A network architecture design method, system and medium for embedding hydrodynamic process representation

Abstract

The present invention relates to the field of hydrological dynamics technology, and specifically to a network architecture design method, system and medium for implanting hydrological dynamics process representation. The present invention can provide more physically consistent prediction results while maintaining the prediction performance of pure deep learning through a model structure with physical interpretability similar to that of a process-based model. It also has more robust generalization and extrapolation capabilities than pure deep learning models and has powerful intelligent reasoning capabilities for unobserved hydrological relationships.

Description

Technical Field

[0001] The present invention relates to the field of hydrodynamic technology, and in particular to a network architecture design method, system and medium for embedding hydrodynamic process representation. Background Art

[0002] For decades, process-based modeling and deep learning methods have been generally regarded as two completely different scientific paradigms by hydrological scientists. Among them, process-based modeling has advantages in physical interpretability and physical consistency, but it is difficult to effectively utilize information in large data sets. Deep learning methods have shown model performance far exceeding the former in predicting most hydrological variables, and have the potential to gain insights from data into some hydrological processes that are not yet fully understood physically. However, since data is the main source of knowledge for deep learning and has a black box nature, it can only learn from the mismatch between model output and observed values, rather than inferring unobserved hydrological relationships. For example, deep learning can easily learn false relationships that look correct within the training data range, but it is difficult to perceive information outside the data range, that is, it is difficult to infer from measured and measured basins to unobserved and unmeasured basins.

[0003] Taking into account the respective advantages of process-based modeling and deep learning methods, it is necessary to actively explore hybrid modeling strategies, that is, how to make the deep learning model maintain high hydrological prediction performance while achieving physical interpretability similar to process-based modeling, so that the deep learning model can answer specific hydrological science questions and have intelligent reasoning capabilities, so that the deep learning model can contain domain knowledge, improve the hydrological awareness and process understanding of the deep learning model, and ultimately enable hydrological research to benefit from both the explicit knowledge previously accumulated in hydrology and the implicit knowledge in big data, and promote knowledge discovery.

[0004] The hybrid modeling strategies in the prior art for the above-mentioned problems include modeling strategies mixed in the data set and modeling strategies mixed in the loss function stage. The former is a hybrid modeling strategy that adds physical variables related to specific hydrological tasks to the data set used for training. This strategy has been successfully used in the modeling and prediction of river temperature and flow, quantitative inversion of near-surface air temperature and basin runoff modeling. The more information exchanged between the process-based model and deep learning, the more accurate the prediction. However, since the current process-based model does not fully understand the complex hydrological process, it is difficult to ensure that the simulated physical variables can truly reflect the physical laws of specific hydrological tasks. In addition, this strategy needs to simulate the corresponding physical variables according to the specific hydrological tasks, and its versatility is poor. The latter aims to add additional regularization terms to the loss function of deep learning so that the training process can achieve a balance between data loss and prior knowledge loss. For example, physical laws such as water balance law, energy dialogue or heuristic methods related to spatiotemporal changes are added to punish solution spaces that do not conform to the added regularization terms. Although it can implicitly provide hydrological prior knowledge and reduce the demand for data volume, the limitation of this strategy is that the loss function implements weak supervision under a "soft" constraint, which is essentially a balance between data loss and prior knowledge loss, and cannot guarantee that the learning results will conform to the introduced physical constraints. In addition, the fact that the deep learning model does not change its black box nature due to the change of the loss function still does not have the physical interpretability similar to the process-based model.

[0005] Although the existing technology attempts to mix the above methods with a variety of hybrid modeling strategies, its essence is still closer to improving the hydrological awareness and process understanding of deep learning models. However, it is lacking and insufficient to directly and effectively integrate physical and deep learning elements in the deep learning framework, design a network architecture for expressing similar process-based hydrological dynamics, and then use it to improve the hydrological awareness and process understanding of deep learning models.

[0006] Therefore, it is necessary to design a network architecture design method, system and medium for embedding hydrological dynamic process representation, and to focus on achieving a more physically realistic deep learning model by building a process-based network architecture, so that it is closer to the purpose of improving the hydrological awareness and process understanding of the deep learning model. Summary of the invention

[0007] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a network architecture design method for implanting hydrological dynamic process representation, which focuses on realizing a more physically realistic deep learning model by constructing a process-based network architecture, making it closer to the purpose of improving the hydrological awareness and process understanding of the deep learning model.

[0008] In order to achieve the above object, the present invention provides a network architecture design method for embedding hydrodynamic process representation:

[0009] The hybrid form of neural differential equations is used to simultaneously train and estimate the parameter vector w and the hydrological model parameter p. The fitting is performed by minimizing the loss function between the data and the solution of the hybrid differential equation. The final model is parameterized by the p and w parameters, and its expression formula is:

[0010] ;

[0011] in, To solve the hydrodynamic control, is the state variable that changes with time t, represents the input forcing variable, p is the hydrological model parameter with physical meaning; To mix in the neural network for solving hydrodynamics, is the input forcing variable, w is the network parameter vector;

[0012] The network architecture includes two Input modules, an Output module and a recurrent neural network (RNN) unit. The Input module includes meteorological forcing data and watershed attribute data.

[0013] The Output module is used to respond to input and output the target variable of interest;

[0014] The RNN unit is set between the Input module and the Output module. The RNN unit is implanted with the process-based hydrodynamic element EXP-HYDRO hydrological model, which is used to include other hydrological sub-processes and hydrodynamic equations related to basin runoff in the architecture, forcing the architecture model to understand and learn the runoff sequence in a process-based manner during training and maintain the water conservation state of the basin;

[0015] The output of the RNN unit includes a neural network as a post-processor to further fit its system response;

[0016] The network architecture estimates both network parameters and hydrological model parameters during training, and continuously updates the model structure through the loss of meteorological forcing data and runoff observations;

[0017] The network architecture includes region-dependent attributes to improve the transferability of the architecture across multiple basins.

[0018] Hybrid forms of neural differential equations also include nested or multiplicative forms.

[0019] The neural network of the post-processor is a fully connected layer, a convolutional layer, or a recurrent layer.

[0020] Region-related attributes include climate and topographic features.

[0021] A computer system is used to execute a network architecture design method for embedding a hydrodynamic process representation.

[0022] A storage medium is used to store a computer system in which a network architecture design method for hydrological dynamic process representation is embedded.

[0023] Compared with the prior art, the present invention has the following beneficial effects:

[0024] The present invention improves the physical interpretability of pure deep learning models by implanting process-based models into deep learning network architectures, and enables deep learning models to have process understanding and hydrological awareness similar to process-based models.

[0025] In order to enhance the data learning capability of the process-based model, the present invention also embeds a neural network at the HODE level, which is combined with the embedded HODE to form a hybrid neural HODE.

[0026] The hybrid neural HODE is based on a forward Euler scheme to learn an integral operator that varies according to the time step. The results of basin runoff modeling show that the present invention can achieve comparable model performance to pure deep learning models while maintaining a good physical interpretability structure.

[0027] In addition to predictive capabilities, the present invention incorporates elements of physical representation that allow accurate inferences to be made about other relevant water dynamics, and therefore also enables good inferences about unobserved hydrological relationships. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 It is a schematic diagram of the model architecture of the present invention; DETAILED DESCRIPTION

[0029] The present invention will now be further described with reference to the accompanying drawings.

[0030] See also Figure 1 The present invention provides a network architecture design method for embedding hydrodynamic process representation:

[0031] The hybrid form of neural differential equations is used to simultaneously train and estimate the parameter vector w and the hydrological model parameter p. The fitting is performed by minimizing the loss function between the data and the solution of the hybrid differential equation. The final model is parameterized by the p and w parameters, and its expression formula is:

[0032] ;

[0033] in, To solve the hydrodynamic control, is the state variable that changes with time t, represents the input forcing variable, p is the hydrological model parameter with physical meaning; To mix in the neural network for solving hydrodynamics, is the input forcing variable, w is the network parameter vector;

[0034] The network architecture includes two Input modules, an Output module and a recurrent neural network (RNN) unit. The Input module includes meteorological forcing data and watershed attribute data.

[0035] The RNN unit is a recurrent neural network (RNN) unit.

[0036] The Output module is used to respond to input and output the target variable of interest;

[0037] The RNN unit is set between the Input module and the Output module. The RNN unit is implanted with the process-based hydrodynamic element EXP-HYDRO hydrological model, which is used to include other hydrological sub-processes and hydrodynamic equations related to basin runoff in the architecture, forcing the architecture model to understand and learn the runoff sequence in a process-based manner during training and maintain the water conservation state of the basin;

[0038] The output of the RNN unit contains a neural network as a post-processor to further fit its system response;

[0039] The network architecture estimates both network parameters and hydrological model parameters during training, and continuously updates the model structure through the loss of meteorological forcing data and runoff observations;

[0040] The network architecture includes region-dependent attributes to improve the transferability of the architecture across multiple basins, which can flexibly generate local / regional models and adapt to various data.

[0041] Hybrid forms of neural differential equations also include nested or multiplicative forms.

[0042] The neural network of the post-processor is a fully connected layer, a convolutional layer, or a recurrent layer.

[0043] Region-related attributes include climate and topographic features.

[0044] The present invention provides a computer system for running a network architecture design method for embedding a hydrodynamic process representation.

[0045] A storage medium is used to store a computer system for running a network architecture design method for embedding a hydrodynamic process representation.

[0046] The above are only preferred embodiments of the present invention, which are only used to help understand the method and core idea of ​​the present application. The protection scope of the present invention is not limited to the above embodiments. All technical solutions under the idea of ​​the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.

[0047] The present invention solves the shortcomings of the hybrid modeling strategy for hydrological observation and prediction in the prior art by introducing the simulation data or physical constraints of the process-based model into the input and loss function of the deep learning to add "solidified" physical knowledge to the deep learning model; taking into full consideration that the hydrological system is a highly dynamic system, the physical principles in hydrology are more differentiated by differential equations for hydrological processes, and these differential equations can be solved in any form in theory by analytical solution. From the perspective of how to directly and effectively integrate physical and deep learning elements in the deep learning framework and design a network architecture for hydrological dynamics, this invention uses a model structure with physical interpretability similar to that of a process-based model to provide more physically consistent prediction results while maintaining the prediction performance of pure deep learning, and has more robust generalization and extrapolation capabilities than pure deep learning models, and has powerful intelligent reasoning capabilities for unobserved hydrological relationships.

Claims

1. A network architecture design method for embedding hydrodynamic process representation, characterized in that: The hybrid form of neural differential equations is used to simultaneously train and estimate the parameter vector w and the hydrological model parameter p. The fitting is performed by minimizing the loss function between the data and the solution of the hybrid differential equation. The final model is parameterized by the p and w parameters, and its expression formula is: ； in, To solve the hydrodynamic control, is the state variable that changes with time t, represents the input forcing variable, p is the hydrological model parameter with physical meaning; To mix in the neural network for solving hydrodynamics, is the input forcing variable, w is the network parameter vector; The network architecture includes two Input modules, an Output module and a recurrent neural network RNN ​​unit, wherein the Input module includes meteorological forcing data and watershed attribute data; The Output module is used to respond to input and output the target variable of interest; The RNN unit is arranged between the Input module and the Output module, and the RNN unit is implanted with the process-based hydrodynamic element EXP-HYDRO hydrological model, which is used to include other hydrological sub-processes and hydrodynamic equations related to the basin runoff in the architecture, and force the architecture model to understand and learn the runoff sequence in a process-based manner during the training process, and maintain the water conservation state of the basin; The output of the RNN unit includes a neural network as a post-processor to further fit its system response; The network architecture estimates both network parameters and hydrological model parameters during training, and continuously updates the model structure through the loss of meteorological forcing data and runoff observations; The network architecture includes region-dependent attributes to improve the portability of the architecture between multiple watersheds.

2. The network architecture design method for embedding hydrodynamic process representation according to claim 1, characterized in that: The mixed form of the neural differential equation is a nested or multiplicative form.

3. The network architecture design method for embedding hydrodynamic process representation according to claim 1, characterized in that: The neural network of the post-processor is a fully connected layer, a convolutional layer or a recurrent layer.

4. The network architecture design method for embedding hydrodynamic process representation according to claim 1, characterized in that: The region-related attributes include climate and terrain characteristics.

Images