CN-121980261-A - Adaptive hydrologic model structure optimization and migration method based on deep learning

Abstract

The invention discloses a self-adaptive hydrologic model structure optimization and migration method based on deep learning, which relates to the technical field of hydrologic forecasting, and aims at key hydrologic flux, a plurality of calculation norms are fused, a model structure is optimized through dynamic weight combination, a deep learning model is built, dynamic meteorological data and basin static attributes are used as inputs, nonlinear mapping relations between learning model parameters and weights and meteorological conditions and basin attributes are used as inputs, a coupling model of the deep learning model and the hydrologic model is built, a model pre-trained on a large-scale multi-basin dataset is applied to Shaanxi Wei river and Han river basin by utilizing a migration learning technology, and model fine adjustment is carried out through a small amount of local data, so that rapid localization migration and high-precision runoff forecasting are realized.

Inventors

• LUO JUNGANG
• JING XIN
• WU YUCHEN
• WANG NI
• WEI NA
• YANG XUE
• ZHANG XIAO
• ZUO GANGGANG

Assignees

• 西安理工大学

Dates

Publication Date

20260505

Application Date

20260112

Claims (6)

1. 1. The adaptive hydrological model structure optimization and migration method based on deep learning is characterized by comprising the following steps of: (1) Designing and constructing a bottom frame based on HMETS hydrologic models, integrating a snow melting calculation module, a soil water calculation module and a slope converging module in the frame, and constructing a normal differential equation set by utilizing a water quantity balance relation and a hydrologic flux calculation formula in each module to form basic hydrologic calculation logic; (2) Aiming at key hydrologic flux which has obvious influence on runoff response in the hydrologic model, including actual evaporation and emission, a runoff producing mechanism and in-soil flow, systematically selecting formulas with the same functions but different calculation principles from a plurality of industry-accepted classical hydrologic models as alternative calculation paradigms; (3) The original formula and all the alternative formulas of each key hydrologic flux are respectively endowed with a dynamically adjustable weight coefficient, and a flexible mixed calculation expression is constructed in a weighted summation mode, so that a mixed structure hydrologic model with dynamic self-adaption capability is formed; (4) Constructing a long short-time memory neural network (LSTM) and a Deep Neural Network (DNN), wherein the input characteristics of the LSTM and the DNN comprise dynamic meteorological data sequences which continuously change along with time and static physical attributes which characterize the inherent characteristics of a river basin and do not change along with time, the output characteristics of the LSTM correspond to the dynamic weight of the hybrid structure hydrologic model, and the output characteristics of the DNN correspond to the static parameters of the hybrid structure hydrologic model; (5) Substituting the time-varying weights and the fixed parameters predicted by the LSTM and DNN in each time step into the mixed structure hydrologic model in real time, and driving the model to complete runoff calculation at the moment to obtain a complete runoff process prediction result; (6) The coupling body of the LSTM and the hydrologic model with a mixed structure is pre-trained by utilizing a large-scale and multi-basin open hydrologic data set, so that the complex nonlinear mapping rule between meteorological-basin attribute and optimal model structure is learned and mastered; (7) Local natural attribute data and hydrometeorologic observation data of the target river basin are prepared, and the pre-trained coupling model parameters are finely adjusted by utilizing a migration learning technology so as to realize rapid migration and localized deployment of model knowledge, and the method is applied to a high-precision runoff prediction task of the target river basin.
2. 2. The adaptive hydrological model structure optimization and migration method based on deep learning according to claim 1, wherein the HMETS hydrological model in step (1) is a bottom frame based on a temperature index method, the soil water calculation module is based on a multilayer soil moisture dynamic balance principle, and the slope confluence module is based on a linear reservoir model, wherein: the snow melting calculation module takes the average temperature and the snow fall amount as input, distinguishes the snow fall and the rainfall through a set temperature threshold value, calculates the snow melting amount by combining the snow melting coefficient and the temperature difference, and simultaneously considers the water holding capacity of a snow layer and the liquid water release process to determine the snow melting amount which can be used for producing flow; the soil water calculation module simulates the permeation process of surface water to multi-layer soil and underground water through a layered soil moisture storage and transfer formula, specifically comprises calculation of surface runoff, moisture transfer between soil layers, evaporation loss and deep soil leakage, and ensures water conservation among all components through dynamic balance constraint of soil moisture; And (1.3) calculating the converging time and flow distribution of the surface runoff and the in-soil flow by the slope converging module through a linear reservoir model, wherein the surface runoff is based on a quick response reservoir simulation, the in-soil flow is based on a slow response reservoir simulation, and the flow distribution proportion of each component on a time scale is determined through a storage-release relation of the reservoir.
3. 3. The adaptive hydrologic model structure optimization and migration method based on deep learning according to claim 1, wherein in the step (2), for critical hydrologic flux in the hydrologic model, which has significant influence on runoff response, including actual evaporation, runoff mechanism and in-soil flow, formulas with identical functions and different calculation principles are systematically selected as alternative calculation paradigms from a plurality of industry-accepted classical hydrologic models including HBV model and VIC model, wherein: (2.1) the actual evapotranspiration calculation paradigm comprises an energy balance-based formula and a simplified experience relation-based formula, wherein the actual evapotranspiration calculation paradigm is respectively selected from different models so as to adapt to evapotranspiration estimation requirements of different weather and ground surface conditions; (2.2) the flow mechanism calculation paradigm comprises a infiltration formula based on SCS-CN method and a infiltration formula based on terrain driving, and a rapid flow formula based on soil moisture state and a rapid flow formula based on hydraulic conduction, which are respectively extracted from an HBV model, a VIC model or other classical models to cover the flow characteristics of different drainage basins; (2.3) the calculation paradigm of the in-soil flow comprises a formula based on Darcy law and a formula based on a linear storage model, and the formula is selected from a classical model to simulate the dynamic response of the in-soil flow under different soil types and topography conditions; (2.4) the alternative calculation paradigm ensures that the input parameters and the output results of all formulas are consistent in physical sense through systematic comparative analysis, and can be seamlessly integrated into the bottom framework of the HMETS hydrologic model.
4. 4. The method for optimizing and migrating adaptive hydrologic model structure based on deep learning according to claim 1, wherein in the step (3), dynamically adjustable weight coefficients are respectively given to the original formula and all the alternative formulas of each key hydrologic flux, and a flexible mixed calculation expression is constructed by means of weighted summation, so as to form a hybrid structure hydrologic model with dynamic self-adaption capability, wherein: The weight coefficient comprises two processing methods of discrete weight values and continuous weight values, wherein the discrete weight values are generated through a one-hot method, weight calculation values of all models are converted into vectors only containing single 1, the continuous weight values are generated through a softmax method, and the weight calculation values are converted into weight vectors with the values between 0 and 1 and the sum of 1; (3.2) generating the discrete weights by a one-hot method, namely selecting a single optimal formula from alternative formulas of each key hydrologic flux, wherein the weight value of the corresponding formula in the weight vector is 1, and the rest is 0, so as to realize the selective application of the single formula; (3.3) generating the continuous weights by a softmax method, specifically mapping the weight calculation value of each alternative formula to a range from 0 to 1, and ensuring the sum of all weights to be 1 so as to realize smooth transition of multi-formula weighted combination; (3.4) constructing the mixed calculation expression in a weighted summation mode, and synthesizing calculation results of all the alternative formulas, wherein the weight meets the constraint condition of the sum, so that the physical consistency and calculation stability of the mixed result are ensured; and (3.5) the adjustment of the dynamic weight coefficient is realized through the following neural network prediction, and the weight is dynamically optimized according to the input meteorological data and the river basin attribute so as to adapt to the calculation requirements of different time steps and river basin conditions.
5. 5. The adaptive hydrological model structure optimization and migration method based on deep learning of claim 1, wherein the dynamic meteorological data sequence and static physical properties in the step (4) include: (4.1) the dynamic meteorological data sequence comprises daily precipitation, highest air temperature, lowest air temperature, relative humidity, solar radiation and wind speed data, wherein the data time resolution is a daily scale; (4.2) static physical attributes including soil type, vegetation coverage, average grade, elevation, terrain complexity, and river basin area, etc.; (4.3) the LSTM model adopts a multi-layer structure comprising at least two hidden layers to capture long time series dependencies, and the DNN model comprises at least three fully connected layers to extract nonlinear features of static properties.
6. 6. The adaptive hydrologic model structure optimization and migration method based on deep learning of claim 1, wherein the coupling of LSTM and hybrid structure hydrologic model is pre-trained by using large-scale, multi-basin open hydrologic dataset in steps (6) and (7), and the pre-trained model is fine-tuned by migration learning technology to realize high-precision runoff prediction of target basin, wherein: (5.1) the pre-training stage uses a large-scale, multi-basin open hydrologic dataset CAMELS dataset, the inputs of which comprise dynamic meteorological data sequences including precipitation, air temperature, relative humidity and static basin attributes such as soil type, vegetation coverage, average slope, output as runoff predicted values, the sliding window length of meteorological data is set to 365 days to cover the meteorological change process of the past year; (5.2) the input features and the output features adopt a flow domain normalization method, and the difference of the produced flow under different flow domain areas is weighed by carrying out transformation treatment on the flow, wherein the transformation flow is used for calculating the square root gradient at a stable zero point; (5.3) optimizing model parameters by minimizing a loss function, wherein the loss function comprehensively considers the difference between the simulation value and the observed value, and adopts weighting parameters, wherein the weighting parameters are set to be 0.25 for balancing the loss contribution of the simulation value and the observed value, and the training small batch comprises 100 drainage basins, and each drainage basin comprises 365 days of sequence data; (5.4) the fine tuning stage uses local natural attribute data and hydrokinetic observation data of the target river basin (such as a stop river or a han river) to optimize the pre-training model parameters by minimizing a loss function of the target river basin, the loss function being based on an error between the observed runoff and the predicted runoff of the target river basin; And (5.5) the trimmed model is used for runoff prediction of the target river basin, and a high-precision runoff prediction result is generated by inputting local dynamic meteorological data and static river basin attributes into the model, so that the model is suitable for real-time hydrologic prediction and water resource management of the target river basin.

Description

Adaptive hydrologic model structure optimization and migration method based on deep learning Technical Field The application belongs to the technical field of hydrologic forecasting, and particularly relates to a self-adaptive hydrologic model structure optimization and migration method based on deep learning. Background Hydrologic forecasting is a key technical support for water resource planning, flood control, drought resistance, reservoir scheduling and other works. The hydrologic model is a core tool for hydrologic forecasting, and a series of complex water circulation processes such as precipitation, evaporation, runoff generation, confluence and the like are simulated through a mathematical physical equation so as to forecast the future runoff change. Traditional hydrologic models, including conceptual models and distributed physical models, are applied generally following the "model build-parameter rating-forecast application" procedure by first selecting a structurally fixed hydrologic model, then using historically observed meteorological and runoff data, rating a set of optimal parameters by an optimization algorithm, and assuming that the set of parameters remains unchanged for the future forecast period. However, the existing hydrologic model has a fundamental limitation in application that the model structure is fixed. In particular, the computational formulas or modules describing critical hydrologic fluxes (e.g., runoff, evapotranspiration, etc.) remain unchanged throughout the simulation once selected. Although studies have attempted to solve the time-varying problem of parameters, the choice of the calculation paradigm of the model floor, i.e., the formula itself, is still static. Any calculation formula is a specific simplification of a complex natural mechanism and is optimal under partial conditions. The fixed calculation structural mode adopted in the prior art cannot be dynamically adapted to different hydrological situations, so that the expression capability of the model on a complex watershed environment is insufficient, and the generalization capability and the prediction precision are limited. The prior art explores both discrete and continuous methods for optimizing hydrologic model structures. The discrete preferred method combines different computational formulas by traversal and determines a set of "optimal" static structures and parameters by optimization algorithms. The continuous optimization method is to assign weights to a plurality of alternative calculation formulas in the same hydrologic process, form a mixed expression through weighted summation, and then solve a fixed optimal weight value by using an optimization algorithm. However, these two methods do not fundamentally solve the problem. Whether it is the final selected single structure or the calculated fixed weight combination, it is still static in nature. Once the basin environment changes, the "optimal solution" obtained based on the historical data loses its effectiveness, and the adaptation capability of the model is still limited. Aiming at the limitations of the prior art, the invention provides a brand new technical scheme, namely, a strong nonlinear mapping capability of a deep learning model is utilized to learn the static attribute of a river basin and the complex relationship between meteorological observation data and an optimal structure of a hydrological model. The deep learning model replaces the traditional optimization algorithm, so that the optimal calculation formula combination or weight describing the key hydrologic process can be dynamically output according to real-time input. The static solidification problem of the model structure is fundamentally overcome, the self-adaptive modeling of different watercourses and different hydrologic situations is realized, and the generalization capability and the prediction precision of the model are obviously improved. Disclosure of Invention Currently, hydrologic models have a general problem of structural solidification, the computational formulas describing the critical hydrologic process cannot be changed once selected, which makes the model difficult to adapt to changing watershed environments. Although the prior art attempts to optimize the model structure by optimization algorithms, the results are still static in nature, and dynamic adaptation cannot be achieved, limiting the generalization ability and prediction accuracy of the model. Accordingly, to overcome the above limitations, an object of the present invention is to provide a method for optimizing and migrating adaptive hydrologic model structures based on deep learning. In order to achieve the above purpose, the technical scheme of the invention is as follows: The invention provides a self-adaptive hydrological model structure optimization and migration method based on deep learning, which comprises the following steps: (1) Designing and constructing a bottom frame based on HMETS hydr