A Deep Learning Multi-Model Runoff Prediction Method and System Embedding Runoff Generation Mechanism

Abstract

This invention discloses a deep learning multi-model runoff prediction method and system embedding runoff generation mechanisms, belonging to the field of hydraulic engineering applications. The method includes: acquiring data, including meteorological and hydrological data; inputting the meteorological data into three neural networks respectively embedded with the GR4J runoff generation mechanism, the ordinary linear runoff generation mechanism, and the exponential bucket runoff generation mechanism to obtain three preliminary runoff generation processes; inputting the three preliminary runoff generation processes into a long short-term memory (LSTM) neural network model to obtain three preliminary outflow processes; the LSTM neural network model incorporates peak size constraints, peak occurrence time constraints, and monotonicity constraints; inputting the three preliminary outflow processes into a Bayesian averaging model to obtain the predicted outflow process, and combining this with the measured outflow process from the hydrological data to calculate the final outflow process. This invention not only minimizes the uncertainty of model predictions but also ensures the accuracy of runoff prediction.

Description

Technical Field

[0001] This invention belongs to the field of hydraulic engineering applications, specifically relating to a deep learning multi-model runoff prediction method and system that embeds runoff generation mechanisms. Background Technology

[0002] Efficient and reliable hydrological models can accurately obtain runoff information, providing crucial technical support for water resource management and flood control and mitigation efforts. Process-based hydrological models are equipped with explicit hydrological knowledge, such as mass and energy conservation and runoff generation at full capacity. Currently, the most widely used physical models are the Exponential Bucket Hydrologic Model (EXP-Hydro), the Simple Linear Model (SLM), and GR4J (modèle du Génie Rural à 4 paramètres Journalier), which achieve good accuracy in predicting runoff processes. However, when physical models are applied to catchment areas where hydrological conditions do not align with model assumptions, they may yield unsatisfactory results.

[0003] With the rapid development of computer technology, machine learning (ML) models, especially deep learning (DL) models, have become key tools for studying watershed responses to meteorological changes and multi-basin runoff forecasting due to their excellent data processing capabilities and computational efficiency. For example, random forests, support vector regression, artificial neural networks, and recurrent neural networks (RNNs) are commonly used machine learning methods for runoff forecasting. Long Short-Term Memory (LSTM) networks are the most advanced method in runoff forecasting, possessing specialized memory cells that can selectively retain information for extended periods. However, many scholars in the hydrological community remain cautious when adopting machine learning methods. This is because machine learning models are entirely data-driven and inherently lack representation of internal states and physical processes. When ML models are applied to watersheds where hydrological processes are affected by external disturbances (such as climate change or human activities), they may produce unreliable predictions and significantly negatively impact prediction accuracy.

[0004] Considering that physical models can provide physical interpretability, while neural network models possess powerful data mining capabilities, the synergy between the two methods has become an attractive research topic in recent years. Physically interpretable neural network models, i.e., hybrid models, are increasingly attracting scholarly attention. Hybrid models can be divided into loosely coupled and tightly coupled types. Loosely coupled models include: 1) using neural network models as post-processors to correct errors in process-based models; 2) using neural network models to replace some hydrological processes in physical models; 3) using the output data of process-based models as synthetic data to train neural network models; and 4) incorporating physical constraints as penalty terms into the loss function of deep learning (DL) models. These computationally efficient methods are relatively easy to implement, but they all require representing physical knowledge in a closed-form form within the ML model. However, since the Earth system is a highly dynamic system, its intermediate processes are difficult to obtain in a closed-form representation. In contrast, tightly coupled models, by adding one or more physical layers, embed hydrological models with numerous parameters into the framework, transforming a neural network that better reflects geophysical processes. However, this method is more complex than loosely coupled methods and places high demands on the programming skills of hydrologists.

[0005] In addition, runoff prediction using neural network models faces various sources of uncertainty, such as parameter uncertainty, model structure uncertainty, and input data uncertainty. These uncertainties directly affect the model's output. Hydrological researchers have proposed multi-model ensemble forecasting methods to quantify and reduce the uncertainty in hydrological forecasts. By applying a set of different models for hydrological forecasting and then integrating the forecast results using post-processing methods, different characteristics of the hydrological sequence can be captured, and structural uncertainties can be considered, thereby effectively improving forecast accuracy. Bayesian Model Averaging (BMA) is a commonly used multi-model ensemble forecasting method that can be applied to runoff prediction to reduce the uncertainty generated during the forecasting process.

[0006] Therefore, it is necessary to design a deep learning multi-model runoff prediction method and system that embeds runoff generation mechanisms to address the above problems. Summary of the Invention

[0007] The purpose of this invention is to address the various sources of uncertainty and accuracy issues in runoff prediction by providing a deep learning multi-model runoff prediction method that embeds runoff generation mechanisms. This method embeds three different runoff generation mechanisms into a neural network, replaces the confluence process with an LSTM neural network, and uses a Bayesian averaging model combined with measured outflow processes from hydrological data to calculate the weights of the three models, thus obtaining the final outflow process. This ensures the integrity and interpretability of the overall process, making the runoff predicted by the deep learning model more consistent with hydrological principles, and improving prediction accuracy, computational efficiency, and user-friendliness.

[0008] According to one aspect of this specification, a deep learning multi-model runoff prediction method embedding runoff generation mechanisms is provided, comprising:

[0009] S1. Acquire data, including meteorological data and hydrological data;

[0010] S2. Meteorological data is input into three neural networks that are respectively embedded with the GR4J runoff generation mechanism, the ordinary linear runoff generation mechanism, and the exponential barrel runoff generation mechanism to obtain three preliminary runoff generation processes;

[0011] S3. Input the three preliminary flow generation processes into the long short-term memory neural network model to obtain three preliminary flow outflow processes; the long short-term memory neural network model introduces peak size constraints, peak occurrence time constraints, and monotonicity constraints;

[0012] S4. Input the three preliminary outflow processes into the Bayesian average model to obtain the predicted outflow process. Combine the measured outflow process in the hydrological data to calculate the final outflow process.

[0013] Furthermore, the process of embedding the flow generation mechanism into the neural network in S2 includes:

[0014] Using physical formulas Replace the nonlinear activation function in the neural network model ,parameter Replace some weight coefficients in the neural network model Soil moisture content Replace state variables We obtained a neural network with an embedded flow generation mechanism.

[0015] Furthermore, the neural network embedding the flow generation mechanism is expressed as follows:

[0016]

[0017]

[0018] in, These are rainfall, sunshine duration, and temperature, respectively. Soil moisture content, The updated formula for soil moisture content, For the update formulas of each flow generation mechanism, These are the physical parameters in each flow generation mechanism.

[0019] Furthermore, S3 also includes:

[0020] By adding peak size constraints, peak occurrence time constraints, and monotonicity constraints to the original loss function of the Long Short-Term Memory (LSTM) neural network model, the new loss function expression is as follows:

[0021]

[0022] in, The original loss function, For peak size constraints, For peak time constraints, As a monotonicity constraint, It is a custom weight.

[0023] Furthermore, S3 also includes:

[0024] The peak size constraint is calculated by comparing the relative error between the annual simulated peak and the measured peak, and then added to the loss function with corresponding weights.

[0025] The peak time constraint is calculated by comparing the relative error between the annual simulated peak time and the measured peak time, and then added to the loss function with corresponding weights.

[0026] The monotonicity constraint is based on the monotonic relationship between production and merging flows, and the loss is calculated using a response function.

[0027] Furthermore, S4 also includes:

[0028] The posterior distribution is obtained based on the predicted outflow process and the measured outflow process in the hydrological data. The mean and variance of the posterior distribution are iteratively calculated using the expectation-maximization algorithm to obtain the weights of the three neural networks.

[0029] Based on the weights of the three neural networks and the three preliminary outflow processes, the final outflow process is obtained.

[0030] Furthermore, the final outflow process can be expressed as:

[0031]

[0032] in, This indicates the final outflow process. This represents the weights of the three neural networks. This represents the initial outflow process of the three neural networks. .

[0033] According to one aspect of this specification, a deep learning multi-model runoff prediction system with embedded runoff generation mechanisms is provided, comprising:

[0034] The data acquisition module is used to acquire data, including meteorological data and hydrological data;

[0035] The preliminary runoff process module is used to input meteorological data into three neural networks that are respectively embedded with the GR4J runoff mechanism, the ordinary linear runoff mechanism, and the exponential bucket runoff mechanism to obtain three preliminary runoff processes;

[0036] The initial outflow process module is used to input the three initial flow generation processes into the long short-term memory neural network model to obtain three initial outflow processes; the long short-term memory neural network model introduces peak size constraints, peak occurrence time constraints, and monotonicity constraints;

[0037] The outflow process module is used to input three preliminary outflow processes into the Bayesian average model to obtain the predicted outflow process, and combine them with the measured outflow process in the hydrological data to calculate the final outflow process.

[0038] According to one aspect of this specification, an electronic device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the deep learning multi-model runoff prediction method with embedded runoff generation mechanism.

[0039] According to one aspect of this specification, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the deep learning multi-model runoff prediction method with embedded runoff generation mechanism.

[0040] Compared with the prior art, the beneficial effects of the present invention are:

[0041] 1. This invention proposes to integrate three different flow generation mechanisms into a neural network, which overcomes the shortcomings of the harsh preconditions of ordinary physical models and the lack of interpretability of machine learning models.

[0042] 2. This invention constructs three neural networks and three physically constrained long short-term memory neural network models by embedding the runoff generation mechanism into the neural network and replacing the runoff confluence process with an LSTM model. This ensures the integrity of the overall process, high computational efficiency, and user-friendliness, making the runoff predicted by the multi-model more consistent with hydrological principles.

[0043] 3. This invention uses a Bayesian average model to process the three preliminary outflow processes and combines them with the measured outflow processes in the hydrological data to calculate the final outflow process, thereby achieving runoff prediction. This not only minimizes the uncertainty of model prediction but also ensures the accuracy of the prediction. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 This is a flowchart of a method according to an embodiment of the present invention;

[0046] Figure 2 This is a diagram of the recurrent neural network architecture with an embedded flow generation mechanism in an embodiment of the present invention;

[0047] Figure 3 This is a schematic diagram of the outflow process of the embedded GR4J flow generation mechanism in an embodiment of the present invention;

[0048] Figure 4 This is a schematic diagram of the outflow process in an embodiment of the present invention, incorporating a conventional linear flow generation mechanism;

[0049] Figure 5 This is a schematic diagram of the outflow process of the embedded exponential bucket flow generation mechanism in an embodiment of the present invention;

[0050] Figure 6 This is a schematic diagram of the outflow process after integration using the BMA model in an embodiment of the present invention. Detailed Implementation

[0051] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0052] like Figure 1As shown, this embodiment of the invention provides a deep learning multi-model runoff prediction method with embedded runoff generation mechanism, including: Step 1, basic data preparation; Step 2, constructing three neural network models based on runoff generation mechanism; Step 3, constructing a physically constrained LSTM model as a runoff confluence module; Step 4, inputting the preliminary outflow process calculated by the three neural network models into the BMA model to calculate the final multi-model runoff prediction result.

[0053] Specifically, the embodiments of the present invention also provide the specific content of step 1. The basic data mainly include the meteorological dataset (daily rainfall, daily maximum / minimum temperature, sunshine duration, etc.) driving the neural network and the hydrological dataset (measured daily runoff) supervising the training of the neural network. The dataset is divided into a training set and a validation set in a ratio of 7:3.

[0054] Specifically, this embodiment of the invention obtains diurnal hydro-meteorological data and watershed attribute data from the open-source, large-sample CAMELS dataset of catchment area attributes and meteorology. The dataset includes 671 US watersheds with minimal human interference, of which only 569 watersheds with complete continuous hydrological observation records from October 1, 1985 to December 31, 2008 are considered. The neural network model requires rainfall, temperature, and sunshine duration as input to predict the target runoff. The training period was 1985-2001, and the validation period was 2002-2008, with a training-to-validation ratio of 7:3. This embodiment of the invention uses only one watershed (numbered 01013500) as an example to apply the method.

[0055] Specifically, this embodiment of the invention also provides the specific content of step 2, which involves embedding the GR4J runoff generation mechanism, the ordinary linear runoff generation mechanism, and the exponential bucket runoff generation mechanism into the neurons of a recurrent neural network, constructing a neural network model embedding the GR4J runoff generation mechanism, a neural network model embedding the ordinary linear runoff generation mechanism, and a neural network model embedding the exponential bucket runoff generation mechanism. Simultaneously, the preprocessed meteorological dataset from step 1 is input as the feature values ​​of the neural network to train the three neural network models, resulting in three preliminary runoff generation processes.

[0056] Specifically, the formulas for the three flow generation mechanisms in step 2 are as follows:

[0057] (1) GR4J flow generation mechanism

[0058] Effective rainfall and remaining evaporation capacity Calculation:

[0059]

[0060]

[0061] in, This represents the rainfall at the current time step t. This represents the actual evaporation amount at the current time step t.

[0062] like >0 indicates effective rainfall Part of The water will flow into the confluence reservoir, and the remaining water will flow into the runoff reservoir. It is the amount of rainfall that replenishes the runoff reservoir, expressed as soil moisture content. composition.

[0063]

[0064] in, This represents a dimensionless parameter.

[0065] like =0, >0, This refers to the evapotranspiration of the reservoir.

[0066]

[0067] Soil moisture content Continuously updated, the calculation formula is as follows:

[0068]

[0069] Therefore, the runoff of the runoff reservoir Calculated using the following formula:

[0070]

[0071] Find the subtraction The subsequent runoff reservoir storage capacity is:

[0072]

[0073] Total output flow The update formula is:

[0074]

[0075] (2) Ordinary linear flow generation mechanism

[0076] Actual evaporation The calculation formula is:

[0077]

[0078]

[0079] in, Indicates potential evaporation. This indicates the groundwater depth at the current time step t. This refers to the maximum permissible storage depth of groundwater.

[0080] Soil moisture content The updated expression is:

[0081]

[0082] Production flow of each part The updated expression is:

[0083]

[0084]

[0085]

[0086] The final output flow rate is:

[0087]

[0088] in, Indicates the maximum soil moisture content. Indicates the maximum permissible storage depth of groundwater. Indicates the maximum capacity of groundwater. Let represent the runoff caused by rainfall, z be a parameter characterizing the amount of groundwater loss, and k be a dimensionless parameter. The parameter is used to characterize the soil moisture resistance coefficient, which affects the soil's ability to resist evapotranspiration.

[0089] (3) Exponential bucket flow generation mechanism

[0090] and Usually caused by daily rainfall Japanese temperature The calculation yielded:

[0091]

[0092]

[0093] in, Indicates the amount of snow melted. This indicates the amount of rainfall.

[0094] Snowmelt With temperature and snow volume related, It is a parameter characterizing the heat-day factor. That is the highest temperature. It is the lowest temperature:

[0095]

[0096] Evaporation Calculated as potential evapotranspiration Part of it, it depends on the current soil moisture content of the watershed. relative to maximum water storage capacity Size:

[0097]

[0098] Snow volume With soil moisture content The updated formula is:

[0099]

[0100]

[0101] The total production flow is the base flow in the collection tank. When the water collection tank is saturated, the superstorage flow rate is generated. The sum of these can be expressed as follows:

[0102]

[0103]

[0104] (twenty four)

[0105] Specifically, the embedded flow generation mechanism process in step 2 is as follows:

[0106] First, the traditional formula for a recurrent neural network is:

[0107] (25)

[0108] (26)

[0109] in, For input, In cellular state, This represents the cell state at the previous moment. For output, These are the nonlinear activation functions for cell state and output, respectively. These are the weight matrices for cell state and output, respectively. This is a bias between cell state and output.

[0110] Then, the flow generation mechanism is embedded into a variant recurrent neural network, as shown in the following formula:

[0111] (27)

[0112] (28)

[0113] in, These are rainfall, sunshine duration, and temperature, respectively. Soil moisture content, This represents the soil moisture content at the previous moment. The updated formula for soil moisture content, The formula for updating the production flow is as follows: These are the physical parameters in each flow generation mechanism.

[0114] Specifically, during the training process, the weight coefficients and physical parameters of the recurrent neural network are updated through the neural network and adjusted by the backpropagation of the loss function to realize the update of soil moisture content (i.e., cell state variables), so that the update of neurons strictly follows the principle of water balance and enhances the interpretability of machine learning.

[0115] Specifically, such as Figure 2 As shown, embodiments of the present invention provide recurrent neural network architectures before and after embedding the flow generation mechanism. Figure 2 (b)~(d) respectively explain the mechanisms of GR4J runoff, ordinary linear runoff, and exponential bucket runoff, and introduce the input variables of each model, the physical parameters that replace the weights of the neural network, and the state variables of the model. By inputting the rainfall, sunshine duration and temperature of the watershed numbered 01013500 into the neural network model, the runoff process and the change process of soil moisture content in the watershed can be obtained.

[0116] Specifically, the embodiments of the present invention also provide the specific content of step 3, which involves connecting two LSTM models after three neural network models that embed the flow generation mechanism, and adding physical constraints such as peak size, peak occurrence time and flow generation monotonicity to the loss function of the LSTM, inputting the preliminary flow generation process into the LSTM model, and calculating the preliminary outflow process of the three models respectively.

[0117] Specifically, the peak value constraint is mainly achieved by calculating the relative error between the annual simulated peak value and the measured peak value, and then adding it to the loss function with appropriate weights, as shown in the following formula:

[0118] (29)

[0119] in, To simulate peak traffic, This represents the peak value of the measured flow rate.

[0120] Specifically, the peak time constraint is mainly achieved by calculating the relative error between the annual simulated peak time and the measured peak time, and then adding it to the loss function with appropriate weights, as shown in the following formula:

[0121] (30)

[0122] in, To simulate the peak time of traffic, This represents the peak time of the measured traffic flow.

[0123] Specifically, the monotonicity constraint of production and sinking flows means that when the production flow increases, the sinking flow also increases accordingly, and the production flow and sinking flow exhibit a monotonic relationship, as shown in the following formula:

[0124] (31)

[0125] (32)

[0126] in, It is the current time step output flow. It is a slight increase in output. It is a response function, excluding monotonic terms and retaining non-monotonic terms.

[0127] All the physical constraints mentioned above are supplementary modifications to the loss function. The unmodified loss function is:

[0128] (33)

[0129] in, It is the measured flow rate at a single time step. It is the simulated flow rate at a single time step. It is the average value of the simulated traffic.

[0130] The loss function after adding physical constraints is:

[0131] (34)

[0132] in, These are custom weights, and the optimal weights for the loss function are determined through trial and error. .

[0133] Specifically, after constructing the physically constrained LSTM model, it is then connected to the neural network embedding the runoff generation mechanism described in step 2. The preliminary runoff generation process is input into the LSTM model, with meteorological elements such as rainfall, temperature, and daily duration as feature values. Measured runoff is used as the supervised training dataset to drive the LSTM training. The hyperparameters during training are determined through trial and error, including: learning rate = 0.001, packet loss rate = 0.1, batch size = 128, and number of iterations = 500. Finally, the outflow process is calculated, as follows: Figures 3-5 As shown.

[0134] Specifically, the embodiments of the present invention also provide the specific content of step 4, which involves inputting the preliminary outflow process calculated by the three neural network models into the Bayesian average model, calculating the posterior distribution of the preliminary outflow process and the measured outflow process, calculating the weight of each model, and combining the weights to calculate the final outflow process.

[0135] Specifically, the three preliminary outflow processes calculated in step 3 are selected as members of the set and input into the BMA model. The probability density function (PDF) of the Bayesian average probability prediction model is shown below:

[0136] (35)

[0137] in, (k=1,2,3) represents the set of predictions from all models. Indicates the forecast variable, For observation data of length t, Given a prediction model and observation data The posterior distribution of the model. Let the model weights be... ,have , for If the variance is such that formula (35) can be expressed as:

[0138] (36)

[0139] The mean and variance of the posterior distribution predicted by the BMA model are expressed as follows:

[0140] (37)

[0141] (38)

[0142] The variance of a BMA model is essentially a measure of BMA uncertainty, and the BMA parameters are typically calculated iteratively using the expectation-maximization algorithm. and .

[0143] After calculating the weights of each model, the final outflow process is calculated by combining these weights:

[0144] (39)

[0145] in, These are the weights of each model. This is the initial outflow process of each model ( ). Calculated Substituting into the above equation, we get:

[0146] (40)

[0147] The outflow processes of the three models are integrated using the BMA model, and the integrated outflow process is finally calculated. This can minimize the uncertainty of deep learning methods and improve prediction accuracy, such as Figure 6 As shown.

[0148] The implementation of the various embodiments of the present invention is based on programmed processing through a device with processor functionality. Therefore, in practical engineering, the technical solutions and functions of the various embodiments of the present invention are encapsulated into various modules. Based on this reality, and building upon the above embodiments, the embodiments of the present invention provide a deep learning multi-model runoff prediction system with embedded runoff generation mechanisms. This system is used to execute a deep learning multi-model runoff prediction method with embedded runoff generation mechanisms from the above method embodiments.

[0149] The system includes: a data acquisition module for acquiring data, including meteorological and hydrological data; a preliminary runoff generation module for inputting meteorological data into three neural networks embedded with GR4J runoff generation mechanism, ordinary linear runoff generation mechanism, and exponential bucket runoff generation mechanism respectively, to obtain three preliminary runoff generation processes; a preliminary outflow process module for inputting the three preliminary runoff generation processes into a long short-term memory neural network model to obtain three preliminary outflow processes; the long short-term memory neural network model introduces peak size constraints, peak occurrence time constraints, and monotonicity constraints; and an outflow process module for inputting the three preliminary outflow processes into a Bayesian averaging model to obtain a predicted outflow process, and combining the predicted outflow process with the measured outflow process in the hydrological data to calculate the final outflow process.

[0150] The deep learning multi-model runoff prediction system with embedded runoff generation mechanisms provided in this invention addresses the various sources of uncertainty and accuracy issues in existing runoff prediction technologies. This system employs several modules to embed three different runoff generation mechanisms into a neural network, using an LSTM neural network to replace the confluence process. Furthermore, a Bayesian averaging model is combined with measured outflow processes from hydrological data to calculate the weights of the three models, resulting in the final outflow process. This ensures the integrity and interpretability of the overall process, making the runoff predicted by the deep learning model more consistent with hydrological principles, and improving prediction accuracy, computational efficiency, and user-friendliness.

[0151] Based on the same inventive concept as the foregoing embodiments, this embodiment of the invention also provides an electronic device, including a memory and a processor. The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize a deep learning multi-model runoff prediction method with embedded runoff generation mechanism as proposed in the above embodiments.

[0152] This invention also provides a computer-readable storage medium storing a computer program thereon. When executed by a processor, this program overcomes various sources of uncertainty and accuracy issues in runoff prediction, ensuring the integrity and interpretability of the overall process. This makes the runoff predicted by the deep learning model more consistent with hydrological principles, improving prediction accuracy, computational efficiency, and user-friendliness.

[0153] The storage medium can be any non-volatile storage device such as a hard disk, solid-state drive, flash drive, or optical disk, used to store computer program code and necessary data files. The stored computer program includes: a data acquisition module, a preliminary flow generation process module, a preliminary outflow process module, and an outflow process module.

[0154] Finally, it should be noted that the above specific embodiments are merely representative examples of the present invention. Obviously, the present invention is not limited to the above specific embodiments and many variations are possible. Any simple modifications, equivalent changes, and alterations made to the above specific embodiments based on the technical essence of the present invention should be considered within the protection scope of the present invention.

Claims

1. A deep learning multi-model runoff prediction method embedding runoff generation mechanisms, characterized in that, include: S1. Acquire data, including meteorological data and hydrological data; S2. Meteorological data is input into three neural networks embedded with the GR4J runoff generation mechanism, the ordinary linear runoff generation mechanism, and the exponential barrel runoff generation mechanism, respectively, to obtain three preliminary runoff generation processes; among them, the process of embedding the runoff generation mechanism into the neural network includes: using physical formulas Replace the nonlinear activation function in the neural network model ,parameter Replace some weight coefficients in the neural network model Soil moisture content Replace state variables A neural network embedding the flow generation mechanism is obtained; the neural network embedding the flow generation mechanism is expressed as: ， ， in, These are rainfall, sunshine duration, and temperature, respectively. Soil moisture content, This represents the soil moisture content at the previous moment. The updated formula for soil moisture content, For the update formulas of each flow generation mechanism, These are the physical parameters in each flow generation mechanism; S3. Input the three preliminary flow generation processes into the long short-term memory neural network model to obtain three preliminary flow outflow processes; the long short-term memory neural network model introduces peak size constraints, peak occurrence time constraints, and monotonicity constraints; S4. Input the three preliminary outflow processes into the Bayesian average model to obtain the predicted outflow process. Combine the measured outflow process in the hydrological data to calculate the final outflow process.

2. The deep learning multi-model runoff prediction method with embedded runoff generation mechanism according to claim 1, characterized in that, The S3 further includes: By adding peak size constraints, peak occurrence time constraints, and monotonicity constraints to the original loss function of the Long Short-Term Memory (LSTM) neural network model, the new loss function expression is as follows: ， in, The original loss function, For peak size constraints, For peak time constraints, As a monotonicity constraint, It is a custom weight.

3. The deep learning multi-model runoff prediction method with embedded runoff generation mechanism according to claim 1, characterized in that, The S3 further includes: The peak size constraint is calculated by comparing the relative error between the annual simulated peak and the measured peak, and then added to the loss function with corresponding weights. The peak time constraint is calculated by comparing the relative error between the annual simulated peak time and the measured peak time, and then added to the loss function with corresponding weights. The monotonicity constraint is based on the monotonic relationship between production and merging flows, and the loss is calculated using a response function.

4. The deep learning multi-model runoff prediction method with embedded runoff generation mechanism according to claim 1, characterized in that, The S4 further includes: The posterior distribution is obtained based on the predicted outflow process and the measured outflow process in the hydrological data. The mean and variance of the posterior distribution are iteratively calculated using the expectation-maximization algorithm to obtain the weights of the three neural networks. Based on the weights of the three neural networks and the three preliminary outflow processes, the final outflow process is obtained.

5. The deep learning multi-model runoff prediction method with embedded runoff generation mechanism according to claim 4, characterized in that, The final outflow process is expressed as: ， in, This indicates the final outflow process. This represents the weights of the three neural networks. This represents the initial outflow process of the three neural networks. .

6. A deep learning multi-model runoff prediction system with embedded runoff generation mechanism, characterized in that, include: The data acquisition module is used to acquire data, including meteorological data and hydrological data; The preliminary runoff generation process module is used to input meteorological data into three neural networks, each embedded with a GR4J runoff generation mechanism, a conventional linear runoff generation mechanism, and an exponential bucket runoff generation mechanism, respectively, to obtain three preliminary runoff generation processes. The process of embedding the runoff generation mechanism into the neural network includes: using physical formulas... Replace the nonlinear activation function in the neural network model ,parameter Replace some weight coefficients in the neural network model Soil moisture content Replace state variables A neural network embedding the flow generation mechanism is obtained; the neural network embedding the flow generation mechanism is expressed as: ， ， in, These are rainfall, sunshine duration, and temperature, respectively. Soil moisture content, This represents the soil moisture content at the previous moment. The updated formula for soil moisture content, For the update formulas of each flow generation mechanism, These are the physical parameters in each flow generation mechanism; The initial outflow process module is used to input the three initial flow generation processes into the long short-term memory neural network model to obtain three initial outflow processes; the long short-term memory neural network model introduces peak size constraints, peak occurrence time constraints, and monotonicity constraints; The outflow process module is used to input three preliminary outflow processes into the Bayesian average model to obtain the predicted outflow process, and combine them with the measured outflow process in the hydrological data to calculate the final outflow process.

7. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the deep learning multi-model runoff prediction method with embedded runoff generation mechanism as described in any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the deep learning multi-model runoff prediction method with embedded runoff generation mechanism as described in any one of claims 1 to 5.

Images