Deep learning runoff simulation method coupled with crop water demand

Abstract

The invention discloses a method for simulating regional runoff based on a coupling model VMD-CNN-LSTM-Attention of irrigation water consumption of main crops. The method comprises the following steps: 1, calculating daily reference crop evapotranspiration ET0 according to meteorological station data of a target region; 2, calculating the irrigation water consumption of different crops in different periods by utilizing the crop growth coefficients of the different crops; and 3, taking the main crop irrigation water consumption, rainfall, temperature, evapotranspiration, soil water content and other monthly elements in the target area calculated in the step 2 as driving factors, and inputting the driving factors into the coupling model for monthly runoff prediction. The method has the following beneficial effects that the irrigation water consumption of main crops with high correlation with the runoff in the research area is calculated, and driving factors with higher correlation with the runoff can be found out according to local conditions as input, so that the prediction precision is improved.

Description

Technical Field

[0001] This invention relates to a method for simulating regional runoff, and in particular to a method for simulating regional runoff based on a coupled model of irrigation water consumption for major crops, VMD-CNN-LSTM-Attention. Background Technology

[0002] In recent years, with rising global temperatures and increased frequency of extreme weather events, climate change has had a significant impact on water cycles and agricultural production in semi-arid regions. Affected by environmental changes, the spatiotemporal distribution of runoff has shifted, leading to a growing imbalance between regional water supply and demand, which poses a serious threat to ecosystem balance and sustainable social development. Because runoff is influenced by multiple factors, including natural factors and human activities, runoff sequences exhibit highly nonlinear, complex, and non-stationary characteristics, making prediction extremely difficult. Therefore, accurate runoff prediction requires a deeper understanding of the mechanisms driving runoff changes and the identification of the complex relationships between runoff and various environmental factors, thereby providing insights for the rational utilization of watershed water resources.

[0003] Currently, runoff prediction models mainly include two types: physical models and data-driven models. Physical models are based on simulations of hydrological processes. However, their accuracy is limited by the high requirements for observational data and hydrological parameters, difficulties in model calibration, and incomplete understanding of the mechanisms of some hydrological processes. Therefore, they still face significant challenges in practical applications. Data-driven prediction models, on the other hand, do not consider the physical mechanisms of hydrological processes. Instead, they establish the data patterns between driving factors and runoff to create an optimal relationship model for runoff prediction. This has gradually become a research hotspot in recent years.

[0004] Currently, the most widely used data-driven models include machine learning and time series analysis. Machine learning models, such as Support Vector Machines (SVMs) and Random Forests, are generally black-box models, unable to provide explicit analytical expressions, and sometimes prone to overfitting. In time series analysis models, such as Autoregressive Moving Averages (ARMAs), only reflect linear relationships between variables, and their prediction accuracy in nonlinear aspects is not high. RNNs suffer from gradient vanishing or exploding during backpropagation, affecting the capture of long-range information. Long Short-Term Memory (LSTM) Neural Networks, proposed by Schmidhuber and Hochreiter in 1997, are a variant of RNNs, possessing nonlinear prediction capabilities, faster convergence speeds, and the ability to effectively handle long-range information, resulting in significant improvements in model performance. Compared to RNNs, LSTMs solve the long memory loss problem, giving them a natural advantage in runoff prediction. Numerous studies have demonstrated that LSTM models outperform traditional machine learning models and physical hydrological models in simulation. Research applying LSTM models to over 500 watersheds has shown that, with training on more sample data, LSTM's prediction results are superior to physical models. In addition, LSTM models were used to conduct research and analysis in areas with insufficient data, which effectively improved runoff prediction in these areas.

[0005] The selection of input variables is crucial for data-driven methods; choosing appropriate input features can improve the accuracy and interpretability of the model. Although the middle reaches of the Heihe River have excellent light and heat conditions and are a major agricultural irrigation area, their hydrological processes are heavily influenced by the unified allocation of water resources within the basin. To ensure the ecological security and survival of downstream areas (especially the Populus euphratica forest in Ejina Banner), the runoff process in the middle reaches is not simply a matter of natural attrition, but is subject to strict artificial regulation and water allocation policies. Therefore, the actual runoff in the middle reaches depends not only on agricultural irrigation extraction but also on the scheduling instructions for ecological water transfer downstream. Based on the coupling characteristics of 'nature-society-ecology', in the 'natural-artificial' dual water cycle, the amount of water extracted for agricultural irrigation becomes one of the key factors affecting the runoff changes in the middle reaches of the Heihe River. It not only directly consumes the main stream resources but is also a core variable that must be considered when implementing ecological regulation. If this anthropogenic water consumption process cannot be accurately quantified, the model will struggle to capture the runoff fluctuation characteristics under the regulatory background. Therefore, accurately calculating agricultural irrigation water use and incorporating it into the model input is key to improving the accuracy of runoff prediction in the middle reaches of the Heihe River.

[0006] However, even with its strong learning ability, the LSTM model may still be unable to effectively capture the main features of hydrological processes when faced with temporal variations in runoff.

[0007] To overcome this challenge, signal decomposition techniques are used to break down the original hydrological sequence into multiple sub-modes, allowing for better observation of its data characteristics and enabling LSTM to better capture data patterns and improve model fitting capabilities. However, using a deep learning-based CNN-LSTM driving model can introduce uncertainty or inaccuracies in simulation results. In practical applications, it is necessary to identify driving factors highly correlated with the runoff sequence as input to the data-driven model, and combine this with variational mode decomposition for more accurate runoff prediction. Summary of the Invention

[0008] To address the aforementioned shortcomings of existing technologies, the purpose of this invention is to provide a method for simulating regional monthly runoff using a coupled data-driven model based on irrigation water consumption of major crops as a driving factor. This method fully utilizes runoff-related factors and combines them with a coupled data-driven model to achieve more accurate runoff prediction.

[0009] A method for simulating monthly runoff in a target region based on a coupled model of irrigation water usage for major crops, namely VMD-CNN-LSTM-Attention, includes the following steps:

[0010] 1. Calculate the daily reference crop evapotranspiration ET0 based on meteorological station data in the target area;

[0011] II. Calculate the irrigation water consumption of different crops at different stages using the crop growth coefficients of different crops;

[0012] Third, the monthly factors such as irrigation water consumption of major crops in the target area, as well as precipitation, temperature, evapotranspiration, and soil moisture content, calculated in step two, are used as driving factors and input into the coupled model for monthly runoff prediction.

[0013] As an optimized calculation method, meteorological station data, NDVI data, and soil moisture content data of the target area are acquired to calculate the daily reference crop evapotranspiration ET0. The calculation expression is as follows:

[0014]

[0015] In the formula: ET0 is the reference crop water requirement; R n G is the net radiation at the crop surface; T is the soil heat flux density; T is the average temperature 2m above the ground; u2 is the wind speed 2m above the ground; e s e is the saturated vapor pressure; a Δ is the actual water vapor pressure; Δ is the slope of the saturated water vapor temperature curve; γ is the hygrometer constant.

[0016] Furthermore, the accuracy of the data used in the calculation of the daily reference crop evapotranspiration ET0 is improved by converting the discrete observation data of meteorological stations into continuous spatial data.

[0017] Furthermore, the data used in the calculation of the daily reference crop evapotranspiration ET0 includes precipitation, maximum temperature, minimum temperature, average air temperature, humidity, sunshine duration, and wind speed.

[0018] As an optimized design method, the coupled model construction method includes: decomposing the original runoff sequence into five runoff pattern components, and using the calculated irrigation water volume of major crops in the target area, as well as monthly data elements such as precipitation, temperature, evapotranspiration, and soil moisture content as driving factors, and dividing them into training and test sets in a 7:3 ratio.

[0019] The present invention has the following beneficial effects:

[0020] Compared to single data-driven models like LSTM or CNN, this invention offers the following specific advantages: it calculates the irrigation water consumption of major crops that is highly correlated with runoff in the study area, allowing for the identification of driving factors with higher correlation to runoff as inputs, thereby improving prediction accuracy. Coupled with LSTM and CNN models and incorporating an attention mechanism, this invention achieves complementary advantages among various models, resulting in more accurate runoff simulation results. Attached Figure Description

[0021] Figure 1 Flowchart of VMD-CNN-LSTM-Attention, a coupled model based on irrigation water usage.

[0022] Figure 2 To calculate irrigation water consumption and input it as a driving factor into the coupled model

[0023] Figure 3 Flowchart of a monthly runoff prediction model based on VMD-CNN-LSTM-Attention Detailed Implementation

[0024] The basic technical concept of this invention is as follows: First, this method uses precipitation, temperature, and wind speed data from meteorological stations in the target area to calculate daily reference crop evapotranspiration. Combined with crop growth coefficients and effective rainfall, it accurately calculates irrigation water consumption for different growth stages of major crops. Then, variational mode decomposition (VMD) is used to decompose the original runoff sequence into multiple modal components to reduce data non-stationarity. The calculated irrigation water consumption, along with monthly factors such as precipitation, temperature, evapotranspiration, and soil moisture content, are used as driving factors and input into a coupled model integrating convolutional neural networks (CNN), long short-term memory networks (LSTM), and attention mechanisms for training and prediction. This invention fully considers the significant impact of agricultural irrigation on runoff in the "natural-social" dual water cycle. Through the complementary advantages of multi-source driving factors and deep learning models, it effectively improves the accuracy and stability of monthly runoff simulation under complex environments.

[0025] The present invention will now be described in further detail with reference to the accompanying drawings and examples. For example... Figure 1 As shown, this invention is a runoff simulation method based on irrigation water consumption of major crops as a driving factor, which includes the following steps:

[0026] I. Obtain basic data on the calculated irrigation water consumption of the target area, as well as other data that are highly correlated with runoff.

[0027] This means acquiring meteorological data, hydrological station runoff data, and NDVI and soil moisture data for the target area. Specifically, (1) meteorological station data for the study area are collected, including precipitation, maximum temperature, minimum temperature, average temperature, humidity, sunshine duration, and wind speed. (2) Measured runoff data from hydrological stations are acquired. (3) NDVI data reflecting crop growth and soil moisture data are acquired. Meteorological station data can be obtained from the National Meteorological Science Data Center. Runoff data originates from the water resources bureau where the target area is located. NDVI data can be obtained from datasets released by NASA for the target area, and soil moisture data can be obtained from the ERA5 dataset.

[0028] Based on the above data, the daily reference crop evapotranspiration ET0 is then calculated using the following expression:

[0029]

[0030] In the formula: ET0 is the reference crop water requirement; R n G is the net radiation at the crop surface; T is the soil heat flux density; T is the average temperature 2m above the ground; u2 is the wind speed 2m above the ground; e s e is the saturated vapor pressure; aΔ is the actual water vapor pressure; Δ is the slope of the saturated water vapor temperature curve; γ is the hygrometer constant.

[0031] Furthermore, the accuracy of the data used in the daily reference crop evapotranspiration (ET0) calculation (including precipitation, maximum temperature, minimum temperature, average air temperature, humidity, sunshine duration, wind speed, etc.) is improved by converting discrete observation data from meteorological stations into continuous spatial data. This method makes the calculated reference crop evapotranspiration more representative of the overall regional situation, and has the advantages of simple calculation, intuitive results, and improved spatial distribution while maintaining data accuracy.

[0032] II. Calculate the irrigation water consumption of different crops at different stages using the crop growth coefficients of different crops;

[0033] Sub-step 2-1: Determine the crop coefficient K for the relevant crops based on the main crop planting situation in the study area. C The growth stages of crops are divided into the initial stage, growth period, middle stage, and late stage. Specifically, based on the main crop planting conditions in the study area, the crop coefficient K of relevant crops can be found in the recommendations of the Food and Agriculture Organization of the United Nations. C .

[0034] Sub-step 2-2: Based on the crop coefficient K for different growth stages of each crop C The crop water requirement ET is calculated by comparing it with the reference crop evapotranspiration ET0. C As shown in equation (4):

[0035] ET c =K c ×ET o (4)

[0036] Where: ET C Daily crop water requirement, mm / d; K c ET0 is the crop coefficient; ET0 is the daily reference crop water requirement, mm / d.

[0037] Among them, based on the different growth stages of crops described in sub-step 2-1, the crop coefficient K is generally the initial stage. C The relative minimum, while the crop coefficient K during the mid-growth stage. C The maximum water requirement is used to determine that the water requirement of crops varies at different growth stages.

[0038] Sub-steps 2-3: Calculate the effective rainfall within the study area and compare it with the crop water requirement to determine the irrigation water requirement IR.

[0039]

[0040] In the formula: Pe P represents daily effective rainfall, in mm / d; P represents daily natural rainfall, in mm / d.

[0041]

[0042] Where IR represents irrigation water requirement, in mm; ET C The daily crop water requirement is expressed in mm. Effective rainfall refers to the portion of rainfall used to meet the crop's evapotranspiration needs; essentially, it's the water from natural precipitation used to replenish the soil around the crop roots. By comparing crop water requirements with rainfall, the effective rainfall within the target area is obtained, thus determining the irrigation water requirements for crops at different times.

[0043] Sub-step 2-4: Calculate the irrigation water demand (IR) for the main crops in the study area based on the irrigation water demand (IR) obtained in sub-step 2-3.

[0044]

[0045] Where η is the agricultural irrigation water utilization coefficient and W is the irrigation water consumption of the main crops. The irrigation water consumption of the main crops in the study area can be calculated. Since some areas are highly concentrated farming areas, the water consumption is particularly large. Therefore, the irrigation water consumption in the study area is calculated. This item is highly correlated with the measured runoff in the area, so it can be used as a feature factor for runoff prediction in the data-driven model to improve the prediction accuracy.

[0046] Third, the monthly factors such as irrigation water consumption of major crops in the target area, as well as precipitation, temperature, evapotranspiration, and soil moisture content, calculated in step two, are used as driving factors and input into the coupled model for monthly runoff prediction.

[0047] (I) Construction of the Coupled Model

[0048] As described in the background section, even with its strong learning ability, the LSTM model may still be unable to effectively capture the main features of hydrological processes when faced with temporal variations in runoff. To overcome this challenge, this invention utilizes signal decomposition technology to decompose the original hydrological sequence into multiple sub-modes, thereby enabling better observation of its data characteristics and allowing the LSTM to better capture data patterns and improve model fitting capabilities.

[0049] Furthermore, due to the insufficient ability of LSTM to extract latent features from nonlinear data, the prediction accuracy of LSTM decreases significantly with extended prediction periods. This invention employs a combination of VMD and CNN-LSTM for runoff prediction, which achieves relatively better results.

[0050] Variational Mode Decomposition (VMD) can effectively separate sub-modes and alleviate mode aliasing problems, and its performance is robust. Currently, many researchers have coupled VMD with LSTM models and optimized the coupled models. However, due to LSTM's insufficient ability to extract latent features from nonlinear data, its prediction accuracy significantly decreases with extended prediction periods. Therefore, a combination of VMD and CNN-LSTM is used for runoff prediction, which yields relatively better results.

[0051] Furthermore, predictive factors, as inputs to machine learning models, directly impact the accuracy of prediction results. If a data-driven model relies solely on its own historical observation data to infer future runoff processes, model construction and training depend heavily on learning the statistical variation characteristics of the runoff sequence. For highly nonlinear and non-stationary time series like runoff sequences, the predictive performance of such single-factor data-driven models remains unsatisfactory. By incorporating meteorological elements such as precipitation and temperature, data-driven runoff prediction models have shifted from single-factor to multi-factor inputs, learning the physical causes of runoff and meteorological elements, thus improving runoff forecast accuracy. Therefore, it is necessary to identify elements with high correlation to the runoff sequence as driving factors for runoff prediction.

[0052] Therefore, this invention proposes a coupled model, VMD-CNN-LSTM-Attention, to improve the accuracy of monthly runoff prediction in the middle reaches of the Heihe River. For example... Figure 3 As shown, the model first applies Variational Mode Decomposition (VMD) to preprocess the original runoff sequence, effectively reducing its non-stationarity and nonlinear characteristics. Next, it extracts features from the data using a CNN to compensate for the limitations of LSTM in feature extraction. Finally, an attention mechanism layer is added, which weights the LSTM output features by calculating time-step weight coefficients, thereby strengthening the model's focus on key temporal features and ultimately improving runoff prediction accuracy. Regarding data selection, this study selected factors highly correlated with monthly runoff in the middle reaches of the Heihe River, including soil moisture, precipitation, runoff at the Yingluoxia hydrological station in the upper reaches of the Heihe River, and the difference between this runoff and the irrigation water consumption of major crops in the middle reaches, as driving factors for the model. The model was applied to monthly runoff prediction at the Zhengyixia hydrological station in the middle reaches of the Heihe River, with data from 2000 to 2016 used for training and data from 2017 to 2020 used for evaluation.

[0053] In summary, LSTM in deep learning models can solve the long memory loss problem and has a significant effect on runoff prediction for long time series. However, due to LSTM's insufficient ability to extract latent features of nonlinear data, its prediction accuracy decreases significantly with the extension of the prediction period. Therefore, this invention couples CNN with LSTM to predict runoff, which can achieve relatively better results. At the same time, variational mode decomposition (VMD) is used to decompose the original runoff series, and signal decomposition techniques are used to decompose the original hydrological series into multiple sub-modes to better observe its data characteristics.

[0054] (II) Selection of Driving Factors

[0055] Predictive factors, as inputs to machine learning models, directly impact the accuracy of predictions. If a data-driven model relies solely on its own historical observation data to infer future runoff processes, model construction and training depend heavily on learning the statistical variation characteristics of runoff sequences. However, for highly nonlinear and non-stationary time series like runoff sequences, the predictive performance of such single-factor data-driven models remains unsatisfactory. By incorporating meteorological elements such as precipitation and temperature, data-driven runoff prediction models have shifted from single-factor to multi-factor inputs, learning the physical causes of runoff and meteorological elements, thus improving runoff forecast accuracy. Therefore, it is necessary to identify elements with high correlation to runoff sequences as driving factors for runoff prediction.

[0056] In the specific implementation process, firstly, multi-source data such as meteorological, agricultural, and soil data of the target area are acquired to construct a candidate factor set. Then, the correlation between each candidate factor and the runoff sequence is calculated using the Pearson correlation coefficient analysis method, and screening is conducted in conjunction with hydrological mechanisms to ensure that the selected factors have both statistical significance and physical rationality. For example, precipitation is the main source of runoff replenishment; temperature affects evapotranspiration and snowmelt processes; evapotranspiration directly affects water balance; soil moisture content controls runoff generation conditions; and agricultural irrigation water consumption alters the regional water resource allocation structure. By screening key driving factors highly correlated with runoff through the above steps, and then using these as model inputs, the model's ability to characterize runoff variation patterns and its prediction accuracy can be effectively improved. This invention selects the main crop irrigation water consumption in the target area, as well as monthly elements such as precipitation, temperature, evapotranspiration, and soil moisture content, as driving factors, and inputs them together into a coupled model for monthly runoff prediction. In summary, existing technologies using deep learning-based CNN-LSTM driving models can introduce uncertainty or inaccuracies in simulation results. Practical applications require identifying driving factors highly correlated with runoff sequences as input to the data-driven model, and combining this with variational mode decomposition (VMD) for more accurate runoff prediction. While LSTM addresses the long memory loss problem in deep learning models and demonstrates significant effectiveness in predicting long-term runoff series, its limited ability to extract latent features from nonlinear data leads to a significant decrease in prediction accuracy as the lead time extends. This invention couples CNN and LSTM for runoff prediction, achieving relatively better results. Furthermore, VMD is used to decompose the original runoff sequence, and signal decomposition techniques are employed to break down the original hydrological sequence into multiple sub-modes for better observation of its data characteristics.

[0057] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for simulating monthly runoff in a target region based on a coupled model of irrigation water consumption for major crops (VMD-CNN-LSTM-Attention), characterized in that... It includes the following steps:

1. Calculate the daily reference crop evapotranspiration ET0 based on meteorological station data in the target area; II. Calculate the irrigation water consumption of different crops at different stages using the crop growth coefficients of different crops; Third, the monthly factors such as irrigation water consumption of major crops in the target area, as well as precipitation, temperature, evapotranspiration, and soil moisture content, calculated in step two, are used as driving factors and input into the coupled model for monthly runoff prediction.

2. The method for simulating regional runoff based on the coupled model VMD-CNN-LSTM-Attention of irrigation water volume for major crops as described in claim 1, characterized in that: Obtain meteorological station data, NDVI data, and soil moisture data for the target area, and calculate the daily reference crop evapotranspiration ET0 using the following expression: In the formula: ET0 is the reference crop water requirement; R n G is the net radiation at the crop surface; T is the soil heat flux density; T is the average temperature 2m above the ground; u2 is the wind speed 2m above the ground; e s e is the saturated vapor pressure; a Δ is the actual water vapor pressure; Δ is the slope of the saturated water vapor temperature curve; γ is the hygrometer constant.

3. The method for simulating regional runoff based on the coupled model VMD-CNN-LSTM-Attention of major crop irrigation water consumption according to claim 2, characterized in that, The accuracy of the data used in the calculation of the daily reference crop evapotranspiration ET0 is improved by converting the discrete observation data of meteorological stations into continuous spatial data.

4. The method for simulating regional runoff based on the coupled model VMD-CNN-LSTM-Attention of major crop irrigation water consumption according to claim 3, characterized in that, The data used in the calculation of the daily reference crop evapotranspiration ET0 includes precipitation, maximum temperature, minimum temperature, average temperature, humidity, sunshine duration, and wind speed.

5. The method for simulating monthly runoff in a target region based on a coupled model of irrigation water consumption for major crops, VMD-CNN-LSTM-Attention, as described in claim 1, is characterized in that... The method for constructing the coupled model includes: decomposing the original runoff sequence into five runoff pattern components, and using the calculated irrigation water volume of major crops in the target area, as well as monthly data elements such as precipitation, temperature, evapotranspiration, and soil moisture content, as driving factors, and dividing them into training and test sets in a 7:3 ratio.

Images