Basin runoff and sediment transport prediction method based on machine learning

Abstract

The invention provides a watershed runoff and sediment transport prediction method based on machine learning, and belongs to the technical field of hydraulics and river dynamics in natural science research. The method comprises the following steps: acquiring historical runoff data of a previous time period N; then preprocessing is carried out to form prediction model standard input data; inputting the preprocessed data into the trained drainage basin runoff prediction model, and outputting normalized runoff data with the time interval of one month in M periods in the future; then carrying out reverse normalization processing to obtain monthly runoff data with the time interval of one month in the future M periods; and constructing a power function of a water-sediment relation curve of the river, and calculating the sediment transport rate according to runoff data output by the model so as to calculate sediment transport data of the future M periods. The data size adopted by the method is far smaller than that adopted by a method based on a physical change model, and effective prediction can be achieved even under the condition that runoff and sediment transportation are extremely changed.

Description

Technical Field

[0001] The present invention belongs to the technical field of hydraulics and river dynamics in natural science research, and particularly relates to a method for predicting basin runoff and sediment transport based on machine learning. Background Art

[0002] In the river system, during the process of runoff and sediment transport from land to the ocean, it regulates the role of the global biogeochemical cycle and has a significant impact on the structure and function of the ecosystem. However, the runoff and sediment transport of most rivers in the world have started to decrease sharply, posing a threat to wetlands and coastlines. In recent decades, under the influence of human activities, the sediment transport of many rivers has decreased significantly. Due to the influence of human factors on river runoff and sediment transport, the changes in river runoff and sediment transport are unstable. However, runoff and sediment transport have an important impact on water and sediment management and ecological governance within the basin. Therefore, the prediction of runoff and sediment transport within the basin has become particularly important. The prediction of runoff and sediment transport within the basin is of great significance for water resource management and reservoir operation.

[0003] Currently, the widely adopted prediction models can be divided into two types: physics-based models and data analysis-based models. Physics-based models can effectively reflect the physical changes in the simulation process, but they are restricted by various assumptions and require a large amount of data support. The data analysis-based model is the most widely used method currently. This method is based on the observation of historical data for simulation, which is simple and easy to implement, does not require data on physical changes, and requires far less data than physics-based models, so it has a wider applicability. Among them, multiple linear regression and autoregressive moving average (AMRA) are commonly used prediction methods in hydrology. Zhang et al. (2016) found through research that when the time series variables show linear or near-linear characteristics, these methods can provide satisfactory prediction results. However, when facing non-linear and non-stationary variables in the time series, these methods often have difficulty effectively capturing hidden features, so it is difficult to obtain satisfactory prediction results.

[0004] Time series such as runoff and sediment transport usually exhibit highly dynamic complexity and non-stationarity characteristics. It is difficult to meet the requirements of reasonable prediction using a single regression prediction method. Therefore, there is an urgent need for a method that can accurately predict the changes in runoff and sediment transport in the basin and provide a theoretical basis for water and sediment management within the basin. Summary of the Invention

[0005] Aiming at the above problems, the present invention proposes a method for predicting basin runoff and sediment transport based on a machine learning model, which can predict the runoff and sediment transport in the basin for a future period of time and provide a theoretical basis for water and sediment management within the basin.

[0006] The present invention provides a method for predicting basin runoff and sediment transport based on machine learning, comprising the following steps:

[0007] Step 1, obtain the historical runoff data of the previous time period N, where N is the periodic data with a time interval of 1 month;

[0008] Step 2, preprocess the historical runoff data to form the standard input data of the prediction model, and the preprocessing is normalization processing;

[0009] Step 3, input the preprocessed data into the trained basin runoff prediction model, and output the normalized runoff data with a time interval of 1 month within the next M periods; the basin runoff prediction model includes a convolutional neural network module and a long short-term memory module;

[0010] Step 4, perform denormalization processing on the obtained normalized runoff data with a time interval of 1 month within the next M periods to obtain the monthly runoff data with a time interval of 1 month within the next M periods;

[0011] Step 5, construct a power function of the water-sediment relationship curve of the river, and calculate the variables c and d of the severity and erosion capacity of the river sediment erosion based on the historical runoff data and historical sediment transport data of the previous time period N; then calculate the sediment transport rate according to the runoff data output by the model, so as to calculate the sediment transport data with a time interval of 1 month within the next M periods.

[0012] Preferably, the historical runoff data input into the basin runoff prediction model is one-dimensional time series data of N years with a time interval of 1 month, and the data output by the basin runoff prediction model is one-dimensional time series data of M years with a time interval of 1 month. The period N and the period M are continuous time series, and the proportion of the period N in the period N + M is 70% - 80%, and the proportion of the period M in the period N + M is 20% - 30%.

[0013] Preferably, the basin runoff prediction model includes a convolutional neural network module and a long short-term memory module; the convolutional neural network module includes a convolutional layer and a global average pooling layer; the convolutional layer is used to extract runoff time series features, and the global average pooling layer converts the two-dimensional runoff time series features extracted by the convolutional layer into one-dimensional runoff time series features and passes them to the long short-term memory module; the long short-term memory module is composed of two long short-term memory LSTM layers; two layers of long short-term memory networks (LSTM) are used to process the one-dimensional runoff time series data to achieve runoff prediction. LSTM efficiently screens and processes the input runoff data through its unique input gate, forget gate, and output gate structures. The input gate is responsible for evaluating the importance of the current input runoff data and determining the degree to which it is written into the memory unit, thereby screening out valuable feature information for prediction. The forget gate controls the forgetting process of the initial runoff data, precisely screening out the information that needs to be discarded in the model to avoid interference from redundant data on the prediction results. The output gate determines the output value based on the current model state and outputs the new runoff data at the current moment. After two layers of deep processing by LSTM, the model can finally generate normalized runoff prediction data within the next M cycles with a time interval of 1 month.

[0014] Preferably, the specific processing process of the basin runoff prediction model is as follows:

[0015] Expand the runoff data into one-dimensional data, and then input a series of runoff data k = 1,....., N through the convolutional layer h K into the convolutional kernel L×1 to achieve data screening and feature extraction; the convolutional kernel processes the data according to the following equation:

[0016]

[0017] where f is a non-linear activation function, W i is the weight of the layer, b is the bias parameter, i is the number of layers included in the convolutional layer, the convolutional layer includes two layers, the upper convolutional layer, the convolutional kernel is 3×1, the stride is 1×1, and the number of channels is 32; the lower convolutional layer, the convolutional kernel is 3×1, the stride is 1×1, and the number of channels is 64; after the data completes feature extraction in the convolutional layer, the global average pooling layer converts the runoff time series features extracted by the convolutional layer into one-dimensional time series and passes them to the long short-term memory module;

[0018] The LSTM layer consists of three gates: an input gate, a forget gate, and an output gate. The input gate determines the degree of influence of the current input, that is, how much new information is written into the memory cell. The forget gate controls the forgetting of old information, that is, determines which information in the model system needs to be discarded. The output gate determines the output value according to the current model state and outputs the important information at the current moment. The LSTM layer works according to formulas (2)-(7): Forgetting stage: LSTM first determines how much information to forget from the previous moment's memory through the forget gate. Input stage: Through the input gate, LSTM determines how much the current input affects the memory state. State update: Based on the results of the first two steps, LSTM updates its memory cell. Output stage: Through the output gate, it determines the influence of the current state on the final output.

[0019] Forget gate f t = σ(W f x t + U f h t-1 + b f )(2)

[0020] Input gate i t = σ(W i x t + U i h t-1 + b i )(3)

[0021] Output gate o t = σ(W o x t + U o h t-1 + b o )(4)

[0022] Hidden state h t = tanh(c t ) ⊙ o t (5)

[0023] Memory cell

[0024] Potential memory cell

[0025] Among them, σ(·) is the sigmoid function, tanh(·) is the hyperbolic tangent function, ⊙ represents element-wise multiplication, W f , U f , b f are the coefficients of the forget gate, W i , U i , b i are the coefficients of the input gate, is the coefficient of the potential unit state, W o , U o , b o are the coefficients of the output gate, f t is the runoff data for forgetting, i t is the runoff data input to the input long short-term memory module, o t is the output runoff data, h t is the runoff data that needs to be hidden for backup, c t is the memory cell, is the potential memory cell.

[0026] Preferably, the lag time q is determined to be 12 and the prediction time step p is 1. Specifically, the monthly runoff data for the past 12 years is used to predict the runoff data for the corresponding month of the next year. For example, based on the runoff data for a certain month from 1960 to 1971, through model analysis and calculation, the runoff data for the corresponding month in 1972 is predicted.

[0027] Preferably, the specific process of step 5 is as follows:

[0028] Use the historical runoff data and historical sediment transport data with a time period of N and a time interval of 1 month to construct a water-sediment relationship curve:

[0029] logQ S = Logc + dlogQ (8)

[0030] where Q S is the sediment transport rate, kg / s, Q is the flow rate, m 3 / s, c and d are variables, representing the severity of sediment erosion and erosion capacity of the river respectively, and the variables c and d are solved through the runoff and sediment transport data with a time period of N;

[0031] Predict the sediment transport data for the next M cycles based on the runoff data for the next M cycles predicted by the runoff prediction model; use the runoff data for the next M cycles as Q and input it into equation (8) to obtain the runoff data Q S .

[0032] Preferably, the specific processes of normalization in step 2 and denormalization in step 4 are as follows:

[0033] Normalization process:

[0034]

[0035] where x is the original runoff data, x min and x max are the minimum and maximum values in the runoff data set respectively. The normalized value x norm is between 0 and 1;

[0036] Inverse normalization process:

[0037] y = x norm ×(x max - x min ) + x min

[0038] y is the predicted runoff data.

[0039] Compared with the prior art, the present invention has the following beneficial effects:

[0040] (1) When dealing with non-linear and non-stationary variables in time series, the present invention can overcome the disadvantage of weak linear relationship in time series. The amount of data used in the present invention is much smaller than that of the method based on physical change model, and the data used in the present invention is only the historical runoff and sediment transport data within the basin.

[0041] (2) The method of combined prediction of machine learning and water-sediment relationship adopted by the present invention shows excellent performance in prediction accuracy. Even in the case of extreme changes in runoff and sediment transport, effective prediction can be achieved.

[0042] (3) The method of combined prediction of machine learning and water-sediment relationship adopted by the present invention makes up for the defect that machine learning cannot effectively predict the sediment transport within the basin. Description of the Drawings

[0043] Figure 1 is a general situation map of the Jialing River Basin studied by the present invention.

[0044] Figure 2 is a flow chart of the CLSTM prediction process.

[0045] Figure 3 is a monthly runoff prediction map of the Jialing River Basin.

[0046] Figure 4 is a monthly sediment transport prediction map of the Jialing River Basin predicted by the water-sediment relationship curve.

[0047] Figure 5 is a prediction performance map of the sediment transport data of the hydrological stations in the Jialing River Basin by the model. Detailed Embodiments

[0048] The following further illustrates the specific implementation process of the present invention in combination with specific embodiments.

[0049] The present invention takes the water systems controlled by Beibei Station, Wusheng Station, Luoduxi Station and Xiaohe Station in the Jialing River Basin as an example ( Figure 1)。The Jialing River is located in the central and western regions of China. It originates from the southern foot of the Qinling Mountains in Shaanxi Province and is a tributary on the left bank of the upper reaches of the Yangtze River Basin. The main stream of the Jialing River is about 1,190 kilometers long, and the basin area is about 160,000 square kilometers. 2 。The Jialing River is a key sediment-producing river in the Yangtze River, and its sediment content ranks first among all tributaries. The water and sediment in the Jialing River Basin mainly come from three major river systems, namely the main stream of the Jialing River, and two tributaries, the Qu River and the Fu River. In this study, the runoff and sediment transport data of Beibei Station, Wusheng Station, Luoduxi Station, and Xiaoheba Station in the Jialing River Basin were used to analyze the temporal variation laws of runoff and sediment transport in the Jialing River Basin. Among them, Beibei Station is the outlet control hydrological station in the Jialing River Basin, Wusheng Station is the hydrological station on the main stream of the Jialing River, Luoduxi Station is the outlet control hydrological station in the Qu River Basin, a tributary of the Jialing River, and Xiaoheba Station is the outlet control hydrological station in the Fu River Basin, a tributary of the Jialing River.

[0050] The present invention collected the annual and monthly data of runoff and sediment transport of Beibei Station, Wusheng Station, Luoduxi Station, and Xiaoheba Station in the Jialing River Basin (Table 1). The annual and monthly data of runoff and sediment transport are from the "Yangtze River Sediment Bulletin" and the "China's Main River Sediment Bulletin" compiled by the Hydrology Bureau of the Yangtze River Water Resources Commission.

[0051] Table 1 Data information of 4 hydrological stations in the Jialing River Basin

[0052]

[0053] The present invention combines the long short-term memory model (LSTM) and the convolutional neural network (CNN) to form a CLSTM prediction method. This method includes five modules: the historical runoff data preprocessing module, the model parameter determination module, the convolutional neural network (CNN) module, the long short-term memory model (LSTM) module, and the water-sediment memory module.

[0054] Among them, the historical runoff data preprocessing module is mainly responsible for classifying and normalizing the historical runoff data to provide a standardized data basis for subsequent analysis. The core task of the model parameter determination module is to determine the lag time q and the prediction time step p to ensure that the model can accurately capture the temporal correlation of the data.

[0055] In the convolutional neural network module, multiple convolutional layers are stacked to effectively capture the temporal features of the variables. The long short-term memory module contains two LSTM layers, and the top layer is a fully connected layer, which is responsible for extracting the temporal features of the data captured by the CNN. Subsequently, the historical runoff data is trained and predicted through two LSTM layers to accurately predict the runoff in the Jialing River Basin for a period of time in the future.

[0056] Preferably, in the water-sediment memory module, the prediction of the sediment transport volume in the Jialing River Basin for a period of time in the future is realized by finding the change relationship between the runoff volume and the sediment transport volume in the Jialing River Basin.

[0057] The method for predicting runoff by CLSTM is as Figure 2 shown. The prediction is carried out according to the following process:

[0058] 1. In the historical runoff data preprocessing module, the monthly runoff data is used as the input variable for prediction. First, the historical monthly runoff data is divided into a training set and a test set. Specifically, the monthly runoff data for the most recent 10 years (from 2014 to 2023) is used as the test set, while the monthly data for the remaining years (from 1960 to 2013) constitutes the training set. Then, the monthly runoff data in the training set is normalized to improve the training efficiency and stability of the model, while the runoff data in the test set remains unchanged for subsequent model evaluation. Finally, the monthly runoff data in the training set is unfolded into one-dimensional data to provide a suitable input format for the convolutional neural network module.

[0059] 2. According to the variation law of the monthly runoff data, the lag time q is determined to be 12 and the prediction time step p is 1. Specifically, the monthly runoff data for the past 12 years is used to predict the runoff data for the corresponding month of the next year. The model will perform cyclic trial calculations and fittings according to this process to achieve accurate prediction of future runoff volumes.

[0060] 3. After the data preprocessing is completed, the data enters the convolutional neural network module. The convolutional neural network (CNN) mainly consists of two types of layers, namely convolutional layers and pooling layers, which work together to process the data. The main task of the convolutional layer is to learn the feature representation of the input data, and it consists of multiple convolutional kernels. Before passing the data to the long short-term memory module, the output of the convolutional layer is processed by a non-linear activation function to enhance the expressive ability of the model. In this method, the runoff data is unfolded into one-dimensional data, and then a series of runoff data k = 1,....., N K is input into the convolutional kernel (L×1) to achieve data screening and feature extraction. The convolutional kernel processes the data according to the following equation:

[0061]

[0062] where f is the non-linear activation function, W i is the weight of the layer, b is the bias parameter, i is the number of layers included in the convolutional layer, and in the present invention, the convolutional layer includes two layers. For the upper convolutional layer, the convolutional kernel is (3×1), the stride is (1×1), and the number of channels is 32; for the lower convolutional layer, the convolutional kernel is (3×1), the stride is (1×1), and the number of channels is 64. After the data completes feature extraction in the convolutional layer, it enters the global average pooling layer of the convolutional neural network. The global average pooling layer converts the runoff time series features extracted by the convolutional layer into a one-dimensional time series and passes it to the long short-term memory module.

[0063] 4. The long short-term memory module consists of two layers of long short-term memory (LSTM) layers. The upper LSTM layer is responsible for connecting to the convolutional neural network module and receiving data from the pooling layer of the convolutional neural network. Then, these data are further processed and analyzed through the two LSTM layers. LSTM is a special type of recurrent neural network (RNN). By introducing "memory cells" and gating mechanisms, LSTM allows the network to selectively remember or forget information, thus alleviating the deficiencies of RNN in dealing with long-term dependence problems. The core of LSTM lies in its gating mechanism, which mainly includes three gates: the input gate, the forget gate, and the output gate. The input gate determines the degree of influence of the current input, that is, how much new information is written into the memory cell; the forget gate controls the forgetting of old information, that is, determines which information in the model system needs to be discarded; the output gate determines the output value according to the current model state and outputs the important information at the current moment. LSTM works according to equations (2)-(7): Forgetting stage: LSTM first determines how much information to forget from the previous moment's memory through the forget gate; Input stage: Through the input gate, LSTM determines how much the current input affects the memory state; State update: According to the results of the first two steps, LSTM updates its memory cell; Output stage: Through the output gate, it determines the influence of the current state on the final output.

[0064] Forget gate f t = σ(W f x t + U f h t-1 + b f ) (2)

[0065] Input gate i t = σ(W i x t + U i h t-1 + b i ) (3)

[0066] Output gate o t = σ(W o x t + U o h t-1 + b o ) (4)

[0067] Hidden state h t = tanh(c t ) ⊙ o t (5)

[0068] Memory cell

[0069] Potential memory cell

[0070] Among them, σ(·) is the sigmoid function, tanh(·) is the hyperbolic tangent function, ⊙ represents element-wise multiplication, and W f , U f , and b f are the coefficients of the forget gate, W i , U i , and b i are the coefficients of the input gate, are the coefficients of the latent cell state, W o , U o , and b o are the coefficients of the output gate, f t is the runoff data to be forgotten, i t is the runoff data of the input long short-term memory module, o t is the output runoff data, h t is the runoff data that needs to be hidden for backup, c t is the memory cell, is the latent memory cell. Through the historical runoff data preprocessing module, model parameter determination module, convolutional neural network (CNN) module, and long short-term memory model (LSTM) module in the present invention, the monthly runoff data of the middle Jialing River Basin in the future for a period of time can be predicted. The prediction of the sediment transport sequence in the Jialing River Basin is realized in the water-sediment memory module.

[0071] 5. In the water-sediment memory module, the water-sediment relationship curve is used to find the potential relationship between the runoff and sediment load in the Jialing River Basin. The water-sediment relationship curve (SRC) of a river is a curve that describes the relationship between the runoff and sediment load at a certain cross-section of the river (Morgan, 1995; Asselman, 2000). The water-sediment relationship curve is presented in the form of a power function:

[0072] logQ S = Logc + dlogQ (8)

[0073] Among them, Q S is the sediment transport rate, kg / s, Q is the flow rate, m 3 / s, c and d are variables, representing the severity of sediment erosion and erosion capacity of the river, respectively. When using the water-sediment relationship curve to predict the sediment load in the Jialing River Basin, the sediment data of the most recent 10 years are also used as the test set to verify the accuracy of the model. To ensure that the sediment data volume is sufficient, the monthly data of the sediment time series in the Jialing River Basin are used for analysis. Table 2 shows the power function equation obtained by fitting the water-sediment relationship curve for verifying the accuracy of the model.

[0074] Table 2 Power function equation fitted by the water-sediment relationship curve SRC

[0075] Hydrological station Equation <![CDATA[R 2 > Beibei <![CDATA[logQ s =-7.10 + 2.88logQ]]> 0.89 Wusheng <![CDATA[logQ s =-7.54 + 3.37 logQ]]> 0.80 Luoduxi <![CDATA[logQ s =-5.70 + 2.73logQ]]> 0.86 Xiaoheba <![CDATA[logQ s =-6.94 + 3.23 logQ]]> 0.85

[0076] The present invention uses the root mean square error (RMSE), Nash-Sutcliffe model efficiency coefficient (NSE), and mean relative error (MRE) to evaluate the performance of the model used, and their definitions are as follows:

[0077]

[0078] where n is the number of data, y pred is the predicted value, y real is the true value, is the average value of the true values.

[0079] The prediction of runoff using CLSTM is shown in Figure 3 . The predicted monthly runoff of Beibei Station, Wusheng Station, Luoduxi Station, and Xiaoheba Station is basically consistent with the actual monthly runoff, and the prediction effect is excellent. The NSE values of the predicted monthly runoff and the actual monthly runoff of Beibei Station, Wusheng Station, Luoduxi Station, and Xiaoheba Station are all close to the 1:1 line, and the data distribution is relatively concentrated. The predicted values are consistent with the actual values, verifying the prediction reliability of the CLSTM method. The evaluation of the monthly runoff prediction results of the four hydrological stations in the Jialing River Basin is summarized in Table 3. The evaluation indicators RMSE, NSE, and MRE of using CLSTM to predict monthly runoff are 1608841747, 0.92, and 0.22 respectively at Beibei Station, 808177381.3, 0.84, and 0.26 respectively at Wusheng Station, 901381511.5, 0.84, and 0.24 respectively at Luoduxi Station, and 526516100.8, 0.86, and 0.23 respectively at Xiaoheba Station.

[0080] Table 3 Prediction performance of three machine learning methods used in the Jialing River Basin

[0081] Hydrological station Models RMSE NSE MRE Beibei CLSTM 1608841747 0.92 0.22 Wusheng CLSTM 808177381.3 0.84 0.26 Luoduxi CLSTM 901381511.5 0.84 0.24 Xiaoheba CLSTM 526516100.8 0.86 0.23

[0082] The sediment transport volume of the Jialing River Basin predicted by the water-sediment relationship curve method is as Figure 4 shown. The predicted monthly sediment transport volume of Beibei Station, Wusheng Station, Luoduxi Station, and Xiaoheba Station is basically consistent with the actual monthly sediment transport volume. In the peak months of monthly sediment transport volume, there are slight deviations between the predicted values and the actual values, but the overall prediction effect is good. When using the water-sediment relationship curve to predict monthly sediment transport volume, the trend change of monthly sediment transport volume can be effectively simulated, basically meeting the requirements of prediction. The model evaluation results of the four hydrological stations in the Jialing River Basin are summarized in Figure 5Among them, the NSE values of the predicted monthly sediment transport volumes and the actual monthly sediment transport volumes at Beibei Station, Wusheng Station, Luoduxi Station, and Xiaoheba Station are all close to the 1:1 line, and the data are relatively concentrated. The predicted values are relatively close to the actual values, verifying the reliability of the prediction of the water-sediment relationship curve. Among them, the NSE values of the predicted monthly sediment transport volumes and the actual monthly sediment transport volumes at the four hydrological stations using the water-sediment relationship curve are all between 0.92 and 0.97, and the MREs are all below 0.38. This indicates that it is feasible to predict the sediment transport volume of the Jialing River Basin using the water-sediment relationship curve in this study.

[0083] The above are only the preferred embodiments of the present application and are not used to limit the present application. For those skilled in the art, various changes and modifications can be made to the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present application shall be included within the protection scope of the present application.

[0084] Although the specific implementation manners of the present invention have been described above, they are not limitations on the protection scope of the present invention. Those skilled in the art should understand that, based on the technical solutions of the present invention, various modifications or deformations that can be made without creative efforts by those skilled in the art are still within the protection scope of the present invention.

Claims

1. A method for predicting watershed runoff and sediment transport based on machine learning, characterized in that: The following steps are involved: Step 1, obtain historical runoff data with a time period of N, where N is periodic data with a time interval of 1 month; Step 2, preprocessing the historical runoff data to form standard input data for the prediction model, the preprocessing includes normalization processing; Step 3, inputting the preprocessed data into the trained watershed runoff prediction model, and outputting normalized runoff data with a time interval of 1 month in the future M cycles; the watershed runoff prediction model includes a convolutional neural network module and a long short-term memory module; Step 4, performing denormalization processing on the normalized runoff data with a time interval of 1 month in the future M cycles, and obtaining the monthly runoff data with a time interval of 1 month in the future M cycles; Step 5, construct the power function of the water-sediment relationship curve of the river, and calculate the variables c and d of the river's sediment erosion severity and erosion capacity based on the historical runoff data of the previous time period N; then calculate the sediment transport rate based on the runoff data output by the model, and thus calculate the sediment transport data with a time interval of 1 month in the future M periods.

2. A method for predicting watershed runoff and sediment transport based on machine learning as claimed in claim 1, characterized in that: The period N and the period M are continuous time series, and the period N accounts for 70%-80% of the period N+M, and the period M accounts for 20%-30% of the period N+M.

3. A method for predicting watershed runoff and sediment transport based on machine learning as claimed in claim 1, characterized in that: The basin runoff prediction model includes a convolutional neural network module and a long short-term memory module; the convolutional neural network module includes a convolution layer and a global average pooling layer; the convolution layer is used to extract the runoff time series features, and the global average pooling layer converts the two-dimensional runoff time series features extracted by the convolution layer into one-dimensional runoff time series features and passes them to the long short-term memory module; the long short-term memory module is composed of two layers of long short-term memory LSTM layers; a two-layer long short-term memory network LSTM is used to process the one-dimensional runoff time series data to realize runoff prediction; LSTM uses its unique input gate, forget gate and output gate The structure can efficiently screen and process the input runoff data; the input gate is responsible for evaluating the importance of the current input runoff data, determining the degree to which it is written into the memory unit, and screening out the feature information that is valuable for prediction; the forgetting gate controls the forgetting process of the initial runoff data, screening out the information that needs to be discarded in the model, and avoiding the interference of redundant data on the prediction results; the output gate determines the output value according to the current model state, and outputs the new runoff data at the current moment; after deep processing by two layers of LSTM, the model can finally generate normalized runoff prediction data within the future M cycles with a time interval of 1 month.

4. A method for predicting watershed runoff and sediment transport based on machine learning as claimed in claim 1 or 3, characterized in that: The specific processing process of the basin runoff prediction model is as follows: The runoff data is expanded into one-dimensional data, and then a series of runoff data k=1, ..., N is transformed into one-dimensional data through the convolution layer h. K Input into the convolution kernel L×1 to achieve data screening and feature extraction; the convolution kernel processes the data according to the following equation: Among them, f is a nonlinear activation function, W i is the weight of the layer, b is the bias parameter, i is the number of layers contained in the convolution layer, and the convolution layer contains two layers. The upper convolution layer has a convolution kernel of 3×1, a step size of 1×1, and a number of channels of 32; the lower convolution layer has a convolution kernel of 3×1, a step size of 1×1, and a number of channels of 64. After the data is extracted from the convolution layer, the global average pooling layer converts the two-dimensional runoff time series features extracted by the convolution layer into one-dimensional runoff time series features and passes them to the long short-term memory module; The LSTM layer consists of three gates: input gate, forget gate and output gate. The input gate determines the influence of the current input, that is, how much new information is written into the memory cell. The forget gate controls the forgetting of old information, that is, determines which information in the model system needs to be discarded. The output gate determines the output value according to the current model state and outputs the important information at the current moment. The LSTM layer works according to formulas (2)-(7): Forgetting stage: LSTM first determines how much information to forget from the memory of the previous moment through the forgetting gate. Input stage: Through the input gate, LSTM determines how much influence the current input has on the memory state. State update: According to the results of the first two steps, LSTM updates its memory unit. Output stage: Determines the influence of the current state on the final output through the output gate. Forget Gate t =σ(W f x t +U f h t-1 +b f ) (2) Input Gate i t =σ(W i x t +U i h t-1 +b i ) (3) Output gate o t =σ(W o x t +U o h t-1 +b o ) (4) Hidden State Memory Unit Latent memory unit Among them, σ(·) is the sigmoid function, tanh(·) is the hyperbolic tangent function, represents element-wise multiplication, W f , U f 、b f is the coefficient of the forget gate, W i , U i 、b i is the coefficient of the input gate, is the coefficient of the potential unit state, W o , U o 、b o is the coefficient of the output gate, f t is the forgotten runoff data, i t is the runoff data input to the long short-term memory module, o t is the output runoff data, h t To hide the spare runoff data, c t For the memory unit, A potential memory unit.

5. A method for predicting watershed runoff and sediment transport based on machine learning as claimed in claim 4, characterized in that: Determine the lag time q as 12 and the prediction time step p as 1; use the monthly runoff data of the past 12 years to predict the runoff data of the corresponding month of the next year.

6. A method for predicting watershed runoff and sediment transport based on machine learning as claimed in claim 1, characterized in that: The specific process of step 5 is as follows: The water-sediment relationship curve is constructed using historical runoff data with a time period of N and a time interval of 1 month and historical sediment transport data: logQ S =Logc+dlogQ (8) Among them, Q S is the sediment transport rate, kg / s, Q is the flow rate, m 3 / s, c and d are variables, representing the severity and erosion capacity of the river's sediment erosion, respectively. The variables c and d are solved by the runoff and sediment transport data with a time period of N; The runoff data in the future M cycles are predicted by the runoff prediction model to predict the sediment transport data in the future M cycles; the runoff data in the future M cycles are input as Q into equation (8) to obtain the runoff data Q in the future M cycles S .

7. The method for predicting watershed runoff and sediment transport based on machine learning according to claim 1, characterized in that: The normalization process is: Among them, x is the original runoff data, x min and x max are the minimum and maximum values ​​in the runoff dataset respectively; the normalized value x norm Between 0 and 1; Denormalization process: y=x norm ×(x max -x min )+x min y is the predicted runoff data.

Images