.

5. According to claim 4, a method for river flow evolution using a physical mechanism and a deep learning algorithm is characterized in that: The flow evolution equation includes: Among them, I jt represents the flow rate at section j at time t, I j-1，t is the flow rate of section j-1 at time t, I j-1，t-1 is the flow rate of section j-1 at time t-1, τ j is the propagation time from section j-1 to section j, Δt is the time step, Q 区jt represents the interval inflow between the j-1th node and the jth node at time t, Q 退jt represents the amount of water withdrawal between the j-1th node and the jth node at time t, Q 引jt represents the water diversion volume between the j-1th node and the jth node at time t, Q 损jt represents the loss between the j-1th node and the jth node at time t, Q 平j,t Represents the unbalanced water volume between the j-1th node and the jth node at time t.

6. According to claim 1, a method for river flow evolution using a physical mechanism and a deep learning algorithm is characterized in that: The evolution model predicts the cross-sectional flow of each river section by dividing each river section into two types, namely, a river section without a reservoir and a river section with a reservoir.

7. The method for river flow evolution based on a physical mechanism and a deep learning algorithm according to claim 1 is characterized in that: Also includes: The root mean square error and mean relative error indicators are selected to evaluate the prediction results of the evolution model; Among them, x i is the measured value, y i is the predicted value, N is the number of measured values; RMSE is the root mean square error; Among them, x i is the measured value, y i is the predicted value, N is the number of measured values; MRE is the mean relative error.

8. The method for river flow evolution based on physical mechanism and deep learning algorithm according to claim 7 is characterized in that: It also includes the selection of evolution model parameters based on evaluation indicators: The number of training times for each river section is set to m times. After the training, the RMSE and MRE evaluation indicators of the m samples are calculated. First, the top n RMSE evaluation indicators among the m samples are selected as candidate samples. Secondly, the samples with the best MRE performance are selected from the candidate samples and stored as the final evolution model parameter information file of the river section.

9. The method for river flow evolution based on a physical mechanism and a deep learning algorithm according to claim 1 is characterized in that: The evolution model predicts the cross-sectional flow of each river section including: Through the evolution model, the river section models are connected in series in the order of upstream and downstream to simulate the evolution process of river flow; Each river section is a separate model unit, and the corresponding model parameter files are stored. The river section models are simulated in series in the order of upstream and downstream. Before simulation and prediction, the corresponding data are input into the evolution model according to the input item list of the river section type. First, the data of the first river section is input, and the model parameter file of the first river section is called through the Python language. The flow process of the lower section of the first river section is simulated and used as the flow process of the upper section of the second river section, and so on, until the flow process of each section downstream is obtained.

10. A river flow evolution system based on physical mechanism and deep learning algorithm, characterized in that: include: Collection and processing module: used to collect hydrological data of river sections, build basic data sets, and perform data preprocessing; Model building module: used to extract features from preprocessed data through convolutional neural networks, and use the extracted features to build an evolutionary model based on long short-term memory neural networks; Model derivation module: used to derive the hydrological physical model by combining the Muskingum trough storage curve equation and the river section water balance equation; Prediction module: used to predict the cross-sectional flow of each river section through hydrological physical model and evolution model respectively; Evolution module: used to evaluate the prediction results of the model in each river section through indicators, and to carry out intelligent evolution of collaborative optimization of each river section based on the evaluation results.

Images