Runoff prediction model training, runoff prediction method and device, and electronic equipment

Abstract

The present disclosure relates to a runoff prediction model training, a runoff prediction method and device, and an electronic device. The method comprises: dividing a target region into a plurality of sub-basins; determining average meteorological driving data, static basin attribute data and river section attribute data corresponding to each sub-basin; determining runoff prediction results corresponding to each sub-basin based on a runoff prediction model coupled by deep learning and ecological hydrological processes, according to the average meteorological driving data, the static basin attribute data and the river section attribute data corresponding to each sub-basin; determining a prediction loss corresponding to the runoff prediction model according to runoff monitoring data of a first hydrological station in the target region and the runoff prediction results corresponding to the sub-basin where each first hydrological station is located; and training the runoff prediction model according to the prediction loss, and the trained runoff prediction model is used for runoff prediction of any one sub-basin in the target region. The present disclosure can accurately and reliably predict runoff in a data-deficient basin.

Description

Technical Field

[0001] This disclosure relates to the field of hydrological forecasting technology, and in particular to a runoff forecasting model training, runoff forecasting method and device, and electronic equipment. Background Technology

[0002] Runoff forecasting is a fundamental research issue in the field of hydrology. High-precision runoff forecasts can provide decision-making basis for water resource management and flood disaster prevention. In existing technologies, forecasting methods based on data-driven models have poor applicability in watersheds with limited data and insufficient reliability in non-steady-state environments; forecasting methods based on process-driven models may suffer from problems such as insufficient process generalization and difficulty in parameterization, resulting in insufficient accuracy of runoff forecasts. Summary of the Invention

[0003] In view of this, this disclosure proposes a technical solution for runoff forecasting model training, runoff forecasting method and device, and electronic equipment.

[0004] According to one aspect of this disclosure, a method for training a runoff forecasting model is provided, comprising: dividing a target area into watersheds to determine multiple sub-watersheds; determining average meteorological driving data, static watershed attribute data, and river segment attribute data corresponding to each sub-watershed, wherein the average meteorological driving data represents the meteorological data of the sub-watershed; determining a runoff forecasting result for each sub-watershed based on a runoff forecasting model coupled with deep learning and eco-hydrological processes, according to the average meteorological driving data, static watershed attribute data, and river segment attribute data corresponding to each sub-watershed; determining a forecasting loss corresponding to the runoff forecasting model based on runoff monitoring data from a first hydrological station within the target area and the runoff forecasting results corresponding to each sub-watershed where the first hydrological station is located, wherein the first hydrological station is a pre-set hydrological station within the target area used for model training and for verifying the accuracy of the runoff forecasting model in the time dimension; and training the runoff forecasting model based on the forecasting loss, wherein the trained runoff forecasting model is used to forecast runoff for any sub-watershed in the target area.

[0005] In one possible implementation, the step of dividing the target area into watersheds and determining multiple sub-watersheds includes: determining the digital elevation model data corresponding to the target area and the location of each hydrological station within the target area; and, based on the digital elevation model data and the location of each hydrological station, dividing the target area into watersheds and extracting the river network to determine the multiple sub-watersheds and the river network corresponding to the target area.

[0006] In one possible implementation, determining the average meteorological driving data corresponding to each sub-basin includes: determining the original meteorological driving data corresponding to each sub-basin based on meteorological data within a preset distance range of each sub-basin and at the preset time scale; and calculating the average value of the original meteorological driving data corresponding to any given sub-basin to determine the average meteorological driving data corresponding to that sub-basin.

[0007] In one possible implementation, the static watershed attribute data includes: climate attributes, vegetation attributes, topographic attributes, and soil attributes; determining the static watershed attribute data corresponding to each sub-watershed includes: for any given sub-watershed, determining the original climate data, original vegetation data, original topographic data, and original soil data within the sub-watershed; determining the average value of the original climate data within the sub-watershed as the climate attribute corresponding to the sub-watershed; determining the average value of the original vegetation data within the sub-watershed as the vegetation attribute corresponding to the sub-watershed; determining the average value, maximum value, minimum value, and standard deviation of the original topographic data within the sub-watershed as the topographic attribute corresponding to the sub-watershed; and determining the average value, maximum value, minimum value, and standard deviation of the original soil data within the sub-watershed as the soil attribute corresponding to the sub-watershed.

[0008] In one possible implementation, determining the river segment attribute data corresponding to each of the sub-basins includes: for any one of the sub-basins, determining the connecting river segment between the sub-basin and adjacent sub-basins based on the river network; and determining the river segment attribute data corresponding to the sub-basin at the connecting river segment based on a preset buffer range.

[0009] In one possible implementation, determining the runoff forecast result for each sub-basin based on the runoff forecast model and according to the average meteorological driving data, static watershed attribute data, and river segment attribute data corresponding to each sub-basin includes: determining an original training dataset based on the average meteorological driving data, static watershed attribute data, and river segment attribute data corresponding to each sub-basin; performing data preprocessing on the original training dataset to determine a standardized training dataset, wherein the data preprocessing includes standardization processing and data sequence partitioning processing, and the standardized training dataset includes multiple time-series data sequences; and determining the runoff forecast result for each sub-basin based on the runoff forecast model and according to the original training dataset and the standardized training dataset.

[0010] In one possible implementation, the runoff forecasting model includes: a deep learning module, a sub-basin runoff generation module, and a river confluence module; the step of determining the runoff forecasting result for each sub-basin based on the runoff forecasting model, according to the original training dataset and the standardized training dataset, includes: inputting the standardized training dataset into the deep learning module to determine the physical parameters of the sub-basin runoff generation module and the river confluence module, wherein the physical parameters include dynamic parameters and static parameters, and the dynamic parameters are used to eliminate the uncertainty of the sub-basin runoff generation module and the river confluence module in generalizing eco-hydrological processes; inputting the original training dataset and the physical parameters into the sub-basin runoff generation module and the river confluence module to determine the runoff forecasting result for each sub-basin.

[0011] In one possible approach, the method further includes replacing a sub-module in the sub-basin runoff generation module and / or the river confluence module with a preset neural network sub-module, wherein the neural network sub-module is used to reduce the generalization error of the eco-hydrological process by the sub-basin runoff generation module and / or the river confluence module.

[0012] In one possible implementation, determining the forecast loss corresponding to the runoff forecasting model based on the runoff monitoring data of the first hydrological station in the target area and the runoff forecasting results corresponding to the sub-basin where each first hydrological station is located includes: determining the runoff forecasting results during the preheating period and the simulation period according to a preset preheating time; determining the runoff monitoring data of each first hydrological station in the target area during the simulation period and the loss weight corresponding to each first hydrological station; and determining the forecast loss based on the runoff monitoring data of each first hydrological station during the simulation period, the runoff forecasting results of the sub-basin where each first hydrological station is located during the simulation period, and the loss weight.

[0013] In one possible implementation, the method further includes: dividing the hydrological stations within the target area into a first hydrological station and a second hydrological station based on the leave-one-out method; for any second hydrological station, verifying the accuracy of the runoff forecast model in the spatial dimension according to the runoff forecast results corresponding to the sub-basin where the second hydrological station is located, and determining the first verification result corresponding to the second hydrological station.

[0014] In one possible implementation, the method further includes: for any one of the sub-basins, determining the remote sensing monitoring data corresponding to the sub-basin; during the runoff forecasting process of the runoff forecasting model for the sub-basin, determining the hydrological variable simulation results corresponding to the sub-basin; and verifying the accuracy of the runoff forecasting model based on the remote sensing monitoring data and the hydrological variable simulation results, thereby determining a second verification result corresponding to the sub-basin.

[0015] According to another aspect of this disclosure, a runoff forecasting method is provided, comprising: dividing a target area into watersheds to determine multiple sub-watersheds; for any one of the sub-watersheds, determining meteorological driving data for that sub-watershed during a forecast period, wherein the meteorological driving data represents meteorological data for that sub-watershed; inputting the meteorological driving data into a runoff forecasting model to determine the runoff forecasting result corresponding to that sub-watershed during the forecast period, wherein the runoff forecasting model is trained using the runoff forecasting model training method described in any one of the above embodiments.

[0016] According to another aspect of this disclosure, a runoff forecasting model training device is provided, comprising: a sub-basin division module for dividing a target area into watersheds and determining multiple sub-basins; a sample data determination module for determining average meteorological driving data, static watershed attribute data, and river segment attribute data corresponding to each sub-basin, wherein the average meteorological driving data represents the meteorological data of the sub-basin; and a runoff forecasting module for determining the average meteorological driving data, static watershed attribute data, and river segment attribute data corresponding to each sub-basin based on a runoff forecasting model coupled with deep learning and eco-hydrological processes, and determining the average meteorological driving data, static watershed attribute data, and river segment attribute data corresponding to each sub-basin. The system includes: a runoff forecasting module and a loss determination module, used to determine the forecast loss corresponding to the runoff forecasting model based on the runoff monitoring data of the first hydrological station in the target area and the runoff forecasting results corresponding to each sub-basin where the first hydrological station is located. The first hydrological station is a pre-set hydrological station in the target area used for model training and accuracy verification of the runoff forecasting model in the time dimension. A runoff forecasting model training module is used to train the runoff forecasting model based on the forecast loss. The trained runoff forecasting model is used to forecast runoff for any sub-basin in the target area.

[0017] According to another aspect of this disclosure, a runoff forecasting device is provided, comprising: a sub-basin division module for dividing a target area into basins and determining multiple sub-basins; a driving data determination module for determining meteorological driving data for any one of the sub-basins during a forecast period, wherein the meteorological driving data represents meteorological data for the sub-basin; and a runoff forecasting module for inputting the meteorological driving data into a runoff forecasting model to determine the runoff forecasting result corresponding to the sub-basin during the forecast period, wherein the runoff forecasting model is trained using the runoff forecasting model training method described in any one of the above embodiments.

[0018] According to another aspect of this disclosure, an electronic device is provided, comprising: a processor; a memory for storing processor-executable instructions; wherein the processor is configured to implement the above-described method when executing instructions stored in the memory.

[0019] According to another aspect of this disclosure, a non-volatile computer-readable storage medium is provided that stores computer program instructions thereon, wherein the computer program instructions, when executed by a processor, implement the above-described method.

[0020] In this embodiment, by dividing the target area into watersheds, multiple sub-watersheds can be identified, and the average meteorological driving data, static watershed attribute data, and river segment attribute data corresponding to each sub-watershed can be determined. The average meteorological driving data represents the meteorological data of the sub-watershed. Based on a runoff forecasting model coupled with deep learning and eco-hydrological processes, the runoff forecasting result for each sub-watershed can be determined according to the average meteorological driving data, static watershed attribute data, and river segment attribute data. Based on the runoff monitoring data of the first hydrological station within the target area and the runoff forecasting results corresponding to the sub-watershed where each first hydrological station is located, the forecast corresponding to the runoff forecasting model can be determined. The loss is defined as follows: the first hydrological station is a pre-set hydrological station within the target area used for model training and accuracy verification of the runoff forecasting model in the time dimension; based on the forecast loss, a runoff forecasting model can be trained, wherein the trained runoff forecasting model is used to forecast runoff for any sub-basin of the target area, realizing the combination of deep learning networks and physical process generalization, determining a runoff forecasting model with the advantages of both data-driven and process-driven models, so as to realize runoff forecasting for any sub-basin within the target area with limited data, with high accuracy and reliability, and can provide decision support for regional water resources management and flood disaster prevention.

[0021] Other features and aspects of this disclosure will become clear from the following detailed description of exemplary embodiments with reference to the accompanying drawings. Attached Figure Description

[0022] The accompanying drawings, which are included in and form part of this specification, illustrate exemplary embodiments, features, and aspects of this disclosure together with the specification and serve to explain the principles of this disclosure.

[0023] Figure 1 A flowchart is shown illustrating a runoff forecasting model training method according to an embodiment of the present disclosure;

[0024] Figure 2 A schematic diagram showing a sub-basin division result according to an embodiment of the present disclosure is provided.

[0025] Figure 3 A schematic diagram showing runoff monitoring data and runoff forecast results for each hydrological station according to an embodiment of the present disclosure is provided.

[0026] Figure 4 A flowchart of a runoff forecasting method according to an embodiment of the present disclosure is shown;

[0027] Figure 5 A block diagram of a runoff forecasting model training apparatus according to an embodiment of the present disclosure is shown;

[0028] Figure 6 A block diagram of a runoff forecasting device according to an embodiment of the present disclosure is shown;

[0029] Figure 7 A block diagram of an electronic device according to an embodiment of the present disclosure is shown. Detailed Implementation

[0030] Various exemplary embodiments, features, and aspects of this disclosure will now be described in detail with reference to the accompanying drawings. The same reference numerals in the drawings denote elements that have the same or similar functions. Although various aspects of the embodiments are shown in the drawings, they are not necessarily drawn to scale unless specifically indicated otherwise.

[0031] The term “exemplary” as used herein means “serving as an example, embodiment, or illustration.” Any embodiment illustrated herein as “exemplary” is not necessarily to be construed as superior to or better than other embodiments.

[0032] In this document, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. Furthermore, the term "at least one" in this document means any combination of at least two of any one or more elements. For example, including at least one of A, B, and C can mean including any one or more elements selected from the set consisting of A, B, and C.

[0033] Furthermore, to better illustrate this disclosure, numerous specific details are set forth in the following detailed description. Those skilled in the art will understand that this disclosure can be practiced without certain specific details. In some instances, methods, means, components, and circuits well known to those skilled in the art have not been described in detail in order to highlight the main points of this disclosure.

[0034] Runoff forecasting is a fundamental research problem in the field of hydrology. High-precision runoff forecasts can provide an important basis for decision-making in water resource management and flood disaster prevention. Currently, commonly used runoff forecasting methods include two types: data-driven model-based forecasting methods and process-driven model-based prediction methods.

[0035] Data-driven model-based forecasting methods typically rely on historical runoff data sequences or relevant runoff forecasting factors to establish data-driven models such as linear regression and autoregression statistical models, machine learning models, or deep learning models to achieve runoff forecasting. Building data-driven models requires a large amount of accurate and real training data. In hydrology, historical runoff data usually comes from hydrological stations within the target area. However, on the one hand, the number of hydrological stations in the target area may be limited, and the historical runoff data provided may not be sufficient to support the establishment of accurate data-driven models. On the other hand, the historical runoff data provided by hydrological stations only reflects the runoff in their respective watersheds and lacks representativeness for watersheds without hydrological stations. Therefore, in watersheds with limited data, data-driven model-based forecasting methods have poor applicability. Furthermore, data-driven models implicitly assume data independence and distribution; therefore, under non-steady-state environments such as climate change or underlying surface changes, the reliability of data-driven model-based forecasting methods is insufficient.

[0036] Forecasting methods based on process-driven models typically establish process-driven models by describing eco-hydrological and hydrodynamic processes in the form of physical mathematical equations. These models are then parameterized using historical runoff data sequences. Meteorological, vegetation, and soil data are used as inputs to generalize the physical processes that generate runoff, thereby enabling runoff forecasting. However, process-driven models may suffer from insufficient generalization of the physical processes, leading to significant errors in the model structure. Furthermore, parameterizing these models is often quite difficult. These structural errors and parameterization uncertainties result in lower forecast accuracy for process-driven model-based forecasting methods.

[0037] The runoff forecasting model training method provided in this disclosure combines deep learning networks with physical process generalization to determine a runoff forecasting model that possesses the advantages of both data-driven and process-driven models. This enables runoff forecasting in watersheds with limited data, achieving high accuracy and reliability. The runoff forecasting model training method provided in this disclosure is described in detail below.

[0038] Figure 1 A flowchart illustrating a runoff forecasting model training method according to an embodiment of this disclosure is shown. Figure 1 As shown, this runoff forecasting model training method can be executed by electronic devices such as terminal devices or servers. Terminal devices can be user equipment (UE), mobile devices, user terminals, terminals, cellular phones, cordless phones, personal digital assistants (PDAs), handheld devices, computing devices, vehicle-mounted devices, wearable devices, etc. This runoff forecasting model training method can be implemented by a processor calling computer-readable instructions stored in memory. Alternatively, the runoff forecasting model training method can be executed by a server. Figure 1 As shown, the training method for this runoff forecasting model includes:

[0039] In step S11, the target area is divided into watersheds to determine multiple sub-watersheds.

[0040] The target area here refers to the river basin area that needs to be forecasted for runoff, such as the Yellow River source basin (the basin area above the Tangnaihai hydrological station on the upper reaches of the Yellow River in the northeastern part of the Qinghai-Tibet Plateau). It can be flexibly set according to actual needs, and this disclosure does not make specific limitations on it.

[0041] The target area is typically large, with its river network consisting of a main stream and multiple tributaries. To achieve detailed spatiotemporal runoff forecasting for the target area and accurate prediction of key cross-sections of the main stream and tributaries within the river network, the target area can be divided into watersheds, identifying multiple sub-watersheds. A sub-watershed refers to the catchment area above a specific cross-section of the main stream or tributary, with runoff converging between adjacent upstream and downstream sub-watersheds via river channels.

[0042] The process of determining multiple sub-basins will be described in detail later in conjunction with the possible implementation methods of this disclosure, and will not be repeated here.

[0043] In step S12, the average meteorological driving data, static watershed attribute data, and river segment attribute data corresponding to each sub-basin are determined, wherein the average meteorological driving data represents the meteorological data of the sub-basin.

[0044] For any given sub-basin, the corresponding average meteorological driving data, static basin attribute data, and river segment attribute data can be determined. The average meteorological driving data represents the meteorological data for that sub-basin. The sampling time for the meteorological data can be flexibly set according to actual usage needs, such as daily-scale meteorological data. The static basin attribute data represents the runoff generation and confluence attributes of the sub-basin, which typically do not change significantly during the forecast period. The river segment attribute data represents the attributes of the connecting river segments between this sub-basin and adjacent sub-basins.

[0045] The following sections will describe in detail the process of determining the average meteorological driving data, static watershed attribute data, and river segment attribute data for each sub-basin, based on possible implementation methods of this disclosure; these details will not be elaborated upon here.

[0046] In step S13, based on the runoff forecasting model coupled with deep learning and eco-hydrological processes, the runoff forecasting results for each sub-basin are determined according to the average meteorological driving data, static watershed attribute data and river segment attribute data corresponding to each sub-basin.

[0047] The runoff forecasting model coupled with deep learning and eco-hydrological processes provided in this disclosure can combine a data-driven deep learning process with a physical process-driven eco-hydrological model based on the average meteorological driving data, static watershed attribute data, and river segment attribute data corresponding to each sub-watershed, to determine the runoff forecasting results for each sub-watershed. The runoff forecasting results can represent the predicted runoff volume at the outlet of each sub-watershed.

[0048] The following sections will describe in detail the process of determining the runoff forecast results for each sub-basin based on the runoff forecast model, according to the average meteorological driving data, static watershed attribute data, and river segment attribute data corresponding to each sub-basin, in conjunction with the possible implementation methods disclosed herein. These details will not be elaborated upon here.

[0049] In step S14, based on the runoff monitoring data of the first hydrological station in the target area and the runoff forecast results corresponding to the sub-basin where each first hydrological station is located, the forecast loss corresponding to the runoff forecast model is determined. The first hydrological station is a hydrological station in the target area that is preset for model training and for verifying the accuracy of the runoff forecast model in the time dimension.

[0050] Typically, at least one hydrological station is set up within the target area to monitor hydrological information such as runoff, water level, flow direction, sediment content, water temperature, ice conditions, groundwater, and water quality. Through these hydrological stations, the actual runoff at the corresponding river cross-section can be monitored.

[0051] Among multiple hydrological stations in the target area, a predetermined number of stations can be selected as the first hydrological station, and runoff monitoring data corresponding to each first hydrological station can be obtained for training the runoff forecasting model and for verifying the accuracy of the runoff forecasting model in the time dimension. The runoff monitoring data can represent the actual runoff volume corresponding to the sub-basin where the first hydrological station is located.

[0052] Based on a pre-defined loss function, the forecast loss corresponding to the runoff forecasting model can be determined using runoff monitoring data from the first hydrological station within the target area during the training period, as well as runoff forecasting results for each sub-basin where the first hydrological station is located. The forecast loss reflects the error between the runoff forecasting results determined by the runoff forecasting model and the runoff monitoring data, facilitating subsequent adjustments to the runoff forecasting model.

[0053] After training the runoff forecasting model, the trained runoff forecasting model can be validated based on runoff monitoring data from the first hydrological station during the test period, thus verifying the accuracy of the runoff forecasting model in the time dimension.

[0054] The following sections will describe in detail the process of determining the forecast loss corresponding to the runoff forecast model based on the runoff monitoring data of the first hydrological station in the target area and the runoff forecast results corresponding to the sub-basin where each first hydrological station is located, in conjunction with the possible implementation methods of this disclosure. These details will not be elaborated here.

[0055] In step S15, a runoff forecasting model is trained based on the forecast loss. The trained runoff forecasting model is used to forecast runoff for any sub-basin of the target area.

[0056] Based on the predicted loss, the runoff forecasting model can be adjusted accordingly to achieve iterative training of the runoff forecasting model. Specific methods for training the runoff forecasting model can be found in related technical implementations, and this disclosure does not impose specific limitations on them.

[0057] In one example, the runoff forecasting model can be trained based on hyperparameter tuning. Specifically, a trial-and-error approach can be used to adjust different hyperparameter values ​​during training, verify the performance of the runoff forecasting model, and determine the hyperparameter settings with excellent performance. The specific content of the hyperparameters can be flexibly set according to actual usage requirements. For example, it may include parameters such as the learning rate, optimizer, dropout, number of network layers, and network size corresponding to the runoff forecasting model, depending on the specific structure of the neural network used in the runoff forecasting model. This disclosure does not impose specific limitations on this.

[0058] In one example, a runoff forecasting model can be trained based on an early cessation strategy. Specifically, the performance of the runoff forecasting model is repeatedly verified during training. If the forecast loss in a certain epoch decreases compared to the forecast loss in the previous epoch, the model parameters obtained in this epoch are retained. Conversely, if the forecast loss does not decrease after a preset number of training epochs, training is stopped. The specific value of the preset number of training epochs can be flexibly set according to actual usage requirements, such as 15 epochs, and this disclosure does not impose a specific limitation on it.

[0059] In this embodiment, by dividing the target area into watersheds, multiple sub-watersheds can be identified, and the average meteorological driving data, static watershed attribute data, and river segment attribute data corresponding to each sub-watershed can be determined. The average meteorological driving data represents the meteorological data of the sub-watershed. Based on a runoff forecasting model coupled with deep learning and eco-hydrological processes, the runoff forecasting result for each sub-watershed can be determined according to the average meteorological driving data, static watershed attribute data, and river segment attribute data. Based on the runoff monitoring data of the first hydrological station within the target area and the runoff forecasting results corresponding to the sub-watershed where each first hydrological station is located, the predicted loss corresponding to the runoff forecasting model can be determined. The loss is defined as follows: the first hydrological station is a pre-set hydrological station within the target area for model training and accuracy verification of the runoff forecasting model in the time dimension; based on the forecast loss, the runoff forecasting model can be trained, wherein the trained runoff forecasting model is used to forecast runoff for any sub-basin of the target area, realizing the combination of deep learning network and eco-hydrological process generalization, determining a runoff forecasting model with the advantages of both data-driven and process-driven models, so as to realize runoff forecasting for any sub-basin within the target area with limited data, with high accuracy and reliability, and can provide decision support for regional water resources management and water disaster prevention and control.

[0060] In one possible implementation, the target area is divided into watersheds to determine multiple sub-watersheds, including: determining the digital elevation model data corresponding to the target area and the location of each hydrological station within the target area; and, based on the digital elevation model data and the location of each hydrological station, dividing the target area into watersheds and extracting the river network to determine multiple sub-watersheds and the river network corresponding to the target area.

[0061] A Digital Elevation Model (DEM) can digitally simulate ground topography (i.e., digitally represent the surface morphology of the terrain) using limited terrain elevation data. It can be represented as a physical ground model that uses an ordered array of numerical values ​​to represent ground elevation.

[0062] By determining the digital elevation model data corresponding to the target area, the topographic trend of the target area can be determined, thereby enabling the sub-basin division and river network extraction of the target area. The specific method for determining the digital elevation model data corresponding to the target area can refer to implementation methods in related technologies, such as downloading Shuttle Radar Topography Mission (SRTM) data, etc., and this disclosure does not specifically limit it in this regard.

[0063] When forecasting runoff in a sub-basin, the determined runoff forecast result represents the runoff at the outlet of the corresponding sub-basin. Therefore, to facilitate subsequent verification of the runoff forecast result, when dividing the target area into sub-basins based on digital elevation model data, it is also necessary to determine the location of each hydrological station within the target area. This ensures that when a hydrological station exists in a sub-basin, its location is the outlet of that sub-basin. Specific methods for dividing sub-basins can refer to implementation methods in related technologies, such as using Geographic Information System (GIS) software for sub-basin division. This disclosure does not specifically limit this method.

[0064] Figure 2 A schematic diagram illustrating a sub-basin partitioning result according to an embodiment of the present disclosure is shown. Figure 2 As shown, the target area is the Yellow River source basin, with the blue solid line representing the river network of the Yellow River source basin. Based on GIS software, the Yellow River source basin can be divided into 69 sub-basins and its river network extracted according to the digital elevation model data. Furthermore, each sub-basin can be assigned a corresponding number to distinguish it from other sub-basins. Each sub-basin corresponds to at least one river segment and includes at least one river outlet.

[0065] In one possible implementation, determining the average meteorological driving data corresponding to each sub-basin includes: determining the original meteorological driving data corresponding to each sub-basin based on meteorological data within a preset time scale and a preset distance range of each sub-basin; and for any given sub-basin, calculating the average value of the original meteorological driving data corresponding to that sub-basin to determine the average meteorological driving data corresponding to that sub-basin.

[0066] For any sub-basin, meteorological data at a preset time scale within that sub-basin can be identified as the original meteorological driving data corresponding to that sub-basin. The specific method for determining the meteorological data can be found in related technical implementations, and this disclosure does not impose specific limitations on it. The specific content of the meteorological data can be flexibly set according to actual usage needs, such as temperature, precipitation, wind speed, atmospheric pressure, relative humidity, and potential evapotranspiration, and this disclosure does not impose specific limitations on it.

[0067] After determining the original meteorological driving data corresponding to each sub-basin, for any given sub-basin, the average value of the original meteorological driving data can be calculated within the range of that sub-basin, and this average value can be determined as the original meteorological driving data corresponding to that sub-basin.

[0068] In one example, data such as temperature, precipitation, wind speed, atmospheric pressure, and relative humidity at a preset scale can be acquired from each meteorological station within the target area and within a preset distance range. The potential evapotranspiration for each meteorological station is then calculated using the Penman-Monteith formula. After determining the meteorological data for each station, these data can be interpolated to a preset grid scale using an interpolation method. Based on the location and extent of each sub-basin, the original meteorological driving data for each sub-basin is determined. Based on the gridded original meteorological driving data, the average value can be directly calculated to determine the average meteorological driving data for each sub-basin.

[0069] In one example, relevant meteorological monitoring products can be used to directly acquire gridded meteorological data corresponding to the target area. Based on the location and extent of each sub-basin, the gridded raw meteorological driving data corresponding to each sub-basin is determined, and the average value is calculated to determine the average meteorological driving data corresponding to each sub-basin.

[0070] In one possible implementation, the static watershed attribute data includes: climate attributes, vegetation attributes, topographic attributes, and soil attributes. Determining the static watershed attribute data for each sub-watershed includes: for any given sub-watershed, determining the original climate data, original vegetation data, original topographic data, and original soil data within that sub-watershed; determining the average value of the original climate data within that sub-watershed as the climate attribute corresponding to that sub-watershed; determining the average value of the original vegetation data within that sub-watershed as the vegetation attribute corresponding to that sub-watershed; determining the average value, maximum value, minimum value, and standard deviation of the original topographic data within that sub-watershed as the topographic attribute corresponding to that sub-watershed; and determining the average value, maximum value, minimum value, and standard deviation of the original soil data within that sub-watershed as the soil attribute corresponding to that sub-watershed.

[0071] For any sub-basin, its corresponding static watershed attribute data can represent the characteristics of climate, vegetation cover, topography, and soil quality within that sub-basin. Specifically, static watershed attribute data can include climate attributes, vegetation attributes, topography attributes, and soil attributes. The specific content of the climate, vegetation, topography, and soil attributes can be flexibly set according to actual usage needs, and this disclosure does not impose specific limitations on them.

[0072] Referring to Table 1, climate attributes may include data such as average daily precipitation, average daily potential evapotranspiration, drought index, proportion of snowfall to precipitation, frequency of high precipitation events (≥5 times average precipitation), duration of high precipitation events, frequency of low precipitation events (<1 mm / day), and duration of low precipitation events within each sub-basin; vegetation attributes may include data such as grassland proportion, snow cover proportion, water body proportion, multi-year average leaf area index, monthly maximum leaf area index, and monthly minimum leaf area index within each sub-basin; topographic attributes may include data such as basin area, average elevation, slope, and river length within each sub-basin; soil attributes may include data such as the proportion of sandy soil, clay, and silt in the soil, soil porosity, soil profile depth, soil depth to bedrock, saturated water content, Log-10 transformation of saturated hydraulic gradient, saturated soil thermal conductivity, thermal conductivity of frozen saturated soil, and thermal conductivity of dry soil within each sub-basin.

[0073] Table 1

[0074]

[0075]

[0076] For any given sub-basin, the original climate data, original vegetation data, original topographic data, and original soil data within that sub-basin can be determined. Specifically, the original climate data can represent the monitored values ​​of each data point included in the aforementioned climate attributes within the sub-basin at a preset time scale; the original vegetation data can represent the monitored values ​​of each data point included in the aforementioned vegetation attributes within the sub-basin at a preset time scale; the original topographic data can represent the monitored values ​​of each data point included in the aforementioned topographic attributes within the sub-basin at a preset time scale; and the original soil data can represent the monitored values ​​of each data point included in the aforementioned soil attributes within the sub-basin at a preset time scale. The specific methods for determining the original climate data, original vegetation data, original topographic data, and original soil data can refer to implementation methods in related technologies, depending on the specific content of each data point; this disclosure does not specifically limit these methods.

[0077] For climate and vegetation attributes, the spatial heterogeneity within each sub-basin is usually small. Therefore, for any sub-basin, the average value of the original climate data within the sub-basin can be calculated as the corresponding climate attribute; the average value of the original vegetation data within the sub-basin can be calculated as the corresponding vegetation attribute.

[0078] Topographic and soil attributes can vary significantly within each sub-basin, effectively reflecting the heterogeneity within each sub-basin. Therefore, for any given sub-basin, the mean, maximum, minimum, and standard deviation of the raw topographic data can be calculated as the corresponding topographic attribute; similarly, the mean, maximum, minimum, and standard deviation of the raw soil data can be calculated as the corresponding soil attribute. By utilizing these topographic and soil attributes for each sub-basin, the heterogeneity within each sub-basin can be fully characterized, leading to more accurate runoff forecasts for each sub-basin.

[0079] In one possible implementation, determining the river segment attribute data corresponding to each sub-basin includes: for any given sub-basin, determining the connecting river segment between the sub-basin and adjacent sub-basins based on the river network; and determining the river segment attribute data corresponding to the sub-basin at the connecting river segment based on a preset buffer range.

[0080] Based on the above Figure 2 For example, Figure 2 As shown, any two adjacent sub-basins, upstream and downstream, are connected by at least one connecting river segment, for example, Figure 2 Sub-basins numbered 38 and 37 are identified. Based on the extracted river network, the connecting river segments corresponding to any given sub-basin and its adjacent sub-basins can be determined. The intersection point between the connecting river segment and the boundary of the sub-basin is then determined, and this intersection point is designated as the outlet of the sub-basin. To accurately predict runoff at the sub-basin outlet, the river segment attribute data at the connecting river segments needs to be determined.

[0081] River segment attribute data can represent the attributes of the river segments connecting each sub-basin with adjacent sub-basins. The specific content of the river segment attribute data can be flexibly set according to actual usage needs. For example, it may include the upstream catchment area, river segment length, gradient, silt ratio, clay ratio, and sand ratio, etc. This disclosure does not impose specific limitations on this.

[0082] A preset buffer range can be determined within the connecting river segment, serving as the basis for determining the river segment attribute data. The specific value of the buffer range can be flexibly set according to actual usage requirements, for example, it can be set to 30m, etc., and this disclosure does not impose a specific limitation on it. The specific method for determining the river segment attribute data can refer to existing implementation methods, depending on the specific content of the river segment attribute data, and this disclosure does not impose a specific limitation on it.

[0083] In one possible implementation, based on the runoff forecasting model, the runoff forecasting results for each sub-basin are determined according to the average meteorological driving data, static watershed attribute data, and river segment attribute data corresponding to each sub-basin. This includes: determining the original training dataset based on the average meteorological driving data, static watershed attribute data, and river segment attribute data corresponding to each sub-basin; performing data preprocessing on the original training dataset to determine the standardized training dataset, wherein the data preprocessing includes standardization and data sequence partitioning, and the standardized training dataset includes multiple time-series data sequences; and determining the runoff forecasting results for each sub-basin based on the runoff forecasting model, according to the original training dataset and the standardized training dataset.

[0084] Based on the average meteorological driving data, static watershed attribute data, and river segment attribute data corresponding to each sub-basin, the original training dataset, original validation dataset, and original test dataset can be determined. Specifically, the average meteorological driving data, static watershed attribute data, and river segment attribute data corresponding to each sub-basin have a preset time range. According to the preset training period, validation period, and test period ranges, the average meteorological driving data, static watershed attribute data, and river segment attribute data corresponding to each sub-basin can be divided into the original training dataset, the original validation dataset, and the original test dataset, respectively. The specific ranges of the training period, validation period, and test period can be flexibly set according to actual usage requirements; this disclosure does not impose specific limitations on them.

[0085] In one example, the time range for the average meteorological driving data, static watershed attribute data, and river segment attribute data corresponding to each sub-basin is from January 1, 1960 to December 13, 2019. The training period can be set to January 1, 1960 to December 31, 1989, the validation period to January 1, 1990 to December 31, 1999, and the test period to January 1, 2000 to December 31, 2019. This divides the average meteorological driving data, static watershed attribute data, and river segment attribute data corresponding to each sub-basin into the original training dataset, the original validation dataset, and the original test dataset.

[0086] Furthermore, the standard deviation and mean of the original training dataset, original validation dataset, and original test dataset can be calculated separately. Based on the standard deviation and mean, the data in the original training dataset, original validation dataset, and original test dataset can be standardized to obtain the standardized training dataset, standardized validation dataset, and standardized test dataset.

[0087] The standardized training dataset, standardized validation dataset, and standardized test dataset can also be processed to partition the data into sequences. Specifically, the sequence length and sliding window length can be preset to partition the data in the standardized training dataset into multiple time-series data sequences.

[0088] In one example, taking the training period from January 1, 1960 to December 31, 1989 as an example, the sequence length can be set to 6 years and the sliding window length to 1 year. The data from January 1, 1960 to December 31, 1965 can be determined as the first time series data sequence; sliding forward one year, the data from January 1, 1961 to December 31, 1966 can be determined as the second time series data sequence, and so on, thus determining 25 time series data sequences with a sequence length of 6 years, which can be used as the standardized training dataset.

[0089] By inputting the original training dataset and the standardized training dataset into the runoff forecasting model, the runoff forecasting results for each sub-basin can be determined.

[0090] In one example, by inputting the data from the original training dataset and the standardized training dataset corresponding to January 1, 2014 to December 13, 2019 into the runoff forecast model, the runoff forecast results for each sub-basin during the period from January 1, 2014 to December 13, 2019 can be determined.

[0091] In one possible implementation, the runoff forecasting model includes: a deep learning module, a sub-basin runoff generation module, and a river confluence module. Based on the runoff forecasting model, and according to the original training dataset and the standardized training dataset, the runoff forecasting results for each sub-basin are determined, including: inputting the standardized training dataset into the deep learning module to determine physical parameters, wherein the physical parameters include dynamic parameters and static parameters, and the dynamic parameters are used to eliminate the uncertainty of the generalization of eco-hydrological processes by the sub-basin runoff generation module and the river confluence module; inputting the original training dataset and physical parameters into the sub-basin runoff generation module and the river confluence module to determine the runoff forecasting results for each sub-basin.

[0092] Runoff forecasting models can include deep learning modules, sub-basin runoff generation modules, and river confluence modules, thereby coupling deep learning data-driven models with process-driven models that generalize eco-hydrological processes, so that runoff forecasting models have both the accuracy of data-driven models and the reliability of process-driven models.

[0093] The deep learning module can be used to uncover patterns in large sample data, determine the mapping relationship between average meteorological driving data and static watershed attribute data and the parameters of the sub-watershed runoff generation module, as well as the mapping relationship between river segment attribute data and the parameters of the river confluence module.

[0094] By inputting the standardized training dataset into the deep learning module, the physical parameters corresponding to the sub-basin runoff generation module and the river confluence module can be determined. These physical parameters can include dynamic and static parameters. Dynamic parameters can be used to eliminate the uncertainty in the generalization of eco-hydrological processes by the sub-basin runoff generation module. The specific content of the dynamic parameters can be flexibly set according to actual usage needs; for example, it can include parameters such as the proportion of frozen water in the soil at each preset time scale and the dynamic impact of vegetation growth. This disclosure does not impose specific limitations on this. The specific content of the static parameters can also be flexibly set according to actual usage needs; for example, it can include parameters such as the maximum and minimum temperatures within each preset time scale. This disclosure does not impose specific limitations on this. The specific form of the deep learning module can refer to the implementation methods in related technologies. It can determine all physical parameters using only one neural network, or it can determine different physical parameters using different neural networks. This disclosure does not impose specific limitations on this.

[0095] In one example, the deep learning module may include a Long Short-Term Memory (LSTM) network for determining dynamic parameters, and a one-dimensional convolutional network (Conv1D) and a fully connected network (FCNN) for determining static parameters. Conv1D is used to extract features from average meteorological driving data; the extracted features can be concatenated with static watershed attribute data and input into the FCNN to determine the static parameters. The dynamic parameters output by the LSTM and the static parameters output by the FCNN can both be constrained to the range (0, 1) using a sigmoid activation function, and then scaled down to the actual parameter range according to practical physical standards.

[0096] The sub-basin runoff generation module can be used to determine the runoff generation of each sub-basin. Specifically, based on a pre-defined lumped eco-hydrological model, the physical equations of eco-hydrological processes such as precipitation, vegetation interception and transpiration, vegetation biomass renewal, snowmelt, evaporation, soil water, surface runoff, and groundwater runoff can be defined in a differentiable form within a deep learning framework to simulate the runoff generation process of each sub-basin. This allows the runoff generation of each sub-basin to be determined based on the input driving data and static watershed attribute data. The specific form of the lumped eco-hydrological model can refer to implementation methods in related technologies, such as the EXP-HYDRO (exponential bucket hydrological model), the Xin'anjiang model, the WEAP (Water Evaluation and Planning model), and the BGM (Bucket Grassland Model), etc., and this disclosure does not specifically limit it. Similarly, the specific form of the deep learning framework can refer to implementation methods in related technologies, such as the PyTorch framework, the Tensorflow framework, and the Keras framework, and this disclosure does not specifically limit it.

[0097] If the lumped eco-hydrological model used in the sub-basin runoff generation module does not include slope runoff calculation, a corresponding slope runoff module can be added to the deep learning framework. The specific methods used in the slope runoff module can refer to implementation methods in related technologies, such as unitline runoff methods based on gamma distribution, etc., and this disclosure does not specifically limit them.

[0098] In one example, the sub-basin runoff generation module can use the EXP-HYDRO hydrological model and calculate the slope runoff within the sub-basin based on the unit hydrograph runoff method with gamma distribution. The corresponding physical equations are defined in a differentiable form within the PyTorch framework. The slope runoff calculation process based on the unit hydrograph runoff method with gamma distribution can be expressed as formulas (1) and (2):

[0099]

[0100]

[0101] Where Q represents the runoff of the sub-basin without considering the confluence; ξ represents the unit curve; Γ represents the gamma function; a represents the shape parameter of the gamma function; b represents the time parameter of the gamma function; Q * This indicates the runoff generated in the sub-basin after the confluence.

[0102] Since the EXP-HYDRO model determines evapotranspiration based on potential evapotranspiration and available soil moisture, neglecting the dynamic impact of vegetation growth on evapotranspiration, a deep learning module can be used to exponentially increase the dynamic parameter β to represent the dynamic impact of vegetation growth on evapotranspiration, thereby reducing the parameterization uncertainty of the sub-basin runoff generation module. Furthermore, the EXP-HYDRO model does not consider the effects of soil freeze-thaw cycles; therefore, a deep learning module can be used to determine the proportion of frozen water in the soil at each time step using the dynamic parameter α. The frozen water will not participate in evapotranspiration and runoff generation calculations, and the maximum soil water storage capacity can be considered equal to the unfrozen capacity minus the frozen water content. Since soil freeze-thaw cycles also directly affect soil hydraulic conductivity, and thus the soil water outflow coefficient, the outflow coefficient F can also be determined as a dynamic parameter. Through the dynamic parameters α, β, and F, as well as related static parameters, the deep learning module can be tightly coupled with the sub-basin runoff generation module, fully combining the accuracy of the data-driven model with the reliability of the process-driven model.

[0103] The river confluence module can be used at each preset time step to determine the final runoff result of each sub-basin after its runoff converges along the river network. Here, the time step can represent the time it takes for the runoff from the upstream sub-basin to converge at the outlet section of the sub-basin whose runoff needs to be predicted. The specific methods used in the river confluence module can refer to implementation methods in related technologies, such as the Muskingen river confluence method, etc., and this disclosure does not specifically limit them.

[0104] In one example, the Muskingan channel confluence method includes a connecting segment of two sub-basins. Based on the inflow and outflow of the previous preset time step and the inflow of the current time step, the outflow of the current time step is determined, which can be expressed as formula (3):

[0105] q out，2 =C 1qin，2 +C2q in，1 +C3q out，1 (3)

[0106] Where, q out，2 Indicates the outflow at the current time step; q in，2 q represents the inflow at the current time step; in，1 q represents the inflow at the previous time step; out，1 This represents the outflow at the previous time step. Parameters C1, C2, and C3 satisfy C1 + C2 + C3 = 1, and can be expressed by formulas (4) to (6) respectively:

[0107]

[0108]

[0109]

[0110] Where K represents the time coefficient of water storage, in seconds (s); x represents the weighting factor, ranging from 0 to 0.5; Δt represents the simulation time step, in seconds (s). The specific value of the time step can be flexibly set according to actual usage requirements, such as 86400s, etc. This disclosure does not impose specific limitations on it. The relationship between K and X can be expressed as the following formula (7):

[0111] 2KX<Δt<2K(1-X)) (7)

[0112] The inflow to each connecting river segment includes runoff from the upstream basin and outflow from the previous segment. Therefore, the confluence of each tributary can be determined first, and then the confluence of the main channel can be calculated based on the upstream and downstream relationships.

[0113] After determining the physical parameters through the deep learning module, the original training dataset and physical parameters can be input into the sub-basin runoff generation module and the river confluence module. Specifically, the physical parameters related to sub-basin runoff generation, along with the average meteorological driving data and static watershed attribute data from the original training dataset, can be input into the sub-basin runoff generation module. The physical parameters related to river confluence, along with the river segment attribute data from the original training dataset, can be input into the river confluence module. Thus, the runoff generation of each sub-basin is determined through the sub-basin runoff generation module, and the runoff volume after confluence in each sub-basin is determined through the river confluence module, thereby determining the runoff forecast result for each sub-basin.

[0114] By defining eco-hydrological processes in a differentiable form within a deep learning framework, sub-basin runoff generation modules and river confluence modules are obtained. Furthermore, using a neural network-based deep learning module, the physical parameters of these modules are determined. This allows for the coupling of deep learning-based data-driven models with eco-hydrological processes and process-driven models based on these processes. This enables runoff forecasting models to combine the accuracy of data-driven models with the reliability of process-driven models. Moreover, the use of dynamic parameters can reduce the uncertainty caused by insufficient generalization of eco-hydrological processes by the sub-basin runoff generation modules, thereby improving the accuracy and reliability of runoff forecasting results.

[0115] In one possible implementation, the method further includes replacing a sub-module in the sub-basin runoff generation module and / or the river confluence module with a pre-defined neural network sub-module, wherein the neural network sub-module is used to reduce the generalization error of the sub-basin runoff generation module and / or the river confluence module in terms of eco-hydrological processes.

[0116] Furthermore, in this embodiment of the present disclosure, a preset neural network can be used to replace some sub-modules in the sub-basin runoff generation module and / or river confluence module, so as to reduce the generalization error of the sub-basin runoff generation module and / or river confluence module to the eco-hydrological process, thereby further increasing the accuracy and reliability of runoff forecasting.

[0117] In one possible implementation, the forecast loss corresponding to the runoff forecasting model is determined based on the runoff monitoring data of the first hydrological station in the target area and the runoff forecast results corresponding to the sub-basin where each first hydrological station is located. This includes: determining the runoff forecast results during the preheating period and the simulation period according to a preset preheating time; determining the runoff monitoring data of each first hydrological station in the target area during the simulation period and the loss weight corresponding to each first hydrological station; and determining the forecast loss based on the runoff monitoring data of each first hydrological station during the simulation period, the runoff forecast results of the sub-basin where each first hydrological station is located during the simulation period, and the loss weight.

[0118] In runoff forecasting models, some parameters related to eco-hydrological processes may have initial values ​​of 0, such as soil water, which does not conform to actual physical conditions. In this case, the runoff forecast results determined by the runoff forecasting model can be predicted to have a large error relative to the actual values. Therefore, by setting a preset warm-up period, each time series data sequence in the standardized training dataset can be divided into a warm-up period and a simulation period. Using the time series data sequences corresponding to the warm-up period, the eco-hydrological process-related parameters in the runoff forecasting model, whose initial values ​​are 0, can be processed in a manner similar to initialization, bringing them to relatively stable values. Furthermore, the runoff forecast results during the warm-up period are not used to determine the forecast loss; only the runoff forecast results during the simulation period are used to determine the forecast loss.

[0119] In one example, taking the training period from January 1, 1960 to December 31, 1989, with a sequence length of 6 years and a moving window of 1 year, a preset warm-up period of 2 years can be set. Then, in the first time series data from January 1, 1960 to December 31, 1965, January 1, 1960 to December 31, 1961 is the warm-up period, and January 1, 1962 to December 31, 1989 is the simulation period. Runoff forecasts generated between January 1, 1960 and December 31, 1961 are not included in the determination of forecast losses.

[0120] Existing runoff forecasting methods typically only consider runoff monitoring data obtained from hydrological stations at the downstream outlet of the watershed during model training or calibration, while neglecting the use of runoff monitoring data obtained from hydrological stations in the middle and upper reaches. This may lead to difficulties in parameter training and inaccuracies.

[0121] In this embodiment of the disclosure, in order to comprehensively consider the runoff monitoring data of hydrological stations at different locations in the basin and improve the accuracy of runoff forecasting results within the basin, different loss weights can be set for the first hydrological stations at different locations in the basin to adjust the level of attention given to each first hydrological station. The specific value of the loss weight can be flexibly set according to actual usage requirements, and this disclosure does not impose specific limitations on it.

[0122] Based on a preset loss function, the forecast loss corresponding to the runoff forecasting model can be determined according to the runoff monitoring data of each first hydrological station during the simulation period, the runoff forecasting results of the sub-basin where each first hydrological station is located during the simulation period, and the loss weight. The specific form of the loss function can be flexibly set according to actual usage requirements, such as the Nash efficiency coefficient or root mean square error index; this disclosure does not impose specific limitations on it.

[0123] In one example, the target area is the Yellow River source basin. The Yellow River source basin has the Tangnaihai Hydrological Station, Jungong Hydrological Station, Maqu Hydrological Station, Mentang Hydrological Station and Jimai Hydrological Station set up from upstream to downstream.

[0124] The Tangnaihai Hydrological Station in the upstream, the Maqu Hydrological Station in the midstream, and the Jimai Hydrological Station in the downstream were selected as the first hydrological stations, and the loss weight corresponding to each first hydrological station was set to be the same, which was 1 / 3.

[0125] The loss function is set as the Nash efficiency coefficient, and the Nash efficiency coefficient corresponding to each first hydrological station can be expressed as formula (8):

[0126]

[0127] Wherein, NSE represents the Nash efficiency coefficient; This represents the runoff monitoring data corresponding to each first hydrological station at time t; This represents the runoff forecast results for each sub-basin where the first hydrological station is located at time t; This represents the average value of runoff monitoring data at each of the first hydrological stations.

[0128] Based on the Nash efficiency coefficients and loss weights corresponding to the three first hydrological stations, the predicted loss can be determined. The predicted loss can be expressed as formula (9):

[0129]

[0130] Where L represents forecast loss; NSE i φ represents the Nash efficiency coefficient corresponding to each first hydrological station. i This represents the loss weight corresponding to each first hydrological station, and φ1 = φ2 = φ3 = 1 / 3.

[0131] By reasonably setting the loss function and the loss weight corresponding to each first hydrological station, the limited amount of runoff monitoring data provided by the first hydrological station can be used to optimize the overall runoff forecasting model and improve the accuracy of runoff forecasting for each sub-basin within the target area.

[0132] In one possible implementation, the method further includes: based on the leave-one-out method, dividing the hydrological stations in the target area into a first hydrological station and a second hydrological station; for any second hydrological station, verifying the accuracy of the runoff forecast model in the spatial dimension according to the runoff forecast results corresponding to the sub-basin where the second hydrological station is located, and determining the first verification result corresponding to the second hydrological station.

[0133] Based on the leave-one-out method, hydrological stations within the target area can be divided into the first hydrological station and the second hydrological station. During the training process and the time-dimensional validation of the trained runoff forecasting model, only the runoff monitoring data corresponding to the first hydrological station is used. The sub-basin where the second hydrological station is located is also considered as a data-scarce basin where real data cannot be provided. Thus, after the runoff forecasting model is trained, the accuracy of the runoff forecasting model in the data-scarce basin can be validated spatially, based on the station scale, using the runoff monitoring data corresponding to the second hydrological station.

[0134] Specifically, for any second hydrological station, the first verification result corresponding to that second hydrological station can be determined based on the runoff monitoring data corresponding to that second hydrological station and the runoff forecast results of the sub-basin where the second hydrological station is located. The specific form of the first verification result can be flexibly set according to actual usage requirements; for example, it can be the Nash efficiency coefficient and the root mean square error index, etc., and this disclosure does not impose specific limitations on it.

[0135] Taking the aforementioned target area as the Yellow River source basin, and the Yellow River source basin having the Tangnaihai Hydrological Station, Jungong Hydrological Station, Maqu Hydrological Station, Mentang Hydrological Station, and Jimai Hydrological Station set up sequentially from upstream to downstream, as an example. Based on the leave-one method, Tangnaihai Hydrological Station, Maqu Hydrological Station, and Jimai Hydrological Station can be selected as the first hydrological station, and Jungong Hydrological Station and Mentang Hydrological Station can be selected as the second hydrological station. The first verification result can be set to include the Nash efficiency coefficient and the root mean square error index. Among them, the root mean square error index can be expressed as formula (10):

[0136]

[0137] RMSE stands for Root Mean Square Error.

[0138] Figure 3 This diagram illustrates runoff monitoring data and runoff forecast results for each hydrological station according to an embodiment of the present disclosure. Figure 3As shown in the figure, the black dashed lines represent the runoff monitoring data of each hydrological station, and the blue solid lines represent the runoff forecast results corresponding to the sub-basin where each hydrological station is located.

[0139] like Figure 3 As shown, the Nash efficiency coefficient (NSE) between the runoff monitoring data of Tangnaihai Hydrological Station and the runoff forecast results corresponding to its sub-basin is 0.92, and the root mean square error index (RMSE) is 0.10 mm / d; the Nash efficiency coefficient (NSE) between the runoff monitoring data of Jungong Hydrological Station and the runoff forecast results corresponding to its sub-basin is 0.90, and the root mean square error index (RMSE) is 0.12 mm / d; the Nash efficiency coefficient (NSE) between the runoff monitoring data of Maqu Hydrological Station and the runoff forecast results corresponding to its sub-basin is 0.92, and the root mean square error index (RMSE) is 0.12 mm / d. The Nash efficiency coefficient (NSE) is 0.94, and the root mean square error index (RMSE) is 0.09 mm / d. The Nash efficiency coefficient (NSE) between the runoff monitoring data of the Mentang hydrological station and the runoff forecast results corresponding to its sub-basin is 0.79, and the root mean square error index (RMSE) is 0.15 mm / d. The Nash efficiency coefficient (NSE) between the runoff monitoring data of the Jimai hydrological station and the runoff forecast results corresponding to its sub-basin is 0.78, and the root mean square error index (RMSE) is 0.11 mm / d.

[0140] Generally, an NSE greater than 0.6 indicates good model performance. As shown above, the runoff forecasting model of this embodiment has an NSE greater than 0.6 at each hydrological station. At the two upstream stations, Jimai and Mentang, the runoff monitoring data is relatively small due to the aridity of the upstream sub-basin, and the signal-to-noise ratio of low-flow observations is relatively low. Therefore, the NSE is slightly lower than at the downstream stations, but still greater than 0.6, and the overall RMSE is low. Therefore, the runoff forecasting model of this embodiment has high accuracy in runoff forecasting for watersheds with limited data.

[0141] By using the leave-one-out method, the runoff forecasting model can be validated at the hydrological station scale for runoff forecasting in watersheds with limited data within the target area. This allows for further adjustments to the runoff forecasting model based on the initial validation results, thereby improving its performance in watersheds with limited data.

[0142] In one possible implementation, the method further includes: for any sub-basin, determining the remote sensing monitoring data corresponding to the sub-basin; during the runoff forecasting process of the runoff forecasting model for the sub-basin, determining the hydrological variable simulation results corresponding to the sub-basin; and verifying the accuracy of the runoff forecasting model based on the remote sensing monitoring data and the hydrological variable simulation results, thereby determining the second verification result corresponding to the sub-basin.

[0143] In addition to validating the accuracy of runoff forecasting models at the hydrological station scale, the accuracy of runoff forecasting models can also be validated at the sub-basin scale based on remote sensing monitoring data of the sub-basin.

[0144] Specifically, for any sub-basin, during the runoff forecasting process of the runoff forecasting model, simulation results of hydrological variables related to eco-hydrological processes are generated. By determining the remote sensing monitoring data corresponding to the sub-basin and comparing the remote sensing monitoring data with the hydrological variable simulation results, the accuracy of the runoff forecasting model can be verified, and a second verification result corresponding to the sub-basin can be determined. The specific content of the hydrological variable simulation results can be flexibly set according to actual usage needs; for example, it may include evapotranspiration data and vegetation transpiration data within the sub-basin, etc., which this disclosure does not specifically limit. The specific method for determining the remote sensing monitoring data can refer to implementation methods in related technologies; for example, it can be obtained based on the Global Land Surface Data Assimilation System (GLDAS), etc., which this disclosure does not specifically limit. The specific form of the second verification result can be flexibly set according to actual usage needs; for example, it may include Pearson correlation coefficient and root mean square error index, etc., which this disclosure does not specifically limit.

[0145] In one example, the hydrological variable simulation results may include evapotranspiration data, and the remote sensing monitoring data may be evapotranspiration data acquired based on GLDAS. The second validation results may include Pearson correlation coefficient, root mean square error index, and spatial performance index SPAEF. Specifically, the Pearson correlation coefficient and root mean square error index can be used to evaluate the dynamic changes of the hydrological variable simulation results in the time dimension; the spatial performance index SPAEF can be used to evaluate the dynamic changes of the hydrological variable simulation results in the spatial dimension.

[0146] The Pearson correlation coefficient and the spatial performance index SPAEF can be expressed by formulas (11) to (14):

[0147]

[0148]

[0149]

[0150]

[0151] Where R represents the Pearson correlation coefficient between the hydrological variable simulation results and the remote sensing monitoring data; the spatial performance index SPAEF ranges from negative infinity to 1, with 1 representing the best performance; B represents the coefficient of variation of the distribution of the hydrological variable simulation results and the proportion of the distribution of the remote sensing monitoring data; and Γ represents the overlap coefficient of the distribution bar chart of the hydrological variable simulation results and the remote sensing monitoring data.

[0152] Taking the Yellow River source basin as an example, the Pearson correlation coefficient between the evapotranspiration data simulated by the runoff forecasting model and the evapotranspiration data based on GLDAS products was 0.853 in 69 sub-basins, and the median RMSE was 0.767 mm / d. The spatial distribution of the evapotranspiration data simulated by the runoff forecasting model was largely consistent with the distribution of evapotranspiration data in remote sensing monitoring data, with a spatial performance index (SAPEF) of 0.56. Therefore, the runoff forecasting model has high accuracy.

[0153] By using remote sensing monitoring data at the sub-basin scale to assist in verifying the forecast accuracy of runoff forecasting models, the problem of poor reliability of runoff forecasting models due to the lack of real and reliable runoff monitoring data can be further reduced.

[0154] In this embodiment, by dividing the target area into watersheds, multiple sub-watersheds can be identified, and the average meteorological driving data, static watershed attribute data, and river segment attribute data corresponding to each sub-watershed can be determined. The average meteorological driving data represents the meteorological data of the sub-watershed. Based on a runoff forecasting model coupled with deep learning and eco-hydrological processes, the runoff forecasting result for each sub-watershed can be determined according to the average meteorological driving data, static watershed attribute data, and river segment attribute data. Based on the runoff monitoring data of the first hydrological station within the target area and the runoff forecasting results corresponding to the sub-watershed where each first hydrological station is located, the forecast corresponding to the runoff forecasting model can be determined. The loss is defined as follows: the first hydrological station is a pre-set hydrological station within the target area used for model training and accuracy verification of the runoff forecasting model in the time dimension; based on the forecast loss, a runoff forecasting model can be trained, wherein the trained runoff forecasting model is used to forecast runoff for any sub-basin of the target area, realizing the combination of deep learning networks and physical process generalization, determining a runoff forecasting model with the advantages of both data-driven and process-driven models, so as to realize runoff forecasting for any sub-basin within the target area with limited data, with high accuracy and reliability, and can provide decision support for regional water resources management and flood disaster prevention.

[0155] This disclosure also provides a method for predicting runoff. Figure 4A flowchart illustrating a runoff forecasting method according to an embodiment of this disclosure is shown. This runoff forecasting method can be executed by an electronic device such as a terminal device or a server. The terminal device can be a user equipment (UE), mobile device, user terminal, terminal, cellular phone, cordless phone, personal digital assistant (PDA), handheld device, computing device, vehicle-mounted device, wearable device, etc. The runoff forecasting method can be implemented by a processor calling computer-readable instructions stored in memory. Alternatively, the runoff forecasting method can be executed by a server. Figure 4 As shown, this runoff forecasting method includes:

[0156] In step S41, the target area is divided into watersheds to determine multiple sub-watersheds.

[0157] The target area here refers to the river basin area that needs to be forecasted for runoff, such as the Yellow River source basin (the basin area above the Tangnaihai hydrological station on the upper reaches of the Yellow River in the northeastern part of the Qinghai-Tibet Plateau). It can be flexibly set according to actual needs, and this disclosure does not make specific limitations on it.

[0158] The target area is typically large, with its river network consisting of a main stream and multiple tributaries. To achieve refined spatiotemporal runoff forecasting for the target area and accurate prediction of key cross-sections of the main stream and tributaries within the river network, the target area can be divided into watersheds, identifying multiple sub-watersheds. A sub-watershed refers to the catchment area above a specific cross-section of the main stream or tributary, with runoff converging between adjacent upstream and downstream sub-watersheds via the river channel. Specific methods for dividing the target area into sub-watersheds can be found in the methods described above and will not be elaborated upon here.

[0159] In step S42, for any sub-basin, the meteorological driving data for that sub-basin during the forecast period are determined, wherein the meteorological driving data represents the meteorological data of that sub-basin.

[0160] For any sub-basin, meteorological data within that sub-basin can be determined within the forecast period and used as the corresponding meteorological driving data for that sub-basin. The forecast period can represent the time period during which future runoff forecasting is required; this disclosure does not specifically limit this period. The specific method for determining the meteorological data can refer to implementation methods in related technologies; this disclosure does not specifically limit this method. The specific content of the meteorological data can be flexibly set according to actual usage needs, such as temperature, precipitation, wind speed, atmospheric pressure, relative humidity, and potential evapotranspiration; this disclosure does not specifically limit this content.

[0161] In step S43, meteorological driving data is input into the runoff forecasting model to determine the runoff forecasting results for the sub-basin during the forecast period. The runoff forecasting model is trained using the method described above.

[0162] For any given sub-basin, by inputting the driving data corresponding to that sub-basin into the runoff forecasting model for the target area, the runoff forecast result for that sub-basin during the forecast period can be determined. The runoff forecasting model is trained using any of the runoff forecasting model training methods described above.

[0163] Before inputting the driving data into the runoff forecasting model, the driving data can be preprocessed. The specific content of the data preprocessing can be flexibly set according to actual usage requirements. For example, it may include downscaling, bias correction, and standardization. This disclosure does not impose specific limitations on this.

[0164] In this embodiment, by dividing the target area into watersheds, multiple sub-watersheds can be determined. For any given sub-watershed, meteorological driving data for that sub-watershed during the forecast period can be determined, whereby the meteorological driving data can represent the meteorological data of that sub-watershed. The meteorological driving data is input into the runoff forecasting model to determine the corresponding runoff forecast result for that sub-watershed during the forecast period. The runoff forecasting model is trained using the above method, which combines deep learning networks with physical process generalization, enabling runoff forecasting for any sub-watershed within a target area with limited data. This model has high accuracy and reliability and provides decision support for regional water resource management and flood disaster prevention.

[0165] It is understood that the various method embodiments mentioned above in this disclosure can be combined with each other to form combined embodiments without violating the principle and logic. Due to space limitations, this disclosure will not elaborate further. Those skilled in the art will understand that in the above methods of specific implementation, the specific execution order of each step should be determined by its function and possible internal logic.

[0166] In addition, this disclosure also provides a runoff forecasting model training device, a runoff forecasting device, an electronic device, a computer-readable storage medium, and a program. All of the above can be used to implement any of the runoff forecasting model training methods and / or runoff forecasting methods provided in this disclosure. The corresponding technical solutions and descriptions are described in the corresponding section of the method and will not be repeated here.

[0167] Figure 5 A block diagram of a runoff forecasting model training apparatus according to an embodiment of the present disclosure is shown. Figure 5 As shown, the device 500 includes:

[0168] The sub-basin division module 501 is used to divide the target area into watersheds and determine multiple sub-basins.

[0169] The sample data determination module 502 is used to determine the average meteorological driving data, static watershed attribute data and river segment attribute data corresponding to each sub-basin, wherein the average meteorological driving data represents the meteorological data of the sub-basin;

[0170] The runoff forecasting module 503 is used to determine the runoff forecasting results for each sub-basin based on the runoff forecasting model coupled with deep learning and eco-hydrological processes, according to the average meteorological driving data, static watershed attribute data and river segment attribute data corresponding to each sub-basin.

[0171] The loss determination module 504 is used to determine the forecast loss corresponding to the runoff forecast model based on the runoff monitoring data of the first hydrological station in the target area and the runoff forecast results corresponding to the sub-basin where each first hydrological station is located. The first hydrological station is a pre-set hydrological station in the target area used for model training and to verify the accuracy of the runoff forecast model in the time dimension.

[0172] The runoff forecasting model training module 505 is used to train the runoff forecasting model based on the forecasting loss. The trained runoff forecasting model is used to forecast runoff for any sub-basin of the target area.

[0173] In one possible implementation, the sub-basin division module 501 is specifically used to: determine the digital elevation model data corresponding to the target area, and the location of each hydrological station in the target area; and, based on the digital elevation model data and the location of each hydrological station, divide the target area into watersheds and extract the river network to determine multiple sub-basins and the river network corresponding to the target area.

[0174] In one possible implementation, the sample data determination module 502 is specifically used to: determine the original meteorological driving data corresponding to each sub-basin based on meteorological data within a preset time scale and a preset distance range of each sub-basin; and for any given sub-basin, calculate the average value of the original meteorological driving data corresponding to that sub-basin to determine the average meteorological driving data corresponding to that sub-basin.

[0175] In one possible implementation, the static watershed attribute data includes: climate attributes, vegetation attributes, topographic attributes, and soil attributes; the sample data determination module 502 is specifically used to: for any sub-watershed, determine the original climate data, original vegetation data, original topographic data, and original soil data within the sub-watershed; determine the average value of the original climate data within the sub-watershed as the climate attribute corresponding to the sub-watershed; determine the average value of the original vegetation data within the sub-watershed as the vegetation attribute corresponding to the sub-watershed; determine the average value, maximum value, minimum value, and standard deviation of the original topographic data within the sub-watershed as the topographic attribute corresponding to the sub-watershed; and determine the average value, maximum value, minimum value, and standard deviation of the original soil data within the sub-watershed as the soil attribute corresponding to the sub-watershed.

[0176] In one possible implementation, the sample data determination module 502 is specifically used to: for any sub-basin, determine the connecting river segment between the sub-basin and the adjacent sub-basin according to the river network; and determine the river segment attribute data corresponding to the connecting river segment of the sub-basin based on a preset buffer range.

[0177] In one possible implementation, the runoff forecasting module 503 is specifically used to: determine the original training dataset based on the average meteorological driving data, static watershed attribute data, and river segment attribute data corresponding to each sub-basin; perform data preprocessing on the original training dataset to determine the standardized training dataset, wherein the data preprocessing includes standardization processing and data sequence partitioning processing, and the standardized training dataset includes multiple time series data sequences; and determine the runoff forecasting result corresponding to each sub-basin based on the runoff forecasting model, according to the original training dataset and the standardized training dataset.

[0178] In one possible implementation, the runoff forecasting model includes: a deep learning module, a sub-basin runoff generation module, and a river confluence module; the runoff forecasting module 503 is specifically used to: input the standardized training dataset into the deep learning module to determine the physical parameters, wherein the physical parameters include dynamic parameters and static parameters, and the dynamic parameters are used to eliminate the uncertainty of the sub-basin runoff generation module and the river confluence module in generalizing the eco-hydrological process; input the original training dataset and physical parameters into the sub-basin runoff generation module and the river confluence module to determine the runoff forecasting results corresponding to each sub-basin.

[0179] In one possible implementation, the runoff forecasting module 503 is further configured to: replace a sub-module in the sub-basin runoff generation module and / or the river confluence module with a preset neural network sub-module, wherein the neural network sub-module is configured to reduce the generalization error of the sub-basin runoff generation module and / or the river confluence module in terms of eco-hydrological processes.

[0180] In one possible implementation, the loss determination module 504 is specifically used to: determine the runoff forecast results during the preheating period and the simulation period according to the preset preheating time; determine the runoff monitoring data of each first hydrological station in the target area during the simulation period, as well as the loss weight corresponding to each first hydrological station; and determine the forecast loss based on the runoff monitoring data of each first hydrological station during the simulation period, the runoff forecast results of the sub-basin where each first hydrological station is located during the simulation period, and the loss weight.

[0181] In one possible implementation, the runoff forecasting model training module 505 is further used to: divide the hydrological stations in the target area into a first hydrological station and a second hydrological station based on the leave-one-out method; for any second hydrological station, verify the accuracy of the runoff forecasting model in the spatial dimension according to the runoff forecasting results corresponding to the sub-basin where the second hydrological station is located, and determine the first verification result corresponding to the second hydrological station.

[0182] In one possible implementation, the runoff forecasting model training module 505 is further configured to: determine the remote sensing monitoring data corresponding to any given sub-basin; determine the hydrological variable simulation results corresponding to the sub-basin during the runoff forecasting process of the runoff forecasting model for that sub-basin; and verify the accuracy of the runoff forecasting model based on the remote sensing monitoring data and the hydrological variable simulation results, thereby determining the second verification result corresponding to the sub-basin.

[0183] Figure 6 A block diagram of a runoff forecasting device according to an embodiment of the present disclosure is shown. Figure 6 As shown, the device 600 includes:

[0184] The sub-basin division module 601 is used to divide the target area into watersheds and determine multiple sub-basins;

[0185] The driving data determination module 602 is used to determine the meteorological driving data of any sub-basin during the forecast period, wherein the meteorological driving data represents the meteorological data of the sub-basin;

[0186] The runoff forecast module 603 is used to input meteorological driving data into the runoff forecast model to determine the runoff forecast results for the sub-basin during the forecast period. The runoff forecast model is trained using the method described above.

[0187] In some embodiments, the functions or modules of the apparatus provided in this disclosure can be used to perform the methods described in the above method embodiments. The specific implementation can be referred to the description of the above method embodiments, and for the sake of brevity, it will not be repeated here.

[0188] This disclosure also proposes a computer-readable storage medium storing computer program instructions that, when executed by a processor, implement the above-described method. The computer-readable storage medium can be volatile or non-volatile.

[0189] This disclosure also proposes an electronic device, including: a processor; and a memory for storing processor-executable instructions; wherein the processor is configured to implement the above method when executing the instructions stored in the memory.

[0190] Electronic devices can be provided as terminals, servers, or other forms of devices.

[0191] Figure 7 A block diagram of an electronic device according to an embodiment of the present disclosure is shown. For example, electronic device 1900 may be provided as a server or terminal device. (Refer to...) Figure 7 The electronic device 1900 includes a processing component 1922, which further includes one or more processors, and memory resources represented by memory 1932 for storing instructions, such as application programs, that can be executed by the processing component 1922. The application programs stored in memory 1932 may include one or more modules, each corresponding to a set of instructions. Furthermore, the processing component 1922 is configured to execute instructions to perform the methods described above.

[0192] Device 1900 may also include a power supply component 1926 configured to perform power management of device 1900, a wired or wireless network interface 1950 configured to connect device 1900 to a network, and an input / output interface 1958 (I / O interface). Device 1900 can operate on an operating system, such as Windows Server, stored in memory 1932. TM macOS X TM Unix TM Linux TM FreeBSD TM Or similar.

[0193] In an exemplary embodiment, a non-volatile computer-readable storage medium is also provided, such as a memory 1932 including computer program instructions that can be executed by a processing component 1922 of the device 1900 to perform the above-described method.

[0194] This disclosure can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of this disclosure.

[0195] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example—but not limited to—electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination thereof. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0196] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0197] Computer program instructions used to perform the operations of this disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, status setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing the status information of the computer-readable program instructions to implement various aspects of this disclosure.

[0198] Various aspects of this disclosure are described herein with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this disclosure. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer-readable program instructions.

[0199] These computer-readable program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that, when executed by the processor of the computer or other programmable data processing apparatus, they create means for implementing the functions / actions specified in one or more blocks of the flowchart and / or block diagram. These computer-readable program instructions can also be stored in a computer-readable storage medium that causes a computer, programmable data processing apparatus, and / or other device to operate in a particular manner; thus, the computer-readable medium storing the instructions comprises an article of manufacture that includes instructions for implementing aspects of the functions / actions specified in one or more blocks of the flowchart and / or block diagram.

[0200] Computer-readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, thereby causing the instructions executed on the computer, other programmable data processing apparatus, or other device to perform the functions / actions specified in one or more boxes of a flowchart and / or block diagram.

[0201] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of an instruction containing one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions marked in the blocks may occur in a different order than those shown in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0202] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or technical improvements to the embodiments in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A method for training a runoff forecasting model, characterized in that, include: The target area is divided into watersheds, and multiple sub-watersheds are identified. Determine the average meteorological driving data, static watershed attribute data, and river segment attribute data corresponding to each sub-basin, wherein the average meteorological driving data represents the meteorological data of the sub-basin; Based on a runoff forecasting model that couples deep learning and eco-hydrological processes, the runoff forecasting results for each sub-basin are determined according to the average meteorological driving data, static watershed attribute data and river segment attribute data corresponding to each sub-basin. Based on the runoff monitoring data of the first hydrological station in the target area and the runoff forecast results corresponding to each sub-basin of the first hydrological station, the forecast loss corresponding to the runoff forecast model is determined. The first hydrological station is a hydrological station in the target area that is preset for model training and for verifying the accuracy of the runoff forecast model in the time dimension. Based on the forecast loss, the runoff forecast model is trained, wherein the trained runoff forecast model is used to forecast runoff for any sub-basin of the target area; The runoff forecasting model is used to determine the runoff forecasting results for each sub-basin based on the average meteorological driving data, static watershed attribute data, and river segment attribute data corresponding to each sub-basin, combined with a data-driven deep learning process and a physical process-driven eco-hydrological model. The model includes a deep learning module, a sub-basin runoff generation module, and a river confluence module. The runoff forecasting model based on deep learning and eco-hydrological processes determines the runoff forecast results for each sub-basin according to the average meteorological driving data, static watershed attribute data, and river segment attribute data corresponding to each sub-basin, including: The original training dataset is determined based on the average meteorological driving data, static watershed attribute data, and river segment attribute data corresponding to each sub-basin; The original training dataset is preprocessed to determine the standardized training dataset. The data preprocessing includes standardization and data sequence partitioning. The standardized training dataset includes multiple time-series data sequences. The standardized training dataset is input into the deep learning module to determine the physical parameters of the sub-basin runoff generation module and the river confluence module. The physical parameters include dynamic parameters and static parameters. The dynamic parameters are used to eliminate the uncertainty of the sub-basin runoff generation module and the river confluence module in generalizing eco-hydrological processes. The original training dataset and the physical parameters are input into the sub-basin runoff generation module and the river confluence module to determine the runoff forecast results for each sub-basin.

2. The method according to claim 1, characterized in that, The process of dividing the target area into watersheds and identifying multiple sub-watersheds includes: Determine the digital elevation model data corresponding to the target area, and the location of each hydrological station within the target area; Based on the digital elevation model data and the location of each hydrological station, the target area is divided into watersheds and river networks are extracted to determine the multiple sub-watersheds and the corresponding river network of the target area.

3. The method according to claim 1 or 2, characterized in that, The determination of the average meteorological driving data corresponding to each of the sub-basins includes: Based on meteorological data within a preset time scale for each sub-basin and within a preset distance range for each sub-basin, determine the original meteorological driving data corresponding to each sub-basin; For any given sub-basin, the average value of the original meteorological driving data corresponding to that sub-basin is calculated to determine the average meteorological driving data corresponding to that sub-basin.

4. The method according to claim 1 or 2, characterized in that, The static watershed attribute data includes: climate attributes, vegetation attributes, topographic attributes, and soil attributes; The step of determining the static watershed attribute data corresponding to each of the sub-watersheds includes: For any one of the sub-basins, determine the original climate data, original vegetation data, original topographic data, and original soil data within the sub-basin area; The average value of the original climate data within the sub-basin is determined as the climate attribute corresponding to the sub-basin; The average value of the original vegetation data within the sub-basin is determined as the vegetation attribute corresponding to the sub-basin; The average, maximum, minimum, and standard deviation of the original topographic data within the sub-basin are determined as the topographic attributes corresponding to the sub-basin. The average, maximum, minimum, and standard deviation of the original soil data within the sub-watershed are determined as the soil attributes corresponding to the sub-watershed.

5. The method according to claim 2, characterized in that, The determination of the river segment attribute data corresponding to each of the sub-basins includes: For any one of the sub-basins, the connecting river segments between the sub-basin and adjacent sub-basins are determined based on the river network; Based on a preset buffer range, the river segment attribute data corresponding to the connecting river segment of the sub-basin is determined.

6. The method according to claim 1, characterized in that, The method further includes: A pre-defined neural network submodule replaces a submodule in the sub-basin runoff generation module and / or the river confluence module, wherein the neural network submodule is used to reduce the generalization error of the eco-hydrological process by the sub-basin runoff generation module and / or the river confluence module.

7. The method according to claim 1 or 2, characterized in that, The step of determining the forecast loss corresponding to the runoff forecasting model based on the runoff monitoring data of the first hydrological station within the target area and the runoff forecasting results corresponding to the sub-basin where each of the first hydrological stations is located includes: Based on the preset preheating time, the runoff forecast results for the preheating period and the simulation period are determined respectively; Determine the runoff monitoring data of each first hydrological station within the target area during the simulation period, and the loss weight corresponding to each first hydrological station; The predicted loss is determined based on the runoff monitoring data of each first hydrological station during the simulation period, the runoff forecast results of each sub-basin where the first hydrological station is located during the simulation period, and the loss weight.

8. The method according to claim 1 or 2, characterized in that, The method further includes: Based on the leave-one method, the hydrological stations within the target area are divided into the first hydrological station and the second hydrological station; For any second hydrological station, the accuracy of the runoff forecast model is verified in the spatial dimension based on the runoff forecast results corresponding to the sub-basin where the second hydrological station is located, and the first verification result corresponding to the second hydrological station is determined.

9. The method according to claim 1 or 2, characterized in that, The method further includes: For any one of the sub-basins, determine the remote sensing monitoring data corresponding to that sub-basin; During the runoff forecasting process of the runoff forecasting model for this sub-basin, the simulation results of the corresponding hydrological variables for this sub-basin are determined; Based on the remote sensing monitoring data and the hydrological variable simulation results, the accuracy of the runoff forecast model is verified, and the second verification result corresponding to the sub-basin is determined.

10. A runoff forecasting method, characterized in that, include: The target area is divided into watersheds, and multiple sub-watersheds are identified. For any one of the sub-basins, determine the meteorological driving data for that sub-basin during the forecast period, wherein the meteorological driving data represents the meteorological data for that sub-basin; The meteorological driving data is input into the runoff forecasting model to determine the runoff forecasting result for the sub-basin during the forecast period, wherein the runoff forecasting model is trained by the method described in any one of claims 1 to 9.

11. A runoff forecasting model training device, characterized in that, include: The sub-basin delineation module is used to divide the target area into watersheds and determine multiple sub-basins; The sample data determination module is used to determine the average meteorological driving data, static watershed attribute data, and river segment attribute data corresponding to each sub-basin, wherein the average meteorological driving data represents the meteorological data of the sub-basin; The runoff forecasting module is used to determine the runoff forecasting results for each sub-basin based on the runoff forecasting model coupled with deep learning and eco-hydrological processes, according to the average meteorological driving data, static watershed attribute data and river segment attribute data corresponding to each sub-basin. The loss determination module is used to determine the forecast loss corresponding to the runoff forecasting model based on the runoff monitoring data of the first hydrological station in the target area and the runoff forecasting results corresponding to the sub-basin where each first hydrological station is located. The first hydrological station is a hydrological station in the target area that is preset for model training and for verifying the accuracy of the runoff forecasting model in the time dimension. The runoff forecasting model training module is used to train the runoff forecasting model based on the forecasting loss, wherein the trained runoff forecasting model is used to forecast runoff for any sub-basin of the target area; The runoff forecasting model is used to determine the runoff forecasting results for each sub-basin based on the average meteorological driving data, static watershed attribute data, and river segment attribute data corresponding to each sub-basin, combined with a data-driven deep learning process and a physical process-driven eco-hydrological model. The model includes a deep learning module, a sub-basin runoff generation module, and a river confluence module. The runoff forecasting module is also used for: The original training dataset is determined based on the average meteorological driving data, static watershed attribute data, and river segment attribute data corresponding to each sub-basin; The original training dataset is preprocessed to determine the standardized training dataset. The data preprocessing includes standardization and data sequence partitioning. The standardized training dataset includes multiple time-series data sequences. The standardized training dataset is input into the deep learning module to determine the physical parameters of the sub-basin runoff generation module and the river confluence module. The physical parameters include dynamic parameters and static parameters. The dynamic parameters are used to eliminate the uncertainty of the sub-basin runoff generation module and the river confluence module in generalizing eco-hydrological processes. The original training dataset and the physical parameters are input into the sub-basin runoff generation module and the river confluence module to determine the runoff forecast results for each sub-basin.

12. A runoff forecasting device, characterized in that, include: The sub-basin delineation module is used to divide the target area into watersheds and determine multiple sub-basins; The driving data determination module is used to determine the meteorological driving data of any one of the sub-basins during the forecast period, wherein the meteorological driving data represents the meteorological data of the sub-basin; The runoff forecasting module is used to input the meteorological driving data into the runoff forecasting model to determine the runoff forecasting result corresponding to the sub-basin during the forecast period, wherein the runoff forecasting model is trained by the method described in any one of claims 1 to 9.

13. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to implement the method of any one of claims 1 to 10 when executing instructions stored in the memory.

14. A non-volatile computer-readable storage medium storing computer program instructions thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 10.

Images