Intelligent simulation method for runoff process in area lacking data

Abstract

The invention discloses an intelligent simulation method for a runoff process in a region lacking data, and belongs to the technical field of hydrological simulation. The method comprises the following steps: step 1, collecting basic data; step 2, constructing an SWAT hydrological model and outputting physical variables; step 3, an SWAT-LSTM-Transform three-level coupling framework is constructed, and the SWAT-LSTM-Transform three-level coupling framework 4, carrying out transfer learning on the SWAT-LSTM-Transform coupling model, and carrying out transfer learning on the SWAT-LSTM-Transform coupling model; and step 5, runoff simulation and performance evaluation. According to the method, the Transform model and the SWAT-LSTM coupling model are creatively linked together, the adaptability and reliability under the condition of data scarcity are remarkably improved, compared with the prior art, the runoff simulation precision of the method in a region lacking data is improved by 20%, and an innovative solution is provided for hydrological research of the region lacking data.

Description

Technical Field

[0001] The present invention relates to the field of basin rainfall runoff simulation, and more specifically to an intelligent simulation method for runoff process in data-deficient areas. Background Art

[0002] Accurate simulation of rainfall and runoff processes in river basins is a fundamental topic in hydrological research, and its results directly impact key areas such as water resources planning, flood and drought warning, and ecological protection. However, many unmonitored river basins around the world have long faced a severe shortage of hydrological observation data due to geographical constraints or underdeveloped economies. Runoff simulation in such data-deficient areas faces two major technical bottlenecks: the data dependence of physical mechanism models and the lack of physical interpretability of purely data-driven models. The current mainstream SWAT model realizes the hydrological response process of the watershed through distributed physical process simulation. It has low requirements for the length of time series observation data, but relies on spatial attribute data. Its application in data-deficient areas has significant limitations. First, the model is highly sensitive to the integrity of spatial data such as soil properties and land use. Missing data can easily lead to systematic deviations in runoff parameters. Second, parameter calibration relies on historical observation data, which makes it difficult to adapt to the dynamic simulation of extreme hydrological events under climate change. In actual applications, it is found that when the key input data is missing by more than 30%, the runoff simulation accuracy of the SWAT model will show an exponential decline. With the development of science and technology, researchers have gradually realized that the application effect of traditional physical mechanism models in data-deficient areas is limited, and new methods must be explored to solve this problem. In recent years, the rapid development of artificial intelligence technology has brought new opportunities for innovation in hydrological models. Machine learning, especially deep learning models, has gradually become an emerging hotspot in the field of hydrological models due to its superior performance in dealing with complex nonlinear problems, spatiotemporal heterogeneity, and high-dimensional data. Among the many machine learning models, LSTM models (long short-term memory networks) and Transformer models are more typical representatives; The recently proposed SWAT-LSTM coupling method attempts to combine the advantages of mechanistic modeling and data-driven approaches by using SWAT outputs as prior constraints for LSTM. However, this method still has key flaws in data-poor scenarios: First, the recursive structure of LSTM cannot effectively capture long-term hydrological dependencies across seasons and years, relying on time-series dynamic data. Second, the model does not explicitly construct the spatiotemporal interaction mechanism between meteorological factors and underlying surface characteristics, and therefore does not adequately represent the spatial distribution of runoff in basins with complex topography. More importantly, when the training data length is shorter than the typical hydrological period (usually requiring 5-8 years of continuous data), the model is prone to overfitting and generating non-physical fluctuations. This flaw is particularly pronounced in daily-scale simulations. While the Transformer model demonstrates advantages in capturing long-range dependencies in time series modeling, its direct application in the hydrological field faces dual contradictions: on the one hand, the purely data-driven self-attention mechanism lacks the constraints of physical laws such as mass conservation and energy balance, and may generate prediction results that violate hydrological mechanisms; on the other hand, the standard Transformer architecture requires massive amounts of training data to support its parameter optimization, which fundamentally conflicts with the data conditions in data-scarce scenarios. How to achieve collaborative modeling of physical laws and data characteristics under small sample conditions has become a technical difficulty that urgently needs to be overcome in this field. In summary, the core contradictions of existing technologies can be summarized as follows: (1) physical models are limited by data completeness requirements; (2) hybrid models (such as SWAT-LSTM) are difficult to adapt to small sample scenarios due to time series modeling defects; (3) pure data-driven models (such as Transformer) face the dual dilemma of physical inaccuracy and data requirements. Therefore, it is urgent to develop a runoff process simulation method for data-deficient areas that can achieve collaborative modeling of physical laws and data characteristics under small sample conditions. Summary of the Invention

[0003] The present invention proposes an intelligent simulation method for runoff processes in data-deficient areas. This method is a lightweight three-level coupling modeling method guided by physical mechanisms. Through the physical process analysis of the SWAT model, the spatiotemporal feature extraction of the improved LSTM, and the long-range optimization of the lightweight Transformer, it realizes the collaborative modeling of physical laws and data features under small sample conditions, and achieves high-precision and interpretable runoff simulation under limited data conditions. It solves the technical contradictions in hydrological simulation in data-deficient areas, such as the strong data dependence of physical models and the physical inaccuracy of machine learning models.

[0004] To achieve the above objectives, the technical solution of the present invention is: a method for intelligent simulation of runoff processes in data-deficient areas, characterized by a three-level coupling of a SWAT model, an improved LSTM model, and a lightweight Transformer model. Through the collaborative innovation of physical constraint injection, spatiotemporal feature enhancement, and small sample training strategies, the method overcomes the technical bottleneck of runoff simulation in data-deficient areas. The specific method includes the following steps: Step 1: Collection of basic data; The basic data collected include measured flow data, spatial data (DEM digital elevation model, land use map, soil type map) and satellite meteorological data (precipitation, temperature, solar radiation, relative humidity, wind speed); Step 2: SWAT hydrological model construction and physical variable output; Hydrological processes were simulated using the hydrological model SWAT to obtain preliminary hydrological parameters and data. Seven key physical variables (actual evapotranspiration (PET), potential evapotranspiration (ET), soil moisture content (SW), percolation (PERC), surface runoff (SURQ), groundwater runoff (GWQ), and lateral flow (LATQ)) derived from SWAT were selected to construct a daily-scale multivariate input database. Step 3: Build a SWAT-LSTM-Transformer three-level coupling architecture; The LSTM-Transformer machine learning model was constructed by embedding the LSTM hydrological model into the Transformer module. The input layer of the LSTM hydrological model selected seven output SWAT physical variables and five meteorological data (precipitation, temperature, solar radiation, relative humidity, and wind speed). Step 4: Transfer learning of the SWAT-LSTM-Transformer coupling model; The SWAT-LSTM-Transformer coupled model was pre-trained and calibrated in a watershed with relatively sufficient data, resulting in a SWAT-LSTM-Transformer coupled model that can be used in the study area. Step 5: Runoff simulation and performance evaluation; By inputting meteorological driving data of any period, the trained SWAT-LSTM-Transformer coupling model can extend and supplement the runoff data of any period in the study area, thus realizing the simulation and prediction of the runoff process in the study area. The simulated runoff and the corresponding known measured flow are used to evaluate the model performance through the Nash efficiency coefficient to determine whether the runoff simulation results are valid. If necessary, the hydrological parameters are adjusted until the runoff simulation results are valid.

[0005] In the above technical solution, in step 1, the measured flow data can be derived using the reservoir control operation log through the water balance principle.

[0006] In the above technical solution, the water balance principle means that within a certain period of time, the change in water volume in the reservoir can be described by the relationship between input flow and output flow. Generally speaking, the water balance equation of the reservoir can be expressed as: ; in, is the change in reservoir water volume; is the inflow to the reservoir; is the outflow from the reservoir; Among them, the outflow of the reservoir It can be determined by the following factors: (1) Water consumption ( ): The amount of water supplied by the reservoir for agriculture, industry or residents; (2) Loss ( ): Water loss caused by evaporation or leakage during the operation of the reservoir; (3) Flood volume ( ): The amount of excess water that flows out when the reservoir water level exceeds the maximum storage capacity; Therefore, the outflow of the reservoir can be expressed as: ; To calculate the inflow , through the water balance equation of the reservoir; first, the inflow can be Expressed as: ; Next, we need to consider the water level changes in the reservoir , which can be obtained from hydrological data; the relationship between the water volume change and water level of the reservoir can usually be calculated using the reservoir capacity curve: ; in, is the reservoir water volume at time t; is the reservoir water volume at time t-1; Therefore, combining the various elements of the water balance principle, we can get the complete water balance equation: ; in, is the reservoir water volume at time t; is the reservoir water volume at time t-1; is the inbound flow; is water consumption; is the amount of water lost; is the amount of flood water overflowed; In step 1, relative humidity is a type of meteorological data that is difficult to collect using a ready-made dataset. It can be calculated using the empirical formula: A method of calculating relative humidity that uses the average temperature for the day and dew point temperature This method is based on the following empirical formula to calculate relative humidity. , the empirical formula is as follows: ; in, Indicates the average temperature of the day in degrees Celsius. Indicates the dew point temperature of the day in degrees Celsius. Expressed as the relative humidity for the day, expressed as a percentage. The average temperature and dew point temperature for the day are easily found in the finished dataset; In step 1, missing input data is allowed. The SWAT-LSTM-Transformer coupling model of the present invention has a certain fault tolerance threshold, namely: (1) Allowing any single variable (such as rainfall and temperature) to be missing for less than 30 consecutive days or with a cumulative missingness of less than 15%, the SWAT-LSTM-Transformer coupling model of the present invention can compensate for this by using physical variable substitution and attention masking methods; (2) When no more than three variables (such as rainfall, temperature, and humidity) are missing at the same time, the SWAT-LSTM-Transformer coupled model of the present invention can still operate as long as at least four SWAT physical variables are retained and the consecutive missing time steps do not exceed 7 days.

[0007] In the above technical solution, in step 2, the SWAT (Soil and Water Assessment Tool) model is based on the physical mechanism of the hydrological cycle and can fully simulate a series of complex processes such as rainfall, evaporation, infiltration, surface runoff, groundwater flow and vegetation growth. The model divides the watershed into multiple sub-basins and hydrological response units (HRUs) by inputting meteorological data (precipitation, temperature, solar radiation, relative humidity, wind speed), elevation data, land use and soil type, thereby capturing the differentiated effects of different land use types and soil conditions on hydrological processes. The present invention uses the actual evapotranspiration (PET), potential evapotranspiration (ET), soil water content (SW), seepage (PERC), surface runoff (SURQ), groundwater runoff (GWQ) and lateral flow (LATQ) among the hydrological variables output by the SWAT model as multivariate input data of the model; The construction method of the SWAT hydrological model includes the following steps: Step 21: Use the collected digital elevation model (DEM) in the SWAT model module of ArcGIS software to generate the corresponding river network and set the basin outlet, and then divide the sub-basin (i.e., the sub-basin division module of the SWAT model); Step 22: Generate hydrological response units (HRUs) in the entire watershed using land use, soil type, and slope data based on the sub-watersheds, which is the hydrological response unit division module of the SWAT model. Step 23: Next, daily meteorological data (including rainfall, temperature, solar radiation, relative humidity, and wind speed) are input for simulation to obtain hydrological variables and output physical variables. The physical variables include actual evapotranspiration, potential evapotranspiration, soil moisture, seepage, surface runoff, subsurface runoff, and lateral flow. Step 24: Construct the physical variables output by the SWAT model into a daily-scale multivariate input database.

[0008] In the above technical solution, in step 23, the hydrological variables of each sub-basin are weighted to obtain the hydrological variable value of the basin area where the measured flow monitoring location is located. The specific calculation formula is as follows: ; in, is the ith simulation variable of sub-basin j; is the area of ​​sub-basin j; a is the area of ​​the basin where the measured flow monitoring location is located; n is the number of sub-basins in the basin where the measured flow monitoring location is located; is the desired mean level for each variable.

[0009] In the above technical solution, in step 3, a SWAT-LSTM-Transformer three-level coupling architecture is constructed, including the following steps: Step 31: The input of the LSTM module is the preprocessed daily-scale multivariate database (actual evapotranspiration (PET), potential evapotranspiration (ET), soil moisture content (SW), percolation (PERC), surface runoff (SURQ), groundwater runoff (GWQ), and lateral flow (LATQ)) and meteorological data (precipitation, temperature, solar radiation, relative humidity, and wind speed). Step 32: The LSTM module performs preliminary processing on these time series data (i.e., multivariate databases and meteorological data) through its internal memory units and gating mechanism, extracting long-term trends and potential temporal correlation information. In the specific implementation process, the model uses two stacked LSTM layers (each containing 64 hidden units) to better extract the temporal characteristics of the data (i.e., hydrological patterns), enhance the model's expressiveness, and prevent overfitting through a Dropout layer. The Dropout value of the present invention is 0.1. This solves the problem that existing technologies have difficulty effectively capturing long-term hydrological dependencies across seasons and years and rely on time series dynamic data. It also addresses the problem that existing technologies are insufficient in representing the spatial differentiation patterns of runoff in watersheds with complex terrain. When the training data length is shorter than the typical hydrological period (usually 5-8 years of continuous data), the model is prone to overfitting and generating non-physical fluctuations. Step 33: The data output by LSTM is used as the input of the Transformer module; the Transformer processes the input data in parallel through its multi-head self-attention mechanism, focusing on the local and global correlations in the time series; each attention part independently focuses on different patterns or features in the time series, thereby improving the model's ability to capture complex spatiotemporal relationships. The self-attention mechanism of the Transformer model can identify the precursory features of extreme events, greatly improving the model's simulation effect on extreme events. At the same time, the Transformer's feedforward neural network performs further nonlinear transformations on the features of each time step, enhancing the model's expressive power. The Transformer module of the present invention has 8 attention layers, 128 hidden dimensions, and 128 feedforward layer dimensions.

[0010] In the above technical solution, in step 4, transfer learning is used as a machine learning method. Transfer learning is based on the knowledge learned in the source field to assist the learning of the target field or target task. Its core idea is to use existing knowledge to reduce the model's dependence on large-scale labeled data in the target task, thereby improving model performance and learning efficiency. Based on the transfer learning principle of hydrological similarity, the present invention migrates the model pre-trained in the data-rich basin to the data-scarce basin, realizes the intelligent simulation of the runoff process in the data-deficient area, and improves the runoff simulation accuracy; solves the problem of decreased runoff simulation accuracy due to missing key input data; at the same time, when pre-trained in the data-sufficient basin, the model has learned the rainstorm-runoff response patterns of different climate zones (such as the difference between typhoon-type rainfall and convective-type rainfall). After migrating to the data-deficient basin, even if the local extreme event samples are missing, it can still maintain a high prediction accuracy; Domains in Transfer Learning and tasks It is generally composed of two tuples, and the specific expression is: ; ; in: is the marginal density probability; is the feature space, and ; Domain The objective function of is the label space; The machine learning model first Central source task The process of learning and accumulating learning knowledge is called the pre-training process of the model; the pre-trained model is used in the target domain. Target tasks Learning prediction functions When ,the knowledge learned in the source domain can be used to improve the recognition and accuracy efficiency of the target task when the number of samples is small; Measured flow data and seven selected parameters from the SWAT hydrological model (i.e., actual evapotranspiration, potential evapotranspiration, soil moisture, seepage, surface runoff, subsurface runoff, and lateral flow) and five meteorological data sets (including rainfall, temperature, solar radiation, relative humidity, and wind speed) were input into the SWAT-LSTM-Transformer coupled model for pre-training. The SWAT parameters and the Transformer position encoding layer were frozen, and the training period was set to a 4:1 ratio with the simulation period. The simulation results were calibrated to complete the transfer learning of the SWAT-LSTM-Transformer coupled model. In step 4, to improve model training performance and stability, residual connections and layer normalization were introduced between the LSTM and Transformer modules, as well as between internal layers, to ensure smooth gradient flow and avoid gradient vanishing or exploding. When dealing with complex nonlinear spatiotemporal relationships, the Transformer model's powerful multi-head self-attention mechanism was leveraged for effective processing. The Transformer model eliminates the need to serialize input data. Instead, it uses a multi-head self-attention mechanism to concurrently analyze dependencies between different data locations, enabling it to more accurately capture the interactions between different factors in complex watershed systems. Its residual connections and layer normalization techniques further enhance the model's convergence and stability, making it particularly suitable for processing large-scale, high-dimensional data.

[0011] In the above technical solution, in step 5, the present invention uses the Nash efficiency coefficient to evaluate model performance. The Nash efficiency coefficient is an indicator widely used in hydrological model evaluation to measure the degree of fit between model simulation results and observed data. This coefficient is based on the Nash-Sutcliffe efficiency coefficient, and its calculation formula is as follows: ; Where E represents the Nash efficiency coefficient, is the observed value, is the analog value, is the mean of the observed values, and n is the total number of data points. The Nash efficiency coefficient ranges from negative infinity to 1, where 1 indicates a perfect fit of the model to the observed data, 0 indicates that the model's predictions are comparable to the observed mean, and negative values ​​indicate poor model performance. By calculating the Nash efficiency coefficient, we can objectively evaluate the performance and accuracy of hydrological models, providing an important reference for model improvement and optimization.

[0012] In the above technical solution, the runoff simulation in step 5 specifically includes the following steps: Using the trained SWAT-LSTM-Transformer coupling model, we input seven hydrological parameters (actual evapotranspiration, potential evapotranspiration, soil moisture, seepage, surface runoff, groundrunoff, and lateral flow) and five meteorological data (precipitation, temperature, solar radiation, relative humidity, and wind speed) corresponding to the required time series in the study area. This allows us to obtain a runoff simulation sequence for the study area, which can then be sorted out to ultimately extend and complete the runoff process in the study area.

[0013] Compared with the prior art, the method of the present invention has the following beneficial effects: (1) Effective integration of physical mechanisms and machine learning: The present invention combines the physical process analysis of the SWAT model with an improved LSTM and lightweight Transformer model to implement a physical mechanism-guided machine learning modeling method. The present invention not only utilizes the physical mechanism of the SWAT model based on the hydrological cycle (such as the simulation of rainfall, evaporation, infiltration, etc.), but also fully leverages the data-driven advantages of LSTM in time series data processing and Transformer in capturing long-range dependencies. The present invention effectively combines the advantages of the three, breaking through the contradiction between data dependency and physical inaccuracy in existing technologies. It also solves the simulation difficulties of traditional physical models in areas with insufficient data due to insufficient data, and the lack of physical interpretability of pure machine learning models. (2) Multi-model coupling improves accuracy and scalability: The present invention adopts a three-level coupled architecture of SWAT-LSTM-Transformer, fully leveraging the physical foundation of SWAT, the time series extraction capability of LSTM, and the long-range dependency capture capability of Transformer, thereby improving the Nash efficiency coefficient of model simulation and effectively enhancing the accuracy of runoff simulation. The present invention can adapt to different watersheds and different data conditions, exhibiting good scalability and wide applicability. The model architecture of the present invention is flexible. After building and training a SWAT-LSTM-Transformer model in an area with sufficient data, it can be used in any region to perform intelligent simulation of runoff processes. It only requires that the region meet the basic conditions for building the SWAT model and supports the adjustment of input variables and parameters (such as the number of LSTM layers and the number of Transformer heads) according to specific application scenarios. (3) Transfer learning and lightweight design: By pre-training the model in a data-rich basin and migrating it to data-deficient areas, the present invention achieves efficient learning under small sample conditions, significantly reducing the present invention's dependence on large-scale data. At the same time, the lightweight Transformer design (8-layer attention mechanism, 128-dimensional hidden layer, Dropout 0.1) reduces computational complexity and facilitates practical deployment, solving the problem of excessive resource consumption of traditional models. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 This is a partial schematic diagram of the LSTM machine learning model of the present invention; Figure 2 It is a flow chart of the SWAT-LSTM-Transformer coupling model of the present invention; Figure 3 This is a diagram showing the running results of the SWAT-LSTM-Transformer coupling model according to an embodiment of the present invention; Figure 4 Flowchart of the present invention. DETAILED DESCRIPTION

[0015] The following detailed description of the embodiments of the present invention is given in conjunction with the accompanying drawings, which do not limit the present invention but are merely examples. The description makes the advantages of the present invention clearer and easier to understand.

[0016] This paper considers the limited hydrological data in data-deficient areas and proposes a method that combines the hydrological model SWAT with the machine learning models LSTM and Transformer. This method uses limited measured flow data and easily movable spatial and meteorological data to simulate the runoff process, thereby extending and supplementing hydrological data and providing a new research approach and method for studying the hydrological situation in data-deficient areas. In this paper, data-deficient areas refer to areas with less than five years of measured runoff data or annual missing data greater than 30%, with missing data greater than 40% for key data such as rainfall and less than 70% spatial coverage. Furthermore, these areas meet the requirements of low basic geographic data accuracy (DEM resolution greater than 90 meters, soil conditions and land use map completeness less than 60%) or lack records of extreme hydrological events (such as a 30-year rainstorm). The present invention first conducts a basic survey of the study area. It is required that the study area is in or mostly in a non-plain river network area to ensure that the main stream river network situation in the study area can be extracted normally; it is required that the study area has at least one place near the main stream where measured flow data can be obtained, which can be obtained through a hydrological station or reservoir water balance. Then the scope of the study area is delineated, and a visit survey is conducted in the study area or the finished data set is used to obtain the required elevation data, land use data and soil data, and the required meteorological data is obtained using a satellite meteorological data set. Based on these basic data, it is determined whether it is necessary to go to other data-rich basins for transfer learning; if so, other basins are selected to construct the SWAT model, and the obtained hydrological parameters are weighted by area to obtain the hydrological parameters of the measured point basin. Finally, the seven processed hydrological variables and meteorological data are used as feature parameters of the LSTM model. The output results are then fed into the Transformer model for continuous learning and calibration with the measured flow data. The hydrological variables and meteorological data for the required period in the study area are then substituted into the trained SWAT-LSTM-Transformer coupled model for simulation to obtain the required flow sequence. If the required flow sequence is not obtained, the study area can be replaced with another basin and the data from the study area can be used for the same modeling, training, and simulation processes. In addition, the coupling process of the machine learning model LSTM model and the Transformer model in the present invention does not need to be repeated. The coupling process of the two models has been completed in the corresponding code. Therefore, when using the present invention in the future, you only need to input data according to the input format of the model and set the sequence length ratio of the trainer and simulator to obtain the required runoff simulation process and efficiency coefficient. In general, the present invention is based on the hydrological model SWAT model, and intelligently combines physical mechanisms and machine learning through the machine learning model LSTM model and Transformer model, breaking through the contradiction between data dependence and physical inaccuracy in existing technologies, and to a certain extent solving the problem of runoff simulation in data-deficient areas.

[0017] Example: The present invention is now described in detail by taking the application of the present invention to the intelligent simulation of runoff process in small and medium-sized watersheds of small and medium-sized reservoirs in a certain province as an example. It also has a guiding role in the application of the present invention to the intelligent simulation of runoff process in other areas with insufficient data.

[0018] This embodiment selects a small and medium-sized river basin with a small and medium-sized reservoir in the north-central part of a certain province as the research object for research. There are no flow stations, rain gauges and other hydrological research stations available for research in the study area, which makes it difficult to obtain flow data, and rainfall data and other meteorological data cannot be obtained directly. Spatial data relies on finished data sets from satellite remote sensing, and it belongs to a data-deficient area. The reservoir is a centralized drinking water source at the county level and above. Ensuring that its water quality is stable and up to standard is crucial to ensuring the drinking water safety of the county, urban areas and surrounding towns. It is also an important online regulating reservoir for the Hubei North Project and the Hubei North Phase II Project. It can accurately simulate the runoff process of the reservoir basin, which is of great significance and can provide a data basis and theoretical basis for subsequent reservoir operation scheduling and water quality research. The present invention is described in detail by simulating the runoff process of the reservoir basin and supplementing the hydrological data of the study area; This reservoir intercepts the Fengjiangkou River, an eastern branch of the Jueshui River in the Fushan River basin. It is located in a town in the central part of a certain city, 17 km from the county seat and 35 km from a certain city. The basin is a hilly area formed by the extension of the Tongbai Mountains. The ground elevation at the dam site is 94.00 meters, with a general relative elevation difference of 40-60 meters. The basin is leaf-shaped, surrounded by mountains on the west, north, and east, with lower elevations in the south. The mountains slope from north to south, creating a deep, incised riverbed 20-60 meters wide with a gradient of 1 / 350-1 / 550. There are no dense forests within the basin, but shrubs and vegetation are abundant. The climate in the basin exhibits north-south transitional characteristics. Influenced by the monsoon, the rainy season is late, mostly occurring between June and August. It accounts for 55% of the annual rainfall, often in the form of torrential rains, with annual precipitation ranging from 581.2 to 1331.4 mm. The reservoir controls a drainage area of ​​460km2 and intercepts a main river channel of approximately 48.3km in length; In this embodiment, the following steps are mainly included: Step 1: Collect basic data, including measured flow data, spatial data and meteorological data; There are no effective hydrological stations such as flow stations in the area. By collecting the control operation logs of the reservoir, the corresponding pre-inflow flow data are obtained using the water balance method, and this data is used as the measured flow data for research; The study area was defined. Using the finished dataset, spatial data for the reservoir basin was obtained, including elevation data, land use data, and elevation data. Meteorological data, including rainfall, temperature, solar radiation, relative humidity, and wind speed, were obtained using the ERA5 global climate reanalysis dataset. The collected data was daily and covered the period from 2016 to 2019. In this example, the daily average temperature and dew point temperature from the ERA5 global climate reanalysis dataset were used to calculate the relative humidity data for the day. Choose to use the basic data of the region for transfer learning; Step 2: SWAT hydrological model construction and physical variable output; In this step, you need to use ArcGIS software to import the collected elevation files into the SWAT model module in ArcGIS software, and then extract the river network of the study area. After completing the river network extraction, based on the extracted river network, check the sub-basin output points and set new output points on the river network. After confirming that they are correct, select the total outlet of the reservoir basin, and then complete the sub-basin division operation. Finally, calculate the sub-basin parameters; In this example, the reservoir basin is divided into 35 sub-basins, of which sub-basin No. 34 is used as the measured flow observation basin, and its basin outlet is used as the measured flow observation point; The next step is to divide the hydrological response units (HRU). In the operation setting of the SWAT model, the division of the reservoir basin hydrological response units is a key step. The division of HRU is one of the unique features of the SWAT model. The SWAT model is based on the division of sub-basins, combined with factors such as land use type, soil type and slope, to divide areas with similar characteristics into the same type of HRU, and assumes that these HRUs behave consistently in hydrological behavior. During the model calculation process, the hydrological process is calculated separately for the sub-basins of different HRUs, and then the results of each HRU are superimposed at the outlet of the sub-basin to obtain the results of the entire sub-basin. Therefore, the number of HRUs directly affects the calculation speed of the model; In this step, the collected land use files and soil condition files were sequentially input into the SWAT model. Then, the slope was defined based on the slope conditions of the study area analyzed by ArcGIS. Three slope categories were defined, namely 0-5%, 5-30% and 30-9999%. HRUS was generated and hydrological characteristics such as land use, soil type and slope were superimposed on each HRU. Then the HRU is defined to avoid generating too many HRUs and improve the accuracy and speed of SWAT operations. In this example, land use, soil distribution, and slope types below 10%, 15%, and 10% will be split and merged into other types. The five collected meteorological data files, namely daily average rainfall, daily maximum / minimum temperature, net solar radiation, relative humidity, and wind speed, are imported into the SWAT model and written into the database. Then the SWAT model is run, where the time step, run time, and warm-up period can be set. This example performs daily-scale calculations from 2016 to 2019, with a warm-up period of 1 year. After running, the simulation results are saved to obtain the simulated runoff process and physical variables. The physical variables are then sorted to form a multivariate input database. Table 1 shows some of the results from the multivariate input database. This example selects seven representative sub-basins around the reservoir. In Table 1, SUB represents the sub-basin number; FLOW is the outlet runoff at the sub-basin, measured in cubic meters per second; TN and TP are the total nitrogen and total phosphorus contents, respectively, measured in kilograms; PET, ET, SW, PERC, SURQ, GWQ, and LATQ are the actual evapotranspiration, potential evapotranspiration, soil moisture, seepage, surface runoff, subsurface runoff, and lateral flow, respectively. The units for these seven hydrological variables are all millimeters. Table 1 Partial results of multivariate input database ; Step 3: Transfer learning of the SWAT-LSTM-Transformer coupling model; The seven hydrological variables obtained from running the SWAT model are calculated for the basin where the observation point is located. The observation point is in sub-basin 34, meaning the seven hydrological variables for sub-basin 34 are the required hydrological variables for the observation point. The seven hydrological variables (PET, ET, PERC, SW, GWQ, SU, RQ, and LATQ) and the corresponding meteorological data are substituted as feature parameters into the LSTM-Transformer coupled model. The corresponding measured flow data is then substituted into the LSTM-Transformer coupled model for training and learning. This example divides the time series from January 1, 2016, to December 31, 2019, into a training period and a simulation period in a ratio of 4:1. Table 2 shows the results of parameter calibration after running the SWAT-LSTM-Transformer coupled model. The learning rate is a key hyperparameter in optimization algorithms (such as gradient descent) that determines the size of each weight update step. The batch size is the number of samples processed by the model during each training session. Dropout is a regularization technique used to prevent model overfitting. The sequence length is the maximum length of each sequence in the input data and determines the number of time steps the model processes at each time. Embed_dim is the dimension of the vector when the model converts input data (such as words, characters, or other types of data) into a vector representation. Num_heads is the number of layers used in the attention mechanism of the Transformer model. Increasing the number of layers allows the model to capture more information. FF_dim is the dimension of the hidden layer of the feedforward neural network in the Transformer model. NSE (Nash efficiency coefficient) is a metric used to evaluate the predictive accuracy of regression models, particularly useful in fields such as hydrology. Its value ranges from negative infinity to 1, with 1 indicating perfect prediction. The results show that the SWAT-LSTM-Transformer coupled model based on the transfer learning principle performs well in simulating measured flow at the reservoir dam site. The model has a Nash efficiency coefficient of 0.8 for multiple simulations and also exhibits high simulation accuracy in simulating peak and valley flows. Table 2 Parameter calibration and simulation results of the SWAT-LSTM-Transformer coupling model ; Figure 3 This is the simulation result after pre-training of the SWAT-LSTM-Transformer coupling model. Figure 3 It is clear that the simulated runoff process line is highly consistent with the measured flow process line, which shows that the model simulation effect is good and reflects the distribution of runoff over time very well; Step 4: Runoff simulation and performance evaluation; Based on the trained SWAT-LSTM-Transformer coupled model, this example can extend and supplement the runoff hydrological data of other sub-basins in the study area. This example randomly selects four time periods for simulation: 1993-1994, 1958-1959, 1970-1971, and 1976-1977, with each period lasting two years. In this example simulation, changes in underlying surface conditions such as land use and soil conditions in the study area due to changes in time are not considered. Using the global climate reanalysis dataset ERA5, meteorological data from 1992 to 1994 were collected, with a warm-up period of 1 year; meteorological data from 1957 to 1959 were collected, with a warm-up period of 1 year; meteorological data from 1969 to 1971 were collected, with a warm-up period of 1 year; meteorological data from 1955 to 1977 were collected, with a warm-up period of 1 year. Here, the average daily temperature and dew point temperature in the global climate reanalysis dataset ERA5 are used to obtain the corresponding daily relative humidity through an empirical formula; Substitute them into the established SWAT model for calculation and obtain their corresponding 7 hydrological variables. The sub-basins numbered 16, 17, 22, 24, 29, 30, 32 and 34 around the reservoir are selected as the simulation objects of the runoff process, that is, these sub-basins are used as the flow of simulated flow, and their sub-basin outlets are used as the simulated flow outlets; Since the measured flow data used for training is four years long, and the simulation data is two years long, the ratio of the sequence lengths between the training and simulation periods in the SWAT-LSTM-Transformer coupled model is still set to 2:1. The hydrological variables and corresponding meteorological data for the four time periods mentioned above are substituted into the SWAT-LSTM-Transformer coupled model as the characteristic parameters for the simulation period. The measured flow data, hydrological variables, and corresponding meteorological data from 2016 to 2019 are substituted into the coupled model as the characteristic parameters for the training period, and then simulation calculations are performed. Table 3 shows the Nash efficiency coefficient results of the simulation results when the data of the four time periods are substituted into the SWAT-LSTM-Transformer coupling model. It can be clearly seen that the NSE result values ​​of all time periods are 0.75 and above, which shows that the SWAT-LSTM-Transformer coupling model of the present invention has a good effect in simulating runoff and stable performance. Therefore, the inventive method used in this example achieves the expected effect, that is, the SWAT-LSTM-Transformer coupling model of the present invention can be used to intelligently simulate the runoff process in the study area, which can effectively extend and supplement the hydrological data of the study area and the performance of simulating runoff is stable. Table 3 Performance of SWAT-LSTM-Transformer coupling model simulation .

[0019] Comparative Example 1: The following example simulates rainfall runoff and calibrates runoff parameters using the existing SWAT model, a traditional hydrological model, for the study area in this example. The Nash efficiency coefficient is used to compare the proposed SWAT-LSTM-Transformer coupled model with the existing SWAT model.

[0020] The existing technology is to calibrate the runoff parameters of the 34th sub-basin in the study area using the SWAT-CUP software, a model calibration and verification tool that is compatible with the SWAT model, after completing the above step 2 to build the SWAT model of the study area; In this comparative example, 13 hydrological parameters were selected for the runoff calibration part of the model with reference to the recommended parameters of the SWAT calibration guide in SWAT-CUP. The parameters are shown in Table 4 below. Table 4 13 hydrological parameters selected for SWAT-CUP runoff calibration ; In the SWAT-CUP software, the runoff calibration process was set to 500 cycles, with four parameter iterations, each taking four hours. The sensitivity ranking and significance of the model runoff parameters are shown in Table 5 below. Table 5 Results of sensitivity analysis of runoff parameters ; In this example, SWAT-CUP software was used for parameter calibration. The parameters were adjusted based on the results of the runoff parameter sensitivity analysis to obtain the best simulation results. The changes in each parameter and the change values ​​are shown in Table 6 below, where r indicates that the parameter is replaced by the initial value of the parameter multiplied by (1 + the change value); v indicates that the parameter is replaced by the change value; Table 6 Changes in runoff parameters and their values ​​after SWAT-CUP calibration ; SWAT-CUP software was used to calibrate the runoff parameters of the 34th sub-basin of the model. The software outputted the Nash efficiency coefficient of the model, NSE = 0.52. That is, the runoff parameters of the 34th sub-basin were calibrated and the runoff simulation was performed. Finally, the Nash efficiency coefficient of the model, NSE = 0.52; In summary, using the existing technology, namely the traditional SWAT model, for runoff simulation and the accompanying SWAT-CUP for calibration, the model's Nash efficiency coefficient (NSE) is 0.52. Furthermore, parameter calibration using existing technology requires iteration. In this case, 500 iterations were performed, four times for four hours each, and a minimum of 16 hours were required to obtain a usable result. Using the SWAT-LSTM-Transformer coupled model of the present invention to simulate runoff in the aforementioned data-deficient region, the model's Nash efficiency coefficient was approximately 0.8, exceeding the 0.52 Nash efficiency coefficient of the model using existing technology. This demonstrates that the SWAT-LSTM-Transformer coupled model of the present invention is superior to existing technology (i.e., the traditional SWAT model) in simulating runoff processes. In terms of model time, the SWAT-LSTM-Transformer coupled model of the present invention takes approximately 10 minutes from learning to simulating the results for a sub-basin, and simulates runoff in other time and space (i.e., other sub-basins and time periods) in approximately 5 minutes (runoff simulation accuracy is approximately 0.7, while the existing technology simulation accuracy is less than 0.5). The overall simulation time of the present invention is much shorter than that of the existing technology, and the simulation accuracy is much higher than that of the existing technology. In summary, from the perspectives of model simulation effect and model simulation time, the SWAT-LSTM-Transformer coupling model of the present invention is superior to the traditional hydrological model SWAT model.

[0021] Comparative Example 2: The following example simulates rainfall runoff using the recently proposed SWAT-LSTM coupled model for the study area of ​​this example. The simulation results are evaluated using the Nash efficiency coefficient (NSE) to compare and analyze the advantages of the proposed SWAT-LSTM-Transformer coupled model compared to the SWAT-LSTM coupled model.

[0022] In Comparative Example 2, the SWAT-LSTM coupled model was used in the same manner as in the present embodiment, with the Transformer model removed from the code. This Comparative Example 2 simulated the runoff process in each sub-basin over the same four time periods as in the present embodiment, and obtained the NSE (Non-Separate Efficiency) of the model simulations under different conditions, which represents the performance of the SWAT-LSTM coupled model. The results are shown in Table 7 below. Table 7 Performance of SWAT-LSTM coupling model simulation ; Comparing Table 7 with Table 3, it can be clearly seen that the NSE results simulated by the SWAT-LSTM coupling model are mostly below 0.6, while the NSE results simulated by the SWAT-LSTM-Transformer coupling model of the present invention are all around 0.8. Compared with the SWAT-LSTM coupling model, the performance of the SWAT-LSTM-Transformer coupling model of the present invention is improved by more than 20%. This shows that the SWAT-LSTM-Transformer coupling model is significantly better than the SWAT-LSTM coupling model in simulating runoff in data-deficient areas. In addition, from the perspective of model principle and structure, compared with the SWAT-Transformer model, the LSTM model of the present invention can better reflect the local characteristics of the data, force the model output load mass conservation, and reduce the load of the Transformer model, requiring less data and less training time. Compared with the LSTM-Transformer model, the present invention uses the SWAT model to provide data, has physical mechanisms and organic constraints, is more convincing, has increased fault tolerance for missing data, and improves the ability to display regional hydrological characteristics. The present invention creatively combines the Transformer model with the SWAT-LSTM coupling model, significantly improving the generalization performance under data-scarce conditions, and improving the adaptability and reliability of the SWAT-LSTM-Transformer coupling model in data-scarce, distribution-different or unknown scenarios. That is, the model can learn universal hydrological laws from limited training data and maintain high prediction accuracy in similar scenarios, achieving high-precision and explainable runoff simulation under limited data conditions; and solving the technical contradictions of strong data dependence of physical models and physical inaccuracy of machine learning models in hydrological simulations in data-scarce areas.

[0023] Other parts not described belong to the prior art.

Claims

1. An intelligent simulation method for runoff processes in data-deficient areas, characterized by: The following steps are included: Step 1: Collection of basic data; The basic data collected include measured flow data, spatial data and satellite meteorological data in the study area; Step 2: SWAT hydrological model construction and physical variable output; The hydrological model SWAT was used to simulate the hydrological process to obtain preliminary hydrological parameters and data and form a multivariate input database; Step 3: Build a SWAT-LSTM-Transformer three-level coupling architecture; The LSTM-Transformer machine learning model is constructed by embedding the LSTM hydrological model into the Transformer module. The input layer of the LSTM hydrological model selects the physical variables and meteorological data output by SWAT. Step 4: Transfer learning of the SWAT-LSTM-Transformer coupling model; The SWAT-LSTM-Transformer coupled model was pre-trained and calibrated in a watershed with relatively sufficient data, resulting in a SWAT-LSTM-Transformer coupled model that can be used in the study area. Step 5: Runoff simulation and performance evaluation; Input meteorological driving data of any period into the trained SWAT-LSTM-Transformer coupling model, extend and supplement the runoff data of any period in the study area, and realize the simulation and prediction of the runoff process in the study area; The simulated runoff and the corresponding known measured flow are used to evaluate the model performance through the Nash efficiency coefficient to determine whether the runoff simulation results are valid. If necessary, the hydrological parameters are adjusted until the runoff simulation results are valid.

2. The intelligent simulation method for runoff process in data-deficient areas according to claim 1 is characterized by: In step 1, the measured flow data is derived from the reservoir control operation log using the water balance principle and combined with various elements of the water balance principle to obtain a complete water balance equation: ； in, is the reservoir water volume at time t; is the reservoir water volume at time t-1; is the inbound flow; is water consumption; is the amount of water lost; is the amount of flood water.

3. The intelligent simulation method for runoff process in data-deficient areas according to claim 1 or 2 is characterized by: In step 2, the SWAT hydrological model simulates the hydrological variables in the study area by inputting the physical parameters of the study area; the hydrological variables include the runoff process in the study area; the physical parameters of the study area include rainfall, temperature, solar radiation, relative humidity, wind speed, elevation data, soil type and land use.

4. The intelligent simulation method for runoff process in data-deficient areas according to claim 1 is characterized by: In step 2, the SWAT hydrological model is constructed and physical variables are output, including the following steps: Step 21: Use the collected digital elevation model of the study area in the SWAT model module of ArcGIS software to generate the corresponding river network and set the basin outlet, and then divide the study area into sub-basins; Step 22: Generate hydrological response units within the entire watershed using land use, soil type, and slope data on a sub-watershed basis; Step 23: Next, daily meteorological data is input into the SWAT model for simulation, and hydrological variables in the study area are calculated and physical variables are output. The physical variables include actual evapotranspiration, potential evapotranspiration, soil moisture, seepage, surface runoff, subsurface runoff, and lateral flow. Step 24: Construct the physical variables output by the SWAT model into a daily-scale multivariate input database.

5. The intelligent simulation method for runoff process in data-deficient areas according to claim 4 is characterized by: In step 23, the meteorological data includes rainfall, temperature, solar radiation, relative humidity, and wind speed; If it is difficult to find the relative humidity in the finished product data set, it can be calculated using the empirical formula: ； in, Indicates the average temperature of the day in degrees Celsius; Indicates the dew point temperature of the day, in degrees Celsius; It is expressed as the relative humidity of the day, expressed as a percentage; The hydrological variables of each sub-basin are weighted to obtain the hydrological variable values ​​of the basin area where the measured flow monitoring location is located. The specific calculation formula is as follows: ； in, is the ith simulation variable of sub-basin j; is the area of ​​sub-basin j; a is the area of ​​the basin where the measured flow monitoring location is located; n is the number of sub-basins in the basin where the measured flow monitoring location is located; is the desired mean level for each variable.

6. The intelligent simulation method for runoff process in data-deficient areas according to claim 5 is characterized by: In step 3, a SWAT-LSTM-Transformer three-level coupling architecture is constructed, which includes the following steps: Step 31: The input of the LSTM module is the daily-scale multivariate database processed in step 24 and the meteorological data in step 23; Step 32: The LSTM module performs preliminary processing on the time series data through its internal memory units and gating mechanism to extract long-term trends and potential temporal correlation information. In the specific implementation, the model uses two stacked LSTM layers and uses a Dropout layer to prevent overfitting. The Dropout value is 0.

1. Step 33: The LSTM output data is used as the input of the Transformer module; the Transformer module has 8 attention layers, 128 hidden dimensions, and 128 feedforward layer dimensions.

7. The intelligent simulation method for runoff process in data-deficient areas according to claim 6 is characterized by: In step 4, the transfer learning of the SWAT-LSTM-Transformer coupling model includes the following steps: Step 41: Input the measured flow data in step 1 and the seven physical variable parameters and five meteorological data selected by the SWAT hydrological model in step 2 into the SWAT-LSTM-Transformer coupled model for pre-training, and freeze the SWAT parameters and the Transformer position encoding layer; Step 42: Set the ratio of the training period length to the simulation period length to 4:1, and use the measured flow data to calibrate the simulation results.

8. The intelligent simulation method for runoff process in data-deficient areas according to claim 1 is characterized by: In step 5, runoff simulation and performance evaluation are carried out, including the following steps: Step 51: Use the trained SWAT-LSTM-Transformer coupled model to simulate the multivariate input database of the study area corresponding to the required time series. Step 52: Use NSE and the corresponding known measured traffic to evaluate the performance of the SWAT-LSTM-Transformer coupling model.

Images