Large-scale basin deep learning flood forecasting method based on runoff lag information
By dividing the watershed into sub-watersheds on a large scale and using convolutional neural networks and deep learning models to extract runoff lag information, the problems of complex and time-consuming model construction and scarce data in existing technologies have been solved, and more accurate flood forecasting has been achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- YUNNAN UNIV
- Filing Date
- 2023-01-29
- Publication Date
- 2026-07-24
AI Technical Summary
In large-scale watersheds, existing technologies struggle to effectively utilize runoff lag information for flood forecasting, especially under conditions of data scarcity. Existing models are complex and time-consuming to construct and neglect the complexity of upstream and downstream river confluence paths and flow lag relationships.
By collecting watershed data, dividing it into sub-watersheds, extracting runoff lag information using convolutional neural networks, and combining it with a deep learning long short-term memory model, an optimal hydrological forecasting model is constructed to extract runoff process lag information between upstream and downstream watersheds for accurate simulation and forecasting.
It improves the accuracy and reliability of large-scale watershed flood forecasting, effectively compensates for the lack of model construction in areas with missing data, and provides key feature information for flood early warning.
Smart Images
Figure CN116205136B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of hydrological simulation and flood forecasting technology using machine learning, and particularly to a large-scale watershed deep learning flood forecasting method based on runoff lag information. Background Technology
[0002] Large-scale watershed hydrological and meteorological data are constrained by topography, regional economic conditions, and other factors, resulting in a scarcity of measured data. The limited data available is insufficient to support and construct watershed hydrological simulations, particularly for effective flood forecasting under future conditions. Against the backdrop of climate change, my country has experienced frequent floods in recent years, causing severe economic losses and casualties. Effective runoff simulation and flood forecasting technologies will provide crucial scientific and technological support for regional water resource utilization, ecosystem protection, agricultural development, and urban planning.
[0003] Hydrological models are classic methods for runoff simulation and flood forecasting. Currently, the main methods include process-based physical mechanism models and data-driven models. Process-based physical mechanism models, such as VIC (Variable Infiltration Capacity) and SWAT (Soil and Water Assessment Tool), are based on the physical theories of the water cycle and a large amount of measured data, making them highly practical. However, these models are largely created using inductive reasoning. While they provide reliable explanations of hydrological processes, the existing theories are incomplete compared to the complex processes that actually occur, and the models themselves are not comprehensive. Furthermore, they involve a large number of complex parameters that require repeated calibration, consuming more time to build regional models. In recent years, machine learning has developed rapidly, and a series of advanced artificial neural network algorithms have been developed. Deep learning, as the foundation of artificial intelligence, has been applied to the field of hydrology, and many studies have explored its technical aspects, demonstrating its broad application prospects.
[0004] Data-driven hydrological models based on machine learning often employ convolutional neural networks (CNNs) and long short-term memory (LSTM) networks. Recurrent convolutional neural networks can efficiently capture and describe dynamic temporal behavior, while LSTMs introduce a "gating" mechanism to control the transmission of feature information from time series, greatly improving the ability to process information features from long time series and the simulation performance of the model.
[0005] In the absence of large-scale watershed data, data-driven deep learning hydrological models can effectively extract the characteristics of the underlying surface's influence on hydrological processes. Their structure differs from the construction mode of traditional hydrological models. They establish the relationship between the objective function and the data from a purely data perspective, which can maximize the accuracy of runoff simulation and flood forecasting in the context of hydrological and meteorological big data.
[0006] Furthermore, flood forecasting often relies on historical information to predict and warn of flood events on an hourly scale, neglecting the daily-scale runoff lag information existing between upstream and downstream sections in large-scale watersheds. Due to the complex river confluence paths and the flow lag relationship between upstream and downstream sections, large-scale watersheds offer an effective opportunity to extract runoff lag information and construct models that accurately simulate and forecast the outlet runoff and flood processes of large-scale watersheds using upstream information. This is of great significance for flood early warning in large-scale watersheds and has enormous potential for application in various watersheds.
[0007] Therefore, proposing a large-scale watershed deep learning flood forecasting method based on runoff lag information to address the difficulties of existing technologies is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0008] In view of this, the present invention provides a large-scale watershed deep learning flood forecasting method based on runoff lag information. It can extract the lag information of runoff processes between upstream and downstream watersheds on a large scale, construct a hydrological model based on deep learning, and provide a large-scale watershed deep learning flood forecasting method based on runoff lag information.
[0009] To achieve the above objectives, the present invention adopts the following technical solution: A large-scale watershed deep learning flood forecasting method based on runoff lag information includes the following steps: S1. Collect information and data from hydrological observation stations in the basin, basin characteristic data, historical hydrological data, and meteorological forecast characteristic data; S2. Based on the information data of the watershed hydrological observation stations and the watershed characteristic data collected in S1, the large-scale watershed is divided into several sub-watersheds; S3. Based on the sub-basins divided in S2, and according to the historical hydrological data and meteorological forecast characteristics collected in S1, the data is preprocessed to obtain runoff lag information. S4. Use a convolutional neural network model to extract the runoff lag information obtained in S3 and construct a runoff lag information database; S5. A deep learning long short-term memory model is used to couple the runoff lag information database in S4 to obtain runoff lag information models for several sub-basins; S6. Based on the model accuracy evaluation index, the optimal runoff lag information parameters of several sub-basins are obtained, and the optimal hydrological forecasting model is obtained. S7. Based on the optimal hydrological forecasting model in S6, achieve accurate flood forecasting for large-scale watersheds.
[0010] The above methods may include, but are not limited to, the following information data and watershed characteristic data in S1: latitude and longitude of the hydrological observation station, altitude, average watershed area, average altitude, average slope, average annual precipitation, average precipitation amount, average evapotranspiration, aridity index, soil moisture, leaf area index, dominant land use, frequency of high precipitation, frequency of low precipitation, soil porosity, and maximum water content. Meteorological forecast data include, but are not limited to: precipitation, maximum temperature, minimum temperature, and evapotranspiration.
[0011] Optionally, the division of the S2 sub-basins can be based on digital elevation models and river network data, using GIS spatial analysis technology combined with the spatial distribution of the above features.
[0012] Optionally, the data preprocessing in S3 includes the following: The flow data is processed into variable values by incorporating average catchment area and average precipitation into the calculation:
[0013] in, The variables obtained after processing For the input traffic data, The average area of the watershed. The average flow rate of the basin; Traffic flow and rainfall data are transformed using the following formula to make the data more closely approximate a Gaussian distribution:
[0014] in, The target variable after transformation has a Gamma distribution. The target variable before transformation; Standardize all input variables:
[0015] in, For the transformed variables, For the input variables, The average value of the variable. is the standard deviation of the variable.
[0016] Optionally, the runoff lag information obtained in S3 in the above method includes the spatial information of the flood peak lag and the temporal information of the flow lag. The specific content of obtaining the spatial information on flood peak time delay between sub-basins is as follows: The peak flow and corresponding dates of all sub-basins and the total outlet of the basin are obtained by using the annual maximum value method, and the difference in the dates of flood peaks between sub-basins is calculated. The specific details of obtaining the time information on flow delay between sub-basins are as follows: Based on historical runoff data for each sub-basin, the relative flow ratio is calculated using runoff characteristic sequences: (4) in, The average discharge sequence at the watershed outlet. These are flow sequences for different sub-basins.
[0017] The above method, optionally, includes the following specific content of the deep learning long short-term memory model in S5: The training data consisted of meteorological forecast data, hydrological observation station information data, basin characteristic data, and runoff lag information data from the upstream sub-basin. The training objective was to simulate the runoff and flood processes in the downstream sub-basin. The validation period meteorological and basin characteristic data, as well as different runoff lag information data, were input to simulate the downstream runoff and flood processes.
[0018] The above method, optionally, includes the following specific content for the model accuracy evaluation index in S6: NSE The calculation formula is as follows: (5) Deviation coefficient PBIAS The calculation formula is as follows: (6) in, For actual measured runoff, To simulate runoff, This represents the average value of the measured runoff.
[0019] The above method, optionally, includes the following specific details regarding the optimal runoff lag information parameters for several sub-basins in S6: Based on the model accuracy evaluation index, the model with the best index, along with meteorological forecast characteristic data and runoff lag information data, is selected as the optimal model for simulating the total basin outlet runoff and flood process in a specific sub-basin.
[0020] As can be seen from the above technical solution, compared with the prior art, the present invention provides a large-scale watershed deep learning flood forecasting method based on runoff lag information: 1) This invention utilizes machine learning to extract spatial information of flood peak time lag and temporal information of flow lag in the upstream and downstream of a large-scale watershed. It effectively characterizes the time lag information of runoff processes between upstream and downstream watersheds, providing key feature information for accurate simulation and forecasting of flood processes at the outlet of the total watershed based on upstream watershed data.
[0021] 2) This invention constructs an optimal hydrological forecasting model by extracting runoff lag information from upstream and downstream basins, thereby improving the model's ability to accurately simulate flood processes at the outlet of the total basin by relying solely on meteorological and hydrological data from the upstream basin.
[0022] 3) This invention effectively improves the ability to extract runoff characteristics and the accuracy and reliability of flood forecasting in large-scale watersheds; it provides an effective modeling technology for flood forecasting in areas lacking measured data in large-scale watersheds, making up for the shortcomings of data limitations in areas with little or no data that prevent the construction of regional hydrological models, and has great potential application value. Attached Figure Description
[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.
[0024] Figure 1 Flowchart of the large-scale watershed deep learning flood forecasting method based on runoff lag information provided by this invention; Figure 2 This is a schematic diagram of the watershed provided by the present invention; Figure 3 The hysteresis information map extracted by CNN provided by this invention; Figure 4 This is a process diagram illustrating the simulation of the total outlet flood of the basin at the upstream station, as provided in an embodiment of the present invention. Detailed Implementation
[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0026] Reference Figure 1 As shown, the present invention provides a large-scale watershed deep learning flood forecasting method based on runoff lag information, comprising the following steps: S1. Collect information and data from hydrological observation stations in the basin, basin characteristic data, historical hydrological data, and meteorological forecast characteristic data; S2. Based on the information data of the watershed hydrological observation stations and the watershed characteristic data collected in S1, the large-scale watershed is divided into several sub-watersheds; S3. Based on the sub-basins divided in S2, and according to the historical hydrological data and meteorological forecast characteristics collected in S1, the data is preprocessed to obtain runoff lag information. S4. Use a convolutional neural network model to extract the runoff lag information obtained in S3 and construct a runoff lag information database; S5. A deep learning long short-term memory model is used to couple the runoff lag information database in S4 to obtain runoff lag information models for several sub-basins; S6. Based on the model accuracy evaluation index, the optimal runoff lag information parameters of several sub-basins are obtained, and the optimal hydrological forecasting model is obtained. S7. Based on the optimal hydrological forecasting model in S6, achieve accurate flood forecasting for large-scale watersheds.
[0027] Furthermore, the information data of the hydrological observation stations in the S1 basin and the basin characteristic data include, but are not limited to: the latitude and longitude of the hydrological observation stations, altitude, average basin area, average altitude, average slope, average annual precipitation, average precipitation amount, average evapotranspiration, aridity index, soil moisture, leaf area index, dominant land use, frequency of high precipitation, frequency of low precipitation, soil porosity, and maximum water content. Furthermore, the meteorological forecast features in S1 include, but are not limited to: precipitation, maximum temperature, minimum temperature, and evapotranspiration.
[0028] Furthermore, the division of the S2 neutron basin is based on digital elevation model and river network data, and GIS spatial analysis technology is used to combine the spatial distribution of the above characteristics for division.
[0029] Furthermore, data preprocessing in S3 specifically includes: The flow data is processed into variable values by incorporating average catchment area and average precipitation into the calculation:
[0030] in, The variables obtained after processing For the input traffic data, The average area of the watershed. The average flow rate of the basin; Traffic flow and rainfall data are transformed using the following formula to make the data more closely approximate a Gaussian distribution:
[0031] in, The target variable after transformation has a Gamma distribution. The target variable before transformation; Standardize all input variables:
[0032] in, For the transformed variables, For the input variables, The average value of the variable. is the standard deviation of the variable.
[0033] Furthermore, the runoff lag information obtained in S3 includes spatial information of peak flow lag and temporal information of flow lag; The specific content of obtaining the spatial information of flood peak time lag between sub-basins is as follows: obtain the peak flow and corresponding date of all sub-basins and the total outlet of the basin through the annual maximum value method, and calculate the difference of the flood peak occurrence date between sub-basins; The specific content for obtaining the time lag information of flow between sub-basins is as follows: Based on the historical runoff data of each sub-basin, the relative flow ratio is calculated through runoff characteristic sequences. (4) in, The average discharge sequence at the watershed outlet. These are flow sequences for different sub-basins.
[0034] Furthermore, the specific content of the deep learning long short-term memory model in S5 is as follows: The training data consisted of meteorological forecast data, hydrological observation station information data, basin characteristic data, and runoff lag information data from the upstream sub-basin. The training objective was to simulate the runoff and flood processes in the downstream sub-basin. The validation period meteorological and basin characteristic data, as well as different runoff lag information data, were input to simulate the downstream runoff and flood processes.
[0035] Furthermore, the specific content of the model accuracy evaluation index in S6 is as follows: NSE The calculation formula is as follows: (5) Deviation coefficient PBIAS The calculation formula is as follows: (6) in, For actual measured runoff, To simulate runoff, This represents the average value of the measured runoff.
[0036] Furthermore, the specific contents of the optimal runoff lag information parameters for several sub-basins in S6 are as follows: Based on the model accuracy evaluation index, the model with the best index, along with meteorological forecast characteristic data and runoff lag information data, is selected as the optimal model for simulating the total basin outlet runoff and flood process in a specific sub-basin.
[0037] In one specific embodiment, the Dulongjiang-Irrawaddy River basin is used as a research example, and the method flow is as follows: Figure 1 As shown.
[0038] The Dulong River-Inovadi River Basin (Du-I Basin) covers an area of 413,700 km². 2 Except for the northern mountainous region, which falls within subtropical and temperate climate zones, most of the area is in a tropical climate zone. The Dui River basin has an annual runoff of approximately 1000 mm, considered relatively higher than other major rivers globally. The Dui River basin experiences its heaviest rainfall from mid-May to October, during which flooding events are frequent.
[0039] This example uses deep learning to extract runoff lag information, then constructs a large-scale watershed hydrological model, and optimizes the model for forecasting watershed outlet floods to achieve accurate forecasting of watershed outlet flood events. The specific steps are as follows: Information from hydrological observation stations in the Duyi River Basin was collected. Based on the basin's digital elevation model and river information, GIS spatial analysis was used to divide the basin into several sub-basins, such as... Figure 2 As shown; Historical hydrological data and meteorological forecast characteristics of the watershed were compiled, including latitude and longitude, altitude, watershed area, average altitude, average slope, average annual precipitation, average precipitation amount, average evapotranspiration, aridity index, soil moisture, leaf area index, dominant land use, frequency of high precipitation, frequency of low precipitation, soil porosity, and maximum water content; as well as precipitation, maximum temperature, minimum temperature, and evapotranspiration. The data underwent preprocessing, primarily focusing on flow and precipitation data: The method for flow preprocessing involves using the average catchment area and average precipitation in the calculation to process the flow data into variable values.
[0040] in, The variables obtained after processing For the input traffic data, The average area of the watershed. The average flow rate of the basin; Traffic flow and rainfall data are transformed using the following formula to make the data more closely approximate a Gaussian distribution:
[0041] in, The target variable after transformation has a Gamma distribution. The target variable before transformation; Standardize all input variables:
[0042] in, For the transformed variables, For the input variables, The average value of the variable. is the standard deviation of the variable.
[0043] The annual maximum value method was used to analyze the flood peak value in the hydrological process of the sub-basin, and the spatial information of the flood peak time lag and the temporal information of the flow lag between the upstream and downstream sub-basins were calculated. Based on the annual division, the peak flow and corresponding dates of all sub-basins and the total outlet of the basin are obtained, and the difference in the peak flood occurrence dates between upstream and downstream sub-basins is calculated to provide the model with spatial information on peak flood time lag.
[0044] The specific content for obtaining the time lag information of flow between sub-basins is as follows: Based on the historical runoff data of each sub-basin, the relative flow ratio is calculated through runoff characteristic sequences. (4) in, The average discharge sequence at the watershed outlet. These are flow sequences for different sub-basins.
[0045] The spatial information of flood peak time lag and the temporal information of flow lag between upstream and downstream sub-basins are extracted using a CNN to construct a runoff lag information database. The CNN includes convolutional layers, pooling layers, and fully connected layers, such as... Figure 3 As shown; A deep learning model coupled with a runoff lag information database is used. The deep learning model employs a Long Short-Term Memory (LSTM) model. The model input consists of preprocessed data from all sub-basins, divided into training and validation data. The model's training data includes meteorological forecast characteristics of the upstream sub-basin, hydrological observation station information, basin characteristics, and runoff lag information data. The training objective is to simulate the runoff and flood processes of the downstream sub-basin. The validation period is used to simulate downstream runoff and flood processes by inputting meteorological and basin characteristic data, as well as different runoff lag information data.
[0046] Based on the model accuracy evaluation index, the optimal runoff lag information parameters for each sub-basin are selected to obtain the corresponding hydrological forecasting model. The model parameters were selected based on optimal principles. Each model was trained 10 times, and the average value was calculated to evaluate the model's simulation performance. Model accuracy evaluation metrics included: Nash-Sutcliffe efficiency (NSE) The calculation formula is as follows: (5) Deviation coefficient PBIAS The calculation formula is as follows: (6) in For actual measured runoff, To simulate runoff, This represents the average value of the measured runoff.
[0047] The evaluation metrics for the model validation period of each sub-basin are shown in Table 1 below: Table 1
[0048] Based on the simulation results of the sub-basin model, the model with the best performance index, along with meteorological forecast characteristic data and runoff lag information data, is selected as the optimal model for simulating the total basin outlet runoff and flood process in a specific sub-basin.
[0049] Based on the model evaluation index results in the table above, the models for each sub-basin performed excellently. NSE All values are greater than 0.8. The optimal model constructed from "Sub-basin 01" is selected. Based on this, the meteorological characteristic factors of the sub-basin and the basin characteristic factors are input into the model to obtain the runoff and flood processes at the outlet section of the total basin, simulating the flow process. The forecast results accurately simulate the peak flow during the flow process. The simulation results are as follows: Figure 4 As shown.
[0050] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A large-scale watershed deep learning flood forecasting method based on runoff lag information, characterized in that, Includes the following steps: S1. Collect information and data from hydrological observation stations in the basin, basin characteristic data, historical hydrological data, and meteorological forecast characteristic data; S2. Based on the information data of the watershed hydrological observation stations and the watershed characteristic data collected in S1, the large-scale watershed is divided into several sub-watersheds; S3. Based on the sub-basins divided in S2, and according to the historical hydrological data and meteorological forecast characteristics collected in S1, the data is preprocessed to obtain runoff lag information. S4. Use a convolutional neural network model to extract the runoff lag information obtained in S3 and construct a runoff lag information database; S5. A deep learning long short-term memory model is used to couple the runoff lag information database in S4 to obtain runoff lag information models for several sub-basins; S6. Based on the model accuracy evaluation index, the optimal runoff lag information parameters of several sub-basins are obtained, and the optimal hydrological forecasting model is obtained. S7. Based on the optimal hydrological forecasting model in S6, achieve accurate flood forecasting for large-scale watersheds; The runoff lag information obtained in S3 includes spatial information of peak flow lag and temporal information of flow lag; The specific content of obtaining the spatial information on flood peak time delay between sub-basins is as follows: The peak flow and corresponding dates of all sub-basins and the total outlet of the basin are obtained by using the annual maximum value method, and the difference in the dates of flood peaks between sub-basins is calculated. The specific details of obtaining the time information on flow delay between sub-basins are as follows: Based on historical runoff data for each sub-basin, the relative flow ratio is calculated using runoff characteristic sequences: （4） in, The average discharge sequence at the watershed outlet. These are flow sequences for different sub-basins.
2. The large-scale watershed deep learning flood forecasting method based on runoff lag information according to claim 1, characterized in that, Information data and watershed characteristic data of hydrological observation stations in the S1 basin include: latitude and longitude of hydrological observation stations, altitude, average basin area, average altitude, average slope, average annual precipitation, average precipitation amount, average evapotranspiration, aridity index, soil moisture, leaf area index, dominant land use, frequency of high precipitation, frequency of low precipitation, soil porosity, and maximum water content. Meteorological forecast data include: precipitation, maximum temperature, minimum temperature, and evapotranspiration.
3. The large-scale watershed deep learning flood forecasting method based on runoff lag information according to claim 1, characterized in that, The division of the S2 neutron basin is based on digital elevation model and river network data. It uses GIS spatial analysis technology to combine the spatial distribution of watershed hydrological observation station information, watershed characteristic data and meteorological forecast characteristic data collected in S1.
4. The large-scale watershed deep learning flood forecasting method based on runoff lag information according to claim 1, characterized in that, Data preprocessing in S3 specifically includes: The flow data is processed into variable values by incorporating average catchment area and average precipitation into the calculation: in, The variables obtained after processing For the input traffic data, The average area of the watershed. The average flow rate of the basin; Traffic flow and rainfall data are transformed using the following formula to make the data more closely approximate a Gaussian distribution: in, The target variable after transformation has a Gamma distribution. The target variable before transformation; Standardize all input variables: in, For the transformed variables, For the input variables, The average value of the variable. is the standard deviation of the variable.
5. The large-scale watershed deep learning flood forecasting method based on runoff lag information according to claim 1, characterized in that, The specific content of the deep learning long short-term memory model in S5 is as follows: The training data consisted of meteorological forecast data, hydrological observation station information data, basin characteristic data, and runoff lag information data from the upstream sub-basin. The training objective was to simulate the runoff and flood processes in the downstream sub-basin. The validation period meteorological and basin characteristic data, as well as different runoff lag information data, were input to simulate the downstream runoff and flood processes.
6. The large-scale watershed deep learning flood forecasting method based on runoff lag information according to claim 1, characterized in that, The specific content of the model accuracy evaluation index in S6 is as follows: NSE The calculation formula is as follows: （5） Deviation coefficient PBIAS The calculation formula is as follows: （6） in, For actual measured runoff, To simulate runoff, This represents the average value of the measured runoff.
7. The large-scale watershed deep learning flood forecasting method based on runoff lag information according to claim 1, characterized in that, The specific contents of the optimal runoff lag information parameters for several sub-basins in S6 are as follows: Based on the model accuracy evaluation index, the model with the best index, along with meteorological forecast characteristic data and runoff lag information data, is selected as the optimal model for simulating the total basin outlet runoff and flood process in the corresponding sub-basin.