Physical mechanism guided deep learning rainfall runoff forecasting method and related equipment

Abstract

The invention relates to the technical field of rainfall runoff forecasting, in particular to a physical mechanism guided deep learning rainfall runoff forecasting method and related equipment, and the method specifically comprises the steps: obtaining meteorological data, underlying surface data and runoff data, and determining target features through statistical analysis and causal analysis; constructing a water balance long-short term memory neural network; performing model training based on the target features, a long-short term memory neural network of water balance and a Parzen estimator of a structure tree so as to determine hyper-parameters, model weight information and threshold information of a prediction model, and obtaining the prediction model; and forecasting rainfall runoff based on the prediction model. Therefore, on the basis of ensuring the precision performance of the model, the problem that the prediction does not conform to the physical consistency can be avoided, and the practicability of the model is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of precipitation runoff forecasting, and in particular to a precipitation runoff forecasting method and related equipment guided by deep learning using a physical mechanism. Background Art

[0002] Deep learning has shown its advantages in the application research of hydrological simulation, such as flexible modeling, outstanding prediction performance, and universal applicability. For example, precipitation runoff can be predicted through deep learning models. However, in the existing technology, when the predicted data exceeds the range of existing data, there are problems such as reduced forecasting performance and model inconsistency with physical consistency. In addition, the forecasting accuracy of deep learning models for precipitation runoff forecasting still needs to be improved. Summary of the invention

[0003] In view of this, the purpose of the present invention is to provide a precipitation runoff forecasting method and related equipment guided by deep learning of physical mechanisms, so as to overcome the current problems existing in the prediction of precipitation runoff through deep learning models, such as reduced forecasting performance and model inconsistency with physical consistency when the predicted data exceeds the range of existing data, and the problem that the forecasting accuracy still needs to be improved.

[0004] To achieve the above objectives, the present invention adopts the following technical solutions:

[0005] In a first aspect, the present application provides a precipitation runoff forecasting method guided by deep learning using a physical mechanism, comprising:

[0006] Obtain meteorological data, underlying surface data and runoff data, and determine target characteristics through statistical analysis and cause-and-effect analysis;

[0007] Constructing a long short-term memory neural network for water balance;

[0008] Model training is performed based on the target characteristics, the long short-term memory neural network of water balance and the Parzen estimator of the structure tree to determine the hyperparameters, model weight information and threshold information of the prediction model to obtain the prediction model;

[0009] Based on the prediction model, precipitation runoff forecast is carried out.

[0010] Furthermore, in some embodiments of the present application, the acquisition of meteorological data, underlying surface data and runoff data, and determination of target characteristics through statistical analysis and causal analysis, include:

[0011] Determine meteorological characteristics and underlying surface characteristics;

[0012] Based on correlation analysis, principal component analysis and stepwise regression analysis, the meteorological characteristics and underlying surface characteristics are statistically analyzed to screen out the characteristics that are strongly correlated with the runoff;

[0013] A causal relationship model is constructed to conduct causal analysis on the meteorological characteristics to screen out the characteristics that have significant causal effects on runoff.

[0014] Furthermore, in some embodiments of the present application, the long short-term memory neural network for constructing water balance includes:

[0015] A water conservation gate is introduced into the long short-term memory neural network, and the memory units and loss functions in the long short-term memory neural network are adjusted so that the input of water volume is equal to the sum of the system state change and the output.

[0016] Furthermore, in some embodiments of the present application, the model training is performed based on the target characteristics, the long short-term memory neural network of the water balance and the Parzen estimator of the structure tree to determine the hyperparameters, model weight information and threshold information of the prediction model to obtain the prediction model, including: determining the hyperparameters of the prediction model based on the Parzen estimator of the structure tree.

[0017] Furthermore, in some embodiments of the present application, the Parzen estimator based on the structure tree determines the hyperparameters of the prediction model, including:

[0018] Determine the objective function based on the performance indicators of the prediction model;

[0019] Use the Parzen estimator of the structure tree to build a probability model;

[0020] Training the prediction model based on the target feature, and iteratively updating the probability model during the training process;

[0021] The optimal hyperparameters of the prediction model are determined based on the objective function.

[0022] Furthermore, in some embodiments of the present application, the prediction model is trained based on the target feature, and during the training process, the probability model is iteratively updated, including:

[0023] Generate new hyperparameter combinations based on the current training data of the prediction model and evaluate the new hyperparameter combinations.

[0024] Furthermore, in some embodiments of the present application, it also includes:

[0025] Construct a comprehensive indicator system of accuracy and physical consistency;

[0026] The prediction model is evaluated by the comprehensive indicator system.

[0027] Furthermore, in some embodiments of the present application, the performance indicators in the comprehensive indicator system include Nash efficiency coefficient, root mean square error, mean absolute error, 2% highest flow in the cumulative frequency curve of runoff, and 30% lowest flow in the cumulative frequency curve of runoff.

[0028] In a second aspect, the present application provides a precipitation runoff forecasting device guided by deep learning using a physical mechanism, comprising:

[0029] The screening module is used to obtain meteorological data, underlying surface data and runoff data, and determine the target characteristics through statistical analysis and cause-effect analysis;

[0030] Building blocks for constructing long-short-term memory neural networks for water balance;

[0031] A training module, used for performing model training based on the target characteristics, the long short-term memory neural network of water balance and the Parzen estimator of the structure tree, so as to determine the hyperparameters, model weight information and threshold information of the prediction model, and obtain the prediction model;

[0032] The prediction module is used to make precipitation runoff forecast based on the prediction model.

[0033] In a third aspect, the present application provides a precipitation runoff forecasting device guided by deep learning through a physical mechanism, including a processor and a memory, wherein the processor is connected to the memory:

[0034] Wherein, the processor is used to call and execute the program stored in the memory;

[0035] The memory is used to store the program, and the program is at least used to execute the above-mentioned physical mechanism guided deep learning precipitation runoff forecasting method.

[0036] The present invention relates to the technical field of precipitation runoff forecasting, and specifically to a precipitation runoff forecasting method and related equipment guided by physical mechanism deep learning, the method specifically comprising: obtaining meteorological data, underlying surface data and runoff data, determining target features through statistical analysis and causal analysis; constructing a long short-term memory neural network of water balance; performing model training based on target features, long short-term memory neural network of water balance and Parzen estimator of structure tree to determine hyperparameters, model weight information and threshold information of the prediction model, and obtain the prediction model; performing precipitation runoff forecasting based on the prediction model. In this way, on the basis of ensuring the accuracy performance of the model, the problem of prediction not being consistent with physical consistency can also be avoided, thereby improving the practicality of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] 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 some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0038] Figure 1 It is a flow chart of the precipitation runoff forecasting method guided by deep learning based on physical mechanism provided in an embodiment of the present invention.

[0039] Figure 2 It is a flow chart of a precipitation runoff forecasting method based on physical mechanism-guided deep learning provided by another embodiment of the present invention.

[0040] Figure 3 This is a structural diagram of the LSTM model.

[0041] Figure 4 It is a structural diagram of the MC-LSTM model.

[0042] Figure 5 It is a schematic diagram of the change of loss function during the training period and the testing period in the application of the precipitation runoff forecasting method guided by deep learning based on the physical mechanism provided by an embodiment of the present invention.

[0043] Figure 6 It is a schematic diagram of the runoff process in the test period of the Liyuan Basin in the application of the precipitation runoff forecasting method guided by deep learning based on the physical mechanism provided in an embodiment of the present invention.

[0044] Figure 7 It is a schematic diagram of the runoff process in the Wunonglong Basin during the test period in the application of the precipitation runoff forecasting method guided by deep learning based on the physical mechanism provided in an embodiment of the present invention.

[0045] Figure 8 It is a structural schematic diagram of a precipitation runoff forecasting device guided by deep learning of a physical mechanism provided in an embodiment of the present invention.

[0046] Fig. 9 It is a structural schematic diagram of a precipitation runoff forecasting device guided by deep learning of a physical mechanism provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0047] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be described in detail below. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other implementation methods obtained by ordinary technicians in this field without making creative work belong to the scope of protection of the present invention.

[0048] Figure 1 This is a flow chart of the precipitation runoff forecasting method guided by deep learning based on physical mechanism provided by an embodiment of the present invention. Figure 1 , this embodiment may include the following steps:

[0049] S101. Obtain meteorological data, underlying surface data and runoff data, and determine target characteristics through statistical analysis and cause-and-effect analysis.

[0050] Specifically, meteorological characteristics such as rainfall and temperature are first determined based on meteorological data, and digital elevation is determined based on underlying surface data to determine characteristics such as slope, and NDVI is determined to determine characteristics such as vegetation. Then, statistical analysis and causal analysis are performed on the above-determined characteristics to determine the final target characteristics, which are used to train the prediction model mentioned later.

[0051] S102. Construct a long short-term memory neural network for water balance.

[0052] Specifically, the concept of water balance is introduced into the long short-term memory neural network to control the training and prediction of the prediction model, and the long short-term memory neural network of water balance is obtained.

[0053] S103, performing model training based on target features, long short-term memory neural network of water balance and Parzen estimator of structure tree to determine hyperparameters, model weight information and threshold information of the prediction model to obtain the prediction model.

[0054] Specifically, a data set is constructed through target features to train the long short-term memory neural network of water balance, and the hyperparameters of the prediction model are determined through the Parzen estimator of the structure tree, and finally the model weight information and threshold information are obtained to obtain the prediction model.

[0055] S104. Precipitation runoff forecast is performed based on the prediction model.

[0056] Specifically, the data to be predicted is input into the trained prediction model to perform precipitation runoff forecasting. It should be noted that the prediction model mentioned in this application can be constructed and trained by the above method using the data of the basin where precipitation runoff forecasting is actually performed.

[0057] The physical mechanism-guided deep learning precipitation runoff forecasting method provided by the present invention obtains meteorological data, underlying surface data and runoff data, and determines the target characteristics through statistical analysis and causal analysis; constructs a long short-term memory neural network of water balance; performs model training based on the target characteristics, the long short-term memory neural network of water balance and the Parzen estimator of the structure tree to determine the hyperparameters, model weight information and threshold information of the prediction model, and obtains the prediction model; performs precipitation runoff forecasting based on the prediction model. In this way, on the basis of ensuring the accuracy performance of the model, the problem of the prediction not being in line with physical consistency can also be avoided, thereby improving the practicality of the model.

[0058] Figure 2 is a flow chart of a precipitation runoff forecasting method guided by deep learning based on a physical mechanism provided by another embodiment of the present invention, with reference to Figure 2 In some embodiments of the present application, meteorological data, underlying surface data and runoff data are obtained, and target characteristics are determined through statistical analysis and causal analysis, specifically including: first determining meteorological characteristics and underlying surface characteristics; then performing statistical analysis on the meteorological characteristics and underlying surface characteristics based on correlation analysis, principal component analysis and stepwise regression analysis to screen out characteristics that have a strong correlation with runoff; and constructing a causal relationship model to perform causal analysis on the meteorological characteristics to screen out characteristics that have a significant causal effect on the runoff.

[0059] Specifically, in the present application, in terms of statistical analysis, correlation analysis, principal component analysis (PCA) and stepwise regression analysis can be used to screen features that are strongly correlated with runoff. Among them, correlation analysis identifies features that are significantly correlated with runoff by calculating the Pearson correlation coefficient and the Spearman rank correlation coefficient, thereby helping to understand the linear or nonlinear relationship therein. PCA is used to reduce the dimension of the data, mapping high-dimensional features to low-dimensional space while retaining the main information, which helps to extract the most explanatory features for runoff forecasts. Stepwise regression analysis gradually introduces or eliminates features to select the features that contribute most to model predictions, thereby avoiding overfitting and optimizing the predictive performance of the prediction model (the features obtained in the above analysis can all be considered to be features that are strongly correlated with runoff).

[0060] In terms of causal analysis, data analysis and domain knowledge can be combined to construct a causal relationship model to identify the causal impact of meteorological characteristics on runoff. Then, the causal relationship between features can be verified through the Granger causality test to ensure that the selected features have a significant causal effect when predicting runoff (when the test fails, it means that the constructed causal relationship model needs to be adjusted and modified). It can be understood that causal analysis can not only help screen out meteorological features that have a significant impact on runoff, but also enhance the explanatory power and prediction accuracy of the prediction model. Through causal analysis, key features, namely target features, can be more reliably identified, providing solid data support for further rainfall runoff simulation prediction.

[0061] Furthermore, in an embodiment of the present application, constructing a long short-term memory neural network for water balance may specifically include: introducing a water conservation gate in the long short-term memory neural network, and adjusting the memory units and loss functions in the long short-term memory neural network so that the input of water volume is equal to the sum of the system state change and the output.

[0062] Specifically, we first briefly introduce the Long Short-Term Memory (LSTM) neural network. Its model structure is as follows: Figure 3 As shown. Long short-term memory neural network (LSTM) is a specially designed recurrent neural network (RNN) to solve the gradient vanishing and gradient exploding problems encountered by standard RNN when processing long sequence data (in this application, meteorological data and runoff data are both time series data). The core of the LSTM network lies in its unique memory unit and gating mechanism, which makes it superior in capturing long-term dependencies in time series data. The basic structure of the LSTM network includes a memory cell, an input gate, a forget gate, and an output gate. These main components work together to help the LSTM network effectively learn and store long-term dependency information.

[0063] Among them, the memory cell is the core of LSTM, which is used to save long-term dependent information. It achieves this by maintaining a state in each time step of the network, which can be updated, read or ignored, thereby effectively controlling the transmission and retention of information. The input gate is responsible for determining which parts of the current input information will be stored in the memory cell. It consists of a sigmoid activation function and a tanh activation function. The sigmoid function controls the "write amount" of information, and the tanh function generates a candidate value vector, which will be added to the state of the memory cell. The forget gate controls whether the information in the memory cell should be forgotten. Through a sigmoid activation function, the forget gate determines how much historical information to delete from the memory cell. This mechanism helps to remove information that is no longer useful, thereby reducing the problem of gradient disappearance. The output gate determines whether the content of the memory cell should be output to the next time step. It generates a gated value through a sigmoid activation function, and combines the memory cell state of the tanh activation function to generate the final output, so that the network can selectively pass the information in the memory to the next time step.

[0064] This application uses an improved long short-term memory network, namely Mass Conserving LSTM (MC-LSTM), to perform model training and then obtain a prediction model.

[0065] Specifically, MC-LSTM is an improved long short-term memory network, which is specially designed to handle tasks that require mass conservation. The core goal of this network model is to ensure that the mass (and flow or concentration, etc.) of the system remains unchanged during the processing of time series data. MC-LSTM combines the powerful sequence modeling capabilities of LSTM with the mass conservation requirements in physical processes. By introducing additional mass conservation mechanisms based on standard LSTM, it ensures that there will be no unreasonable increase or decrease of substances during the simulation process. In this application, the following operations are performed:

[0066] (1) Introducing the water conservation gate:

[0067] A water conservation gate is introduced in MC-LSTM to adjust the update process of the memory unit. This gate is used to control the inflow and outflow of information in the memory unit to ensure that when processing time series data, the state change of the memory unit meets the requirements of mass conservation. In a physical sense, it means keeping the input (i.e., precipitation) equal to the sum of the change in the system state and the output.

[0068] (2) Adjust the memory unit:

[0069] At each time step, the memory cells are adjusted accordingly to ensure that the updated state maintains water conservation, which means that during the forward propagation of the model, the state of the memory cells is modified to satisfy the water conservation constraint.

[0070] (3) Modify the loss function:

[0071] In order to achieve water conservation, a water conservation term can also be introduced into the loss function. This additional term is used to penalize predictions that violate the conservation principle, thereby guiding the prediction model to maintain water conservation during training.

[0072] It should be noted that this part of the content can be understood by referring to the improvements of MC-LSTM and LSTM in the prior art, and will not be described in more detail here.

[0073] Furthermore, in the embodiment of the present application, it also includes determining the hyperparameters of the prediction model based on the Parzen estimator of the structure tree, including determining the objective function based on the performance index of the prediction model; using the Parzen estimator of the structure tree to build a probability model; training the prediction model based on the target feature, and iteratively updating the probability model during the training process; determining the optimal hyperparameters of the prediction model based on the objective function. Among them, training the prediction model based on the target feature, and iteratively updating the probability model during the training process, includes: generating a new hyperparameter combination based on the current training data of the prediction model, and evaluating the new hyperparameter combination.

[0074] Specifically, the hyperparameters that need to be determined in the MC-LSTM model used in this application include common hyperparameters in ordinary LSTM, such as the number of hidden units, learning rate, and batch size, as well as some parameters specific to MC-LSTM, such as water conservation gate settings and conservation constraint strength. The reasonable selection and adjustment of these hyperparameters have an important impact on the training effect, computational efficiency, and accuracy of physical process simulation of the prediction model. In this application, the tree-structured Parzen Estimator (TPE) method is used to select model hyperparameters. TPE is a powerful Bayesian optimization method for efficiently searching for the optimal hyperparameter combination. In this application, the method of finding the best hyperparameters in TPE is based on probability modeling, which specifically includes the following steps:

[0075] (1) Define the objective function and select the performance indicators of the prediction model on the validation set, such as accuracy and physical consistency (various performance indicators involved in the subsequent evaluation of the prediction model) to generate the objective function.

[0076] (2) Use the tree-structured Parzen estimator to build a probability model and estimate the distribution of hyperparameters based on historical experimental data (data generated during the training of the prediction model). Use TPE to divide the hyperparameter space into "good" and "bad" areas, and adjust and optimize the hyperparameter search strategy through the probability density estimation of these two areas.

[0077] (3) Optimize the probability model. The construction and update of the probability model are carried out dynamically. Specifically, TPE uses the previous experimental results (i.e., the hyperparameter combinations and objective function values ​​involved in the prediction model training process) to update the probability model and generate new hyperparameter candidate combinations. In practical applications, TPE can use the hyperparameter distribution model and the overall hyperparameter distribution model to find those hyperparameter combinations with higher probabilities in the areas with better objective function performance. This can effectively focus on possible high-quality areas and gradually approach the optimal hyperparameter combination.

[0078] (4) Iterative optimization, using TPE to continuously adjust the distribution model of hyperparameters through an iterative process. Specifically, in each iteration, TPE is used to update new hyperparameter candidates based on the current distribution model, and then experiments and evaluations are conducted (i.e., training and evaluation of the prediction model), and the distribution model and optimization strategy are further adjusted based on the results. This iterative optimization process ensures a gradual improvement in the hyperparameter selection and ultimately finds the best hyperparameter combination.

[0079] It should be noted that the above steps (3) and (4) also include calculating the boosting strategy function value, and performing multiple iterations here according to the preset iterations, so as to determine the optimal hyperparameter combination and obtain the hyperparameters of the prediction model.

[0080] Furthermore, in the embodiment of the present application, it also includes: constructing a comprehensive index system of accuracy and physical consistency; and evaluating the prediction model through the comprehensive index system. Among them, the performance indicators in the comprehensive index system include Nash efficiency coefficient, root mean square error, mean absolute error, 2% of the highest flow in the cumulative frequency curve of runoff, and 30% of the lowest flow in the cumulative frequency curve of runoff.

[0081] Specifically, in order to evaluate the performance of the prediction model and compare the adaptability of the prediction model in hydrological forecasting in different river basins, this application evaluates the simulation results of the prediction model through multiple indicators such as Nash efficiency coefficient (Ens), root mean square error (RMSE), and mean absolute error (MAE). Among them, Ens reflects the overall effect of the model simulation; the two metrics RMSE and MAE focus on evaluating the overall error of the series. RMSE is used to measure the deviation between the observed value and the true value, and MAE expresses the mean degree of absolute error. Both are sensitive to the maximum or minimum errors in the sequence. FHV and FLV are both derived from the flow duration curve, which is the cumulative frequency curve of the flow. The H of FHV and FLV correspond to 2% of the highest flow and 30% of the lowest flow, respectively. The specific formulas of the above indicators are as follows:

[0082]

[0083] Where n is the number of forecast periods; i is the period number; Q f is the predicted runoff value; Q 0 is the measured runoff value; H is the total number of time periods corresponding to the highest 2% flow; h is the number of a runoff value among the H flow values; Indicates the highest 2% traffic prediction value; y h It represents the measured value of the highest 2% flow; H represents the total number of time periods corresponding to the lowest 30% flow; l represents the number of a certain runoff value among the H flow values; Indicates the lowest 30% flow observation value; y l Indicates the lowest 30% flow measured value; y L Indicates the maximum value of the lowest 30% flow rate measured value; Indicates the maximum value of the lowest 30% flow prediction values.

[0084] The precipitation runoff forecasting method guided by deep learning based on physical mechanisms provided in this application is tested through two specific scenarios as follows:

[0085] The average precipitation, temperature, snow thickness, and runoff data of 5051 days in the Liyuan Basin from 2010 to 2023 and 1764 days in the Wunonglong Basin from 2019 to 2023 were arranged in time series to form a data set (the underlying surface can be fixed), and divided into training and test sets in a ratio of 7:3. The TPE method was used to optimize the model hyperparameters and verify the changes in the loss function during the training and test periods (such as Figure 5 As shown in the figure, overall, the loss function gradually stabilizes with the increase of training samples, and the loss function in the training period is slightly lower than that in the test period (validation period), indicating that the model is not overfitting.

[0086] In addition, if Figure 6 As shown in the figure (which is the runoff process of the test period in the Liyuan Basin), it can be seen that the physical mechanism-guided deep learning precipitation runoff forecasting method and the corresponding generated prediction model provided in this application have good simulation results of daily runoff in the Liyuan Basin, with a Nash efficiency coefficient of 0.73, MAE of 90.86, FHV and FLV of 13.6 and -14.8 respectively. The overall simulation effect is good and can better restore the flow fluctuation process of the Liyuan Basin. And as Figure 7 As shown (which is the runoff process in the test period of the Wunonglong Basin), the physical mechanism-guided deep learning precipitation runoff forecasting method and the corresponding generated prediction model provided in this application have good daily-scale runoff simulation results in the Wunonglong Basin, with a Nash efficiency coefficient of 0.82, MAE of 76.28, FHV and FLV of 11.6 and -9.8 respectively. The overall simulation effect is excellent and can effectively simulate the flow fluctuations and some flood peak processes in the Wunonglong Basin.

[0087] The physical mechanism-guided deep learning precipitation runoff forecasting method provided in this application can effectively maintain the accuracy of the model, while also enhancing the consistency of the prediction model at the physical level, where the prediction range of the prediction model is reasonably constrained in a space that conforms to physical laws, ensuring the credibility of the results. The reasonable selection and adjustment of hyperparameters further improve the performance of the prediction model, including training effects, computational efficiency, and simulation accuracy of physical processes, which improves the computational efficiency and interpretability of the prediction model in practical applications, and contributes to the generalization ability and wide promotion of the prediction model.

[0088] Based on the same inventive concept, the present application also provides a precipitation runoff forecasting device guided by deep learning using a physical mechanism, such as Figure 8 As shown, the device comprises:

[0089] The screening module 11 is used to obtain meteorological data, underlying surface data and runoff data, and determine target characteristics through statistical analysis and cause-effect analysis.

[0090] The construction module 12 is used to construct a long short-term memory neural network for water balance.

[0091] The training module 13 is used to perform model training based on the target characteristics, the long short-term memory neural network of water balance and the Parzen estimator of the structure tree to determine the hyperparameters, model weight information and threshold information of the prediction model to obtain the prediction model.

[0092] The prediction module 14 is used to make precipitation runoff forecast based on the prediction model.

[0093] Regarding the device in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the method, and will not be elaborated here.

[0094] The present invention also provides a precipitation runoff forecasting device guided by deep learning of physical mechanism, which is used to implement the above method embodiment, such as Fig. 9 As shown, the precipitation runoff forecasting device guided by deep learning of physical mechanism in this embodiment includes a processor 21 and a memory 22, and the processor 21 is connected to the memory 22. The processor 21 is used to call and execute the program stored in the memory 22; the memory 22 is used to store the program, which is at least used to execute the precipitation runoff forecasting method guided by deep learning of physical mechanism in the above embodiment.

[0095] The specific implementation scheme of the precipitation and runoff forecasting device guided by deep learning based on the physical mechanism provided in the embodiment of the present application can refer to the implementation scheme of the precipitation and runoff forecasting method guided by deep learning based on the physical mechanism of any of the above embodiments, which will not be repeated here.

[0096] It can be understood that the same or similar parts of the above embodiments can be referenced to each other, and the contents not described in detail in some embodiments can refer to the same or similar contents in other embodiments.

[0097] It should be noted that, in the description of the present invention, the terms "first", "second", etc. are only used for descriptive purposes and cannot be understood as indicating or implying relative importance. In addition, in the description of the present invention, unless otherwise specified, the meaning of "plurality" refers to at least two.

[0098] Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, segment or portion of code that includes one or more executable instructions for implementing the steps of a specific logical function or process, and the scope of the preferred embodiments of the present invention includes alternative implementations in which functions may not be performed in the order shown or discussed, including performing functions in a substantially simultaneous manner or in the reverse order depending on the functions involved, which should be understood by those skilled in the art to which the embodiments of the present invention belong.

[0099] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, a plurality of steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0100] A person skilled in the art may understand that all or part of the steps in the method for implementing the above-mentioned embodiment may be completed by instructing related hardware through a program, and the program may be stored in a computer-readable storage medium, which, when executed, includes one or a combination of the steps of the method embodiment.

[0101] In addition, each functional unit in each embodiment of the present invention may be integrated into a processing module, or each unit may exist physically separately, or two or more units may be integrated into one module. The above-mentioned integrated module may be implemented in the form of hardware or in the form of a software functional module. If the integrated module is implemented in the form of a software functional module and sold or used as an independent product, it may also be stored in a computer-readable storage medium.

[0102] The storage medium mentioned above can be a read-only memory, a magnetic disk or an optical disk, etc.

[0103] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "examples", "specific examples", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner.

[0104] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations of the present invention. A person skilled in the art may change, modify, replace and vary the above embodiments within the scope of the present invention.

Claims

1. A precipitation runoff forecasting method guided by deep learning based on physical mechanism, characterized in that: include: Obtain meteorological data, underlying surface data and runoff data, and determine target characteristics through statistical analysis and cause-and-effect analysis; Constructing a long short-term memory neural network for water balance; Model training is performed based on the target characteristics, the long short-term memory neural network of water balance and the Parzen estimator of the structure tree to determine the hyperparameters, model weight information and threshold information of the prediction model to obtain the prediction model; Based on the prediction model, precipitation runoff forecast is carried out.

2. The precipitation runoff forecasting method based on physical mechanism guided deep learning according to claim 1 is characterized in that: The acquisition of meteorological data, underlying surface data and runoff data, and determination of target characteristics through statistical analysis and causal analysis, include: Determine meteorological characteristics and underlying surface characteristics; Based on correlation analysis, principal component analysis and stepwise regression analysis, the meteorological characteristics and underlying surface characteristics are statistically analyzed to screen out the characteristics that are strongly correlated with the runoff; A causal relationship model is constructed to conduct causal analysis on the meteorological characteristics to screen out the characteristics that have significant causal effects on runoff.

3. The precipitation runoff forecasting method based on physical mechanism guided deep learning according to claim 1 is characterized in that: The long short-term memory neural network for constructing water balance includes: A water conservation gate is introduced into the long short-term memory neural network, and the memory units and loss functions in the long short-term memory neural network are adjusted so that the input of water volume is equal to the sum of the system state change and the output.

4. The precipitation runoff forecasting method based on physical mechanism guided deep learning according to claim 1 is characterized in that: The model training is performed based on the target characteristics, the long short-term memory neural network of water balance and the Parzen estimator of the structure tree to determine the hyperparameters, model weight information and threshold information of the prediction model to obtain the prediction model, including: determining the hyperparameters of the prediction model based on the Parzen estimator of the structure tree.

5. The precipitation runoff forecasting method based on physical mechanism guided deep learning according to claim 4 is characterized in that: The Parzen estimator based on the structure tree determines the hyper parameters of the prediction model, including: Determine the objective function based on the performance indicators of the prediction model; Use the Parzen estimator of the structure tree to build a probability model; Training the prediction model based on the target feature, and iteratively updating the probability model during the training process; The optimal hyperparameters of the prediction model are determined based on the objective function.

6. The precipitation runoff forecasting method based on physical mechanism guided deep learning according to claim 5 is characterized in that: The prediction model is trained based on the target feature, and the probability model is iteratively updated during the training process, including: Generate new hyperparameter combinations based on the current training data of the prediction model and evaluate the new hyperparameter combinations.

7. The precipitation runoff forecasting method based on physical mechanism guided deep learning according to claim 1 is characterized in that: Also includes: Construct a comprehensive indicator system of accuracy and physical consistency; The prediction model is evaluated by the comprehensive indicator system.

8. The precipitation runoff forecasting method based on physical mechanism guided deep learning according to claim 7 is characterized in that: The performance indicators in the comprehensive indicator system include Nash efficiency coefficient, root mean square error, mean absolute error, 2% highest flow in the runoff cumulative frequency curve and 30% lowest flow in the runoff cumulative frequency curve.

9. A precipitation runoff forecasting device guided by deep learning based on physical mechanism, characterized in that: include: The screening module is used to obtain meteorological data, underlying surface data and runoff data, and determine the target characteristics through statistical analysis and cause-effect analysis; Building blocks for constructing long-short-term memory neural networks for water balance; A training module, used for performing model training based on the target characteristics, the long short-term memory neural network of water balance and the Parzen estimator of the structure tree, so as to determine the hyperparameters, model weight information and threshold information of the prediction model, and obtain the prediction model; The prediction module is used to make precipitation runoff forecast based on the prediction model.

10. A precipitation runoff forecasting device guided by deep learning based on physical mechanism, characterized in that: The invention comprises a processor and a memory, wherein the processor is connected to the memory: Wherein, the processor is used to call and execute the program stored in the memory; The memory is used to store the program, and the program is at least used to execute the precipitation runoff forecasting method guided by deep learning based on physical mechanism as described in any one of claims 1-8.

Images