An unmeasured basin runoff prediction method based on space modeling and improved LSTM

Abstract

The application discloses an unmeasured basin runoff prediction method based on space modeling and improved LSTM, and comprises the following steps: analyzing static characteristic data of a target basin and introducing prior knowledge; performing space modeling on the unmeasured basin, and dividing data of an original measuring station into a plurality of sub-basins according to a watershed in data processing; inputting these gridded meteorological forcing data into a network model of the prediction method as independent variables; building an LSTM network model with a static input gate, which is used for predicting runoff of the target basin; optimizing network model training by using a Nash efficiency function and an Adam strategy; applying the trained network model to the target basin to predict runoff, and calibrating the prediction result by using a SAC-SMA model. The method has high precision, strong physical interpretability and strong parameter transferability.

Description

Technical Field

[0001] This invention relates to the field of water conservancy big data technology, specifically to a method for predicting runoff in unmeasured watersheds based on spatial modeling and improved LSTM. Background Technology

[0002] Generally, data-driven modeling for runoff forecasting relies heavily on a large data base, and data scarcity has long been a significant obstacle to the development of data-driven runoff forecasting. Most river basins worldwide suffer from a lack of or inadequate measurements, and even in some developed countries, the number of measured river basins is decreasing. In China, frequent floods in recent years have caused enormous losses to people's lives and livelihoods. Especially in some remote areas, the incomplete construction of hydrological observation station networks and the scarcity of hydrological data make flood early warning work extremely difficult. Therefore, conducting runoff forecasting in unmeasured river basins is of great significance for disaster prevention and mitigation.

[0003] Existing neural runoff prediction methods are mainly based on deep learning models such as LSTM and Transformer. Chinese patent application (CN1114372631A) "A runoff prediction method for data-scarce areas based on few-shot learning and LSTM" uses few-shot learning + LSTM network model to predict runoff in data-scarce areas. This algorithm is suitable for watersheds with limited data for runoff prediction. However, when the watershed has not been measured, this method has some shortcomings, mainly in the following aspects: (1) no prior knowledge is introduced, and the prediction results of the network model lack interpretability; (2) the watershed data is not gridded, which is not conducive to the transfer learning of the network model; (3) the LSTM network model used is not improved for unmeasured areas, cannot fully utilize prior knowledge, has low prediction accuracy, and is prone to overfitting. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a method for predicting runoff in unmeasured watersheds based on spatial modeling and an improved LSTM. The method includes: analyzing the static characteristics of the target watershed and introducing prior knowledge to give the network model a hydrological foundation; performing gridding and noise reduction on runoff data from watersheds near the target watershed to obtain a spatial runoff dataset for network model training and testing; constructing an LSTM network model with a static input gate to enhance its ability to learn prior knowledge in the hydrological domain; using an improved Nash efficiency as the training objective function and optimizing the network model training using the Adam strategy; and applying the trained LSTM network model to predict runoff in the target (unmeasured) watershed. This invention offers high accuracy, strong physical interpretability, and excellent parameter transferability.

[0005] To solve the above-mentioned technical problems, the present invention adopts the following technical solution.

[0006] A method for predicting runoff in unmeasured watersheds based on spatial modeling and improved LSTM includes the following steps:

[0007] Step 1: Analyze the static characteristic data of the target watershed and introduce prior knowledge to give runoff prediction a hydrological basis;

[0008] Step 2: Spatial modeling of the unmeasured watershed. In data processing, the original data from the measurement station is divided into multiple sub-watersheds according to the watershed. Each sub-watershed includes all areas where water flows to the observation station. These gridded meteorological forcing data, including daily meteorological variables such as temperature, precipitation, air pressure, wind speed, specific humidity, shortwave and longwave radiation, are input as independent variables into the network model of the prediction method.

[0009] Step 3: Build an LSTM network model with a static input gate for predicting runoff in the target watershed;

[0010] Step 4: Optimize network model training using the Nash efficiency function and the Adam policy;

[0011] Step 5: Apply the trained network model to predict runoff in the target watershed, and use the SAC-SMA model to calibrate the prediction results.

[0012] Step one includes:

[0013] 1-1. Analyze the static characteristic data of the target watershed: Collect watershed characteristic attributes from the CAMELS watershed dataset, and use the Morris global sensitivity analysis method to make simple estimates of the most important static characteristics:

[0014]

[0015] Where i represents the static attribute being analyzed, EE i Indicates the basic effect, x i Δ represents the static characteristic attribute value of the watershed. i The pre-defined amount of change is used; then, the global sensitivity measure for each watershed and each static feature data is derived using the mean absolute gradient.

[0016] 1-2. Introduction of prior knowledge: The top six static feature attribute values ​​in terms of global sensitivity are selected as prior knowledge of the watershed and remain unchanged during the validation period; the top six static feature variable attribute values ​​in terms of global sensitivity include: p_mean daily average precipitation, pet_mean average daily potential evapotranspiration, aridity ratio of average evapotranspiration to average precipitation, forest_frac forest distribution, geol_permeability surface permeability, and max_water_content maximum soil moisture content.

[0017] Specifically, step 2 includes the following process:

[0018] 2-1. Spatialization of watersheds: In data processing, the original data from the measurement stations are divided into multiple sub-watershed data according to the watershed. Each sub-watershed includes all areas where water flows to the observation station. These gridded watershed meteorological forcing data, including daily meteorological variables such as temperature, precipitation, air pressure, wind speed, specific humidity, shortwave and longwave radiation, are used as independent variables.

[0019] 2-2. Discrete Representation of Watershed Data: Semi-distributed modeling of watershed data is performed, and the data is represented by h×w grid cells:

[0020]

[0021] Among them, X t This represents the watershed data for t;

[0022] 2-3. Construct a Convolutional Long Short-Term Memory Neural Network (Conv-LSTM) model to fuse geospatial input data. This network model consists of four convolutional LSTM layers and four convolutional non-recursive layers. To obtain predictions, the historical precipitation, minimum and maximum temperatures of the past 8 days, as well as a thermally encoded month, are represented as a grid matrix and input into the first convolutional LSTM layer.

[0023] 2-4. The output of the last layer of the non-recursive convolutional layer is concatenated with the watershed identifier, and the resulting tensor is passed through a non-recursive convolution using a released ReLU activation function; finally, this last layer outputs a prediction for several grid cells.

[0024] Specifically, in step 3, the construction of an LSTM network model with a static input gate for predicting runoff in the target watershed includes:

[0025] 3-1. Construct a system that can directly learn watershed features from multiple watershed data, using the input sequence x = [x[1], ..., x[t]] given in step 2 with T time steps, where each element x[t] is an input feature containing time step t, where 1≤t≤T;

[0026] 3-2. To modify the traditional LSTM network model, in addition to the weather forcing of the time series, a static embedding layer should be designed in the input layer of the network model, adding static features as additional input. Therefore, the formula is adjusted as follows:

[0027] i = σ(W i x s +b i (3)

[0028] Where σ represents the sigmoid activation function, i is still the input gate but no longer changes over time, and W... i It is the weight, x s It is a static feature, b i It is the activation value; in this way, the static properties of the watershed are introduced into the prior knowledge from this watershed by controlling the input cell state of the LSTM through the input gate i, and similar watersheds are also allowed to share the input cell state of the network model.

[0029] 3-3. Determine the different similar types of rainfall-runoff behavior that distinguish each entity in different watersheds, such that the static input gate of the improved LSTM network model contains a series of true values ​​in the range of (0,1), allowing certain parts of the static input gate to be active by simulating any single watershed;

[0030] 3-4. Match the six most important static features of the watershed, based on the top six static feature attributes with the highest global sensitivity, to the corresponding watersheds, and use them as the input gate x of the LSTM. s Static input;

[0031] 3-5. The gridded meteorological forcing data, including daily meteorological variables such as temperature, precipitation, air pressure, wind speed, specific humidity, shortwave and longwave radiation, are used as the dynamic input of the LSTM at each time step to train the network model for runoff prediction.

[0032] Specifically, in step 4, the optimization of network model training using the Nash efficiency function and the Adam policy includes:

[0033] 4-1. Let the improved average Nash efficiency NSE* for the unmeasured watershed be the objective function for training the network model, and its formula is:

[0034]

[0035] Where B is the number of watersheds, and N is the number of samples in each watershed B. y is the predicted value for sample n. n s(b) is the observed value, and s(b) is the standard deviation b of the flow in the basin calculated from the training period, where 1≤n≤N and 1≤b≤B;

[0036] 4-2. Configure all 256 cell states of the Long Short-Term Memory Neural Network. Before entering the non-recursive convolutional layer, the dropout rate of the LSTM output is 0.4. The Adam optimization strategy is adopted to update the variables according to the historical gradient oscillation of the LSTM network model and the true historical gradient after filtering the oscillation, so as to reduce the impact of the network model initialization weights.

[0037] Specifically, in step 5, the application of the trained network model to predict runoff in the target watershed includes:

[0038] 5-1. Perform static feature learning on unmeasured watersheds to introduce prior knowledge;

[0039] 5-2. Perform out-of-sample testing on the LSTM network model, namely: simulate the prediction of runoff in unmeasured watersheds; use the k-fold validation method to train and test sub-watersheds, that is, randomly divide the watersheds in the dataset into 12 groups of equal size, where k=12; use all watersheds in the k-1 groups to train the network model, and then test the network model on the watersheds in a single group.

[0040] 5-3. Train all LSTM network models based on the previous 15 years of CAMELS watershed data, including water-bearing years from 1981 to 1995;

[0041] 5-4. All calibration models SAC-SMA were evaluated based on the most recent 15 years of CAMELS data, including water years from 1996 to 2010; LSTM was also trained and evaluated using the k-fold validation method.

[0042] 5-5. Use the Nash efficiency loss function to guide the optimization of training results;

[0043] 5-6. Use SAC-SMA to evaluate the preliminary prediction results of step 5-4; if the prediction effect is better than the calibrated model SAC-SMA, retain it; otherwise, continue training and evaluation.

[0044] 5-7. If the prediction has ended, output the result; if it has not ended, proceed to step 3-1 to perform runoff prediction again.

[0045] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0046] 1. This invention uses the Morris global sensitivity analysis method to make simple estimates of the most important static features of the watershed and introduces these as prior knowledge into the network model, which can effectively address the problem of the lack of interpretability in network model predictions.

[0047] 2. This invention achieves spatial modeling of watersheds by discretizing watershed data in the form of h×w grid cells. This modeling method can be used for transfer learning of network models.

[0048] 3. This invention improves the input gate of LSTM and combines the average Nash efficiency (NSE*) and Adam optimization strategy designed specifically for unmeasured watersheds to effectively address the problems of low prediction accuracy of LSTM network models and overfitting during static feature learning of watersheds. Attached Figure Description

[0049] Figure 1 This is a schematic diagram of watershed data spatialization according to an embodiment of the present invention.

[0050] Figure 2 This is a diagram of an improved input gate LSTM network structure according to an embodiment of the present invention.

[0051] Figure 3 This is a diagram illustrating the overall algorithm implementation process of one embodiment of the present invention. Detailed Implementation

[0052] This invention discloses a method for predicting runoff in unmeasured watersheds based on spatial modeling and an improved LSTM, comprising: analyzing the static characteristics of the target watershed and introducing prior knowledge to give the runoff prediction a hydrological basis; performing gridding and data denoising on runoff data from watersheds adjacent to the target watershed to obtain a spatial runoff dataset for network model training and testing; constructing an LSTM network model with a static input gate to enhance its ability to learn prior knowledge in the hydrological domain; using an improved Nash efficiency as the training objective function and optimizing the network model training using the Adam strategy; applying the trained LSTM network model to predict runoff in the target (unmeasured) watershed and calibrating the prediction results using a SAC-SMA model. This invention offers high accuracy, strong physical interpretability, and high parameter transferability.

[0053] The present invention will now be described in further detail with reference to the accompanying drawings.

[0054] Figure 1 This is a flowchart illustrating a method for predicting unmeasured watershed runoff based on spatial modeling and improved LSTM, as described in this invention. Figure 1 As shown, the method in this embodiment includes the following steps:

[0055] Step 1: Analyze the static characteristic data of the target watershed and introduce prior knowledge to give the runoff prediction method a hydrological basis;

[0056] 1-1. Analyze the static characteristic data of the target watershed: Collect watershed characteristic attributes from the CAMELS watershed dataset, and use the Morris global sensitivity analysis method to make simple estimates of the most important static characteristics:

[0057]

[0058] Where i represents the static attribute being analyzed, EE i Indicates the basic effect, x i Δ represents the static characteristic attribute value of the watershed. i The pre-defined amount of change is used. Then, a global sensitivity measure for each watershed and each static feature data is derived using the average absolute gradient.

[0059] 1-2. Introduction of prior knowledge: The six static feature attribute values ​​with the highest global sensitivity are selected as the prior knowledge of the watershed (as shown in Table 1) and remain unchanged during the verification period.

[0060] Table 1

[0061]

[0062] Step 2: Spatial modeling of the unmeasured watershed. In data processing, the original measurement station data is divided into multiple sub-watersheds according to watersheds, such as... Figure 1 As shown, each sub-basin consists of all areas where water flows toward the observation station, and these gridded meteorological forcing data (daily meteorological variables such as temperature, precipitation, air pressure, wind speed, specific humidity, shortwave and longwave radiation, etc.) are input as independent variables into the network model of the prediction method.

[0063] 2-1. Spatialization of the watershed: In data processing, the original data from the measuring stations are divided into multiple sub-watershed data according to the watershed, such as... Figure 2 As shown in Table 2, each sub-basin consists of all areas where water flows toward the observation station, and these gridded basin meteorological forcing data are used as independent variables.

[0064] Table 2

[0065]

[0066] 2-2. Discrete Representation of Watershed Data: Semi-distributed modeling of watershed data is performed, and the data is represented by h×w grid cells:

[0067]

[0068] Among them, X t This represents the watershed data for t;

[0069] 2-3. Construct a Convolutional Long Short-Term Memory Neural Network (Conv-LSTM) model to fuse geospatial input data. This network model consists of four convolutional LSTM layers and four convolutional non-recursive layers. To obtain predictions, the historical precipitation, minimum and maximum temperatures of the past 8 days, as well as a thermally encoded month, are represented as a grid matrix and input into the first convolutional LSTM layer.

[0070] 2-4. The output of the last layer of the non-recursive convolutional layer is concatenated with the watershed identifier, and the resulting tensor is passed through a non-recursive convolution using a released ReLU activation function; finally, this last layer outputs a prediction for several grid cells;

[0071] Step 3: Build an LSTM network model with a static input gate for predicting runoff in the target watershed;

[0072] 3-1. Construct an LSTM network model that can directly learn watershed characteristics from multiple watershed datasets, such as... Figure 2 As shown, the input sequence x = [x[1], ..., x[t]] with T time steps given in step 2 is used, where each element x[t] is an input feature containing a time step t (1≤t≤T);

[0073] 3-2. To modify the traditional LSTM network model, in addition to the weather forcing of the time series, a static embedding layer should be designed in the input layer of the network model, adding static features as additional input. Therefore, the formula is slightly adjusted as follows:

[0074] i = σ(W i x S +b i (3)

[0075] Where o represents the sigmoid activation function, which is still the input gate but does not change over time, Wi is the weight, which is a static feature, and bi is the activation value; in this way, the static properties of the watershed are introduced into the input cell state of the LSTM through the input gate, thus introducing prior knowledge from this watershed, and also allowing similar watersheds to share the input cell state of the network model.

[0076] 3-3. Determine the dynamic behaviors (i.e. rainfall-runoff processes) that distinguish different similar types between individual entities (here, different watersheds) so that the static input gate of the improved LSTM network model contains a series of true values ​​in the range of (0, 1), allowing certain parts of the static input gate to be active by simulating any single watershed;

[0077] 3-4. Match the six most important static features of the watershed, based on the top six static feature attributes with the highest global sensitivity, to the corresponding watersheds, and use them as the input gate x of the LSTM. sStatic input.

[0078] 3-5. Use the gridded meteorological forcing data in Table 2 as the dynamic input for each time step of the LSTM to train the network model for runoff prediction.

[0079] Step 4: Optimize network model training using the Nash efficiency function and the Adam policy;

[0080] 4-1. The objective function for training the network model is the improved average Nash efficiency (NSE*) of the unmeasured watershed, which is as follows:

[0081]

[0082] Where S is the number of watersheds, and N is the number of samples in each watershed B. It is the predicted value of sample n (1≤n≤N), y n s(b) is the observed value, and s(b) is the standard deviation b of the flow in the basin calculated from the training period (1≤b≤B).

[0083] 4-2. Configure all 256 cell states of the Long Short-Term Memory (LSTM) neural network. Before entering the non-recursive convolutional layer, the dropout rate of the LSTM output is 0.4. The Adam optimization strategy is adopted to update variables based on the historical gradient oscillation of the LSTM network model and the true historical gradient after filtering the oscillations, in order to reduce the impact of the network model's initial weights.

[0084] Step 5: Apply the trained network model to predict runoff in the target watershed. The algorithm flow is as follows: Figure 3 As shown:

[0085] 5-1. Based on step 4, perform static feature learning on the unmeasured watershed to introduce prior knowledge;

[0086] 5-2. Perform out-of-sample testing on the LSTM network model (simulating the prediction of runoff in unmeasured watersheds). Use the k-fold (k=12) validation method to train and test sub-watersheds. That is, randomly divide the watersheds to be measured into 12 groups of equal size, use all watersheds in k-1 groups to train the network model, and then test the network model on the watersheds of a single group.

[0087] 5-3. All LSTM network models were trained based on the previous 15 years of CAMELS watershed data (1981-1995, years with water).

[0088] 5-4. All calibration models (SAC-SMA) were evaluated based on the most recent 15 years of CAMELS data (water years 1996-2010), and the LSTM was also trained and evaluated using the k-fold validation method (k=12).

[0089] 5-5. Use the Nash efficiency loss function to guide the optimization of training results;

[0090] 5-6. Use SAC-SMA to evaluate the preliminary prediction results in 5-4. If the prediction effect is better than that of the SAC-SMA model, retain it; otherwise, continue training and evaluation.

[0091] 5-7. If the prediction has ended, output the result; if it has not ended, proceed to step 3-1 to perform runoff prediction again.

Claims

1. A method for predicting unmeasured watershed runoff based on spatial modeling and improved LSTM, characterized in that, Includes the following steps: Step 1: Analyze the static characteristic data of the target watershed and introduce prior knowledge to give runoff prediction a hydrological basis; Step 2: Spatial modeling of the unmeasured watershed. In data processing, the original data from the measurement station is divided into multiple sub-watersheds according to the watershed. Each sub-watershed includes all areas where water flows to the observation station. These gridded meteorological forcing data, including daily meteorological variables such as temperature, precipitation, air pressure, wind speed, specific humidity, shortwave and longwave radiation, are input as independent variables into the network model of the prediction method. Step 3: Build an LSTM network model with a static input gate for predicting runoff in the target watershed; Step 4: Optimize network model training using the Nash efficiency function and the Adam policy; Step 5: Apply the trained network model to predict runoff in the target watershed, and use the SAC-SMA model to calibrate the prediction results. Step 1 includes: 1-1. Analyze the static characteristic data of the target watershed: Collect watershed characteristic attributes from the CAMELS watershed dataset, and use the Morris global sensitivity analysis method to make simple estimates of the most important static characteristics: (1) Where i represents the static attribute being analyzed. Indicates the basic effect. Represents the static characteristic attribute value of the watershed. The pre-defined amount of change is used; then, the global sensitivity measure for each watershed and each static feature data is derived using the mean absolute gradient. 1-2. Introduction of Prior Knowledge: The top six static feature attribute values ​​in terms of global sensitivity are selected as prior knowledge of the watershed and remain unchanged during the validation period; the top six static feature variable attribute values ​​in terms of global sensitivity include: p_mean daily average precipitation, pet_mean average daily potential evapotranspiration, aridity (ratio of average evapotranspiration to average precipitation), forest_frac (forest distribution), geol_permeability (surface permeability), and max_water_content (maximum soil moisture content).

2. The method for predicting unmeasured watershed runoff based on spatial modeling and improved LSTM according to claim 1, characterized in that, Step 2, Includes the following processes: 2-1. Spatialization of watersheds: In data processing, the original data from the measurement stations are divided into multiple sub-watershed data according to the watershed. Each sub-watershed includes all areas where water flows to the observation station. These gridded watershed meteorological forcing data, including daily meteorological variables such as temperature, precipitation, air pressure, wind speed, specific humidity, shortwave and longwave radiation, are used as independent variables. 2-2. Discrete Representation of Watershed Data: Semi-distributed modeling of watershed data is performed, and the data is represented by h×w grid cells: (2) in, express Watershed data; 2-3. Construct a Convolutional Long Short-Term Memory Neural Network (Conv-LSTM) model to fuse geospatial input data. This network model consists of four convolutional LSTM layers and four convolutional non-recursive layers. To obtain predictions, the historical precipitation, minimum and maximum temperatures of the past 8 days, as well as a thermally encoded month, are represented as a grid matrix and input into the first convolutional LSTM layer. 2-4. The output of the last layer of the non-recursive convolutional layer is concatenated with the watershed identifier, and the resulting tensor is passed through a non-recursive convolution using a released ReLU activation function; finally, this last layer outputs a prediction for several grid cells.

3. The method for predicting unmeasured watershed runoff based on spatial modeling and improved LSTM according to claim 1, characterized in that, In step 3, the construction of an LSTM network model with a static input gate for predicting runoff in the target watershed includes: 3-1. Construct an LSTM network model that can directly learn watershed features from multiple watershed data. Use the input sequence x=[x[1],…,x[t]] given in step 2, which contains T time steps, where each element x[t] is an input feature containing time step t, and 1≤t≤T; 3-2. To modify the traditional LSTM network model, in addition to the weather forcing of the time series, a static embedding layer should be designed in the input layer of the network model, adding static features as additional input. Therefore, the formula is adjusted as follows: (3) in, σ This represents the sigmoid activation function. It's still an input gate, but it doesn't change over time. W i It's weight. It is a static feature. b i It is the activation value; thus, the static properties of the watershed are passed through the input gate. Controlling the input cell state of LSTM introduces prior knowledge from this watershed, and also allows similar watersheds to share the input cell state of the network model; 3-3. Determine the different similar types of rainfall-runoff behavior that distinguish each entity in different watersheds, such that the static input gate of the improved LSTM network model contains a series of true values ​​in the range of (0,1), allowing certain parts of the static input gate to be active by simulating any single watershed; 3-4. Match the six most important static features of the watershed, based on the top six static feature attributes with the highest global sensitivity, to the corresponding watersheds and use them as the input gate for the LSTM. Static input; 3-5. The gridded meteorological forcing data, including daily meteorological variables such as temperature, precipitation, air pressure, wind speed, specific humidity, shortwave and longwave radiation, are used as the dynamic input of the LSTM at each time step to train the network model for runoff prediction.

4. The method for predicting unmeasured watershed runoff based on spatial modeling and improved LSTM according to claim 1, characterized in that, In step 4, the optimization of network model training using the Nash efficiency function and the Adam policy includes: 4-1. Let the improved average Nash efficiency NSE* for the unmeasured watershed be the objective function for training the network model, and its formula is: (4) in, It is the number of watersheds. That is the number of samples in each watershed. It is a sample n The predicted value, These are observed values. It is the standard deviation of the flow within the watershed calculated from the training period. b ,in, 1≤n≤N and 1≤b≤B ; 4-2. Configure all 256 cell states of the Long Short-Term Memory Neural Network. Before entering the non-recursive convolutional layer, the dropout rate of the LSTM output is 0.

4. The Adam optimization strategy is adopted to update the variables based on the historical gradient oscillation of the LSTM network model and the true historical gradient after filtering the oscillation, so as to reduce the impact of the network model initialization weights.

5. The method for predicting unmeasured watershed runoff based on spatial modeling and improved LSTM according to claim 3, characterized in that, In step 5, applying the trained network model to predict runoff in the target watershed includes: 5-1. Perform static feature learning on unmeasured watersheds to introduce prior knowledge; 5-2. Perform out-of-sample testing on the LSTM network model, namely: simulate the prediction of runoff in unmeasured watersheds; use the k-fold validation method to train and test sub-watersheds, that is, randomly divide the watersheds in the dataset into 12 groups of equal size, where k=12; use all watersheds in the k-1 groups to train the network model, and then test the network model on the watersheds in a single group. 5-3. Train all LSTM network models based on the previous 15 years of CAMELS watershed data, including water-bearing years from 1981 to 1995; 5-4. All calibration models SAC-SMA were evaluated based on the most recent 15 years of CAMELS data, including water years from 1996 to 2010; LSTM was also trained and evaluated using the k-fold validation method. 5-5. Use the Nash efficiency loss function to guide the optimization of training results; 5-6. Use SAC-SMA to evaluate the preliminary prediction results of step 5-4; if the prediction effect is better than the calibrated model SAC-SMA, retain it; otherwise, continue training and evaluation. 5-7. If the prediction has ended, output the result; if it has not ended, proceed to step 3-1 to perform runoff prediction again.

Images