Flood runoff forecasting method based on LSTM-SWAT coupling model

Abstract

The invention provides a flood runoff forecasting method based on an LSTM-SWAT coupling model, and the method comprises the steps: constructing an SWAT distributed hydrological model, and analyzing the calibration and parameter sensitivity of the SWAT distributed hydrological model; a BiLSTM deep learning model is introduced, the constructed SWAT distributed hydrological model and the BiLSTM deep learning model are coupled, and based on physical mechanism operation of the SWAT distributed hydrological model, expansion of input meteorological data is completed in space and time; and taking the multi-feature data output by the SWAT distributed hydrological model as the input data of the BiLSTM deep learning model, and forecasting the runoff. According to the method, the hydrological elements are disassembled, the input characteristic parameter screening problem caused by poor interpretability of a deep learning model is solved, the runoff forecasting model effect is further improved, meanwhile, the influence of multi-dimensional characteristic parameters on forecasting performance is explored, and a new solution can be provided for research of the estuary environment in the flood period.

Description

Technical Field

[0001] The present invention relates to the technical field of flood runoff forecasting, and in particular to a flood runoff forecasting method based on an LSTM-SWAT coupling model. Background Art

[0002] With the continuous intensification of climate change and human activities, the frequency and scale of extreme flood events have increased dramatically, causing huge population and economic losses. my country has a long and winding coastline, rich and diverse coastal terrain, a large number of important port cities distributed along the coast, and a large number of populations and industrial facilities gathered in coastal areas, which are the main disaster carriers in Asia. The developed coastal economy and dense population make it inevitable that my country has the characteristics of insufficient disaster resilience and weak disaster bearing capacity. The research on efficient forecasting of flood runoff is of great significance to my country's economic development and population security.

[0003] The current mainstream hydrological models can be divided into two categories: lumped hydrological models (Xin'anjiang model, Tank model, Sacramento model) and distributed hydrological models (SWAT model, SHE model, IHDM model, DTVGM model). Lumped hydrological models are conceptual models that do not consider the spatial heterogeneity of hydrological phenomena and underlying surface elements. The parameters in the model usually take the average value of the basin, and the basin is regarded as a whole, which simplifies the modeling process. Distributed hydrological models are mathematical models that calculate the hydrological process of the basin based on physical mechanisms. They fully consider the uneven characteristics of hydrological phenomena and underlying surface space, and can fully reflect the impact of rainfall and spatial changes of underlying surface elements in the basin on runoff generation. Both lumped hydrological models and distributed hydrological models are widely used in hydrological research, but their application in runoff forecasting is relatively limited.

[0004] Deep learning models can be an advanced method to solve the runoff forecasting problem because of their ability to solve nonlinear problems. Among them, the long short-term memory model (LSTM) is widely used to deal with time series problems. However, due to its lack of interpretability and inability to accurately depict the real physical process, it is difficult for the model to accurately select input feature quantities in complex phenomena such as flood processes, which still has limitations in its application in flood runoff forecasting. Summary of the invention

[0005] According to the technical problems raised above, a flood runoff forecasting method based on LSTM-SWAT coupling model is provided. The present invention breaks through the problem of input feature parameter screening caused by the poor interpretability of deep learning models by disassembling hydrological elements, further improves the effect of runoff forecasting model, and explores the influence of multidimensional feature parameters on forecasting performance, which can provide new solutions for the study of estuary environment during floods.

[0006] The technical means adopted by the present invention are as follows:

[0007] A flood runoff forecasting method based on LSTM-SWAT coupling model, including:

[0008] S1. Construct the SWAT distributed hydrological model and analyze the calibration and parameter sensitivity of the SWAT distributed hydrological model;

[0009] S2. Introduce the BiLSTM deep learning model, and couple the constructed SWAT distributed hydrological model with the BiLSTM deep learning model. Based on the physical mechanism calculation of the SWAT distributed hydrological model, complete the expansion of the input meteorological data in space and time;

[0010] S3. The multi-feature data output by the SWAT distributed hydrological model is used as the input data of the BiLSTM deep learning model to predict runoff.

[0011] Furthermore, in step S1, the constructed SWAT distributed hydrological model includes: a hydrological module, the hydrological module is used for hydrological simulation, and the hydrological simulation includes: a land phase of the hydrological cycle and a river channel calculation phase of the hydrological cycle, wherein:

[0012] The terrestrial phase of the hydrological cycle, which controls the amount of water, sediment, and nutrients entering rivers;

[0013] The river routing stage of the hydrological cycle is defined as the process of water and sediment movement in the river channel to the outlet, and the entire basin system follows the law of water balance.

[0014] Furthermore, in step S1, the SWAT distributed hydrological model is calibrated and parameter sensitivity analysis is performed using the SWAT-CUP software, and the SUFI-2 uncertainty analysis algorithm built into the SWAT-CUP software is used as an optimization method for parameter estimation. The specific process includes:

[0015] Determine the relevant parameters and initial parameter ranges according to the simulation content of the SWAT distributed hydrological model, and determine the return method according to the parameter type, including addition, subtraction, multiplication, and substitution;

[0016] Set the start and end dates and warm-up period parameters for the SWAT distributed hydrological model simulation;

[0017] Run the SWAT-CUP calibration program and save the calculation results;

[0018] Check the calculation results. If the accuracy requirement cannot be met, replace the old range with the new parameter range and repeat the previous step until the accuracy requirement is met.

[0019] After obtaining satisfactory calibration results, save the calibration results and global parameter sensitivity analysis results;

[0020] Replace the initial files in the corresponding path of the SWAT distributed hydrological model with the files in the SWAT-CUP calibration path to complete the calibration.

[0021] Furthermore, in step S2, the introduced BiLSTM deep learning model considers both the positive and negative bidirectional sequence features of the input sequence, and captures more implicit rules in the runoff sequence by inputting bidirectional time series features.

[0022] Furthermore, step S2 also includes a process of calculating the LSTM model, including:

[0023] Calculate the forget gate and select the information to be forgotten. The calculation formula is as follows:

[0024] f t =σ(W f ⊙[h t-1 ,x t ]+b f )

[0025] Among them, σ represents the Sigmoid function, and the output interval is (0, 1); W f The calculation weight of the forget gate; ⊙ represents the product of vector elements; h t-1 Indicates the external state at the previous moment; x t Indicates the input at the current moment; b f Represents the bias parameter of the forget gate;

[0026] Calculate the input gate and select the information to be memorized. The calculation formula is as follows:

[0027] i t =σ(W i ⊙[h t-1 ,x t ]+b i )

[0028]

[0029] Among them, W i Represents the calculation weight of the input gate; W c represents the calculation weight of the memory storage unit; b i represents the bias parameter of the input gate; b c Represents the bias parameter of the memory storage unit; U c represents learnable parameters; Represents the temporary memory unit state obtained by nonlinear function;

[0030] Calculate the cell state at the current moment. The calculation formula is as follows:

[0031]

[0032] Among them, f t ∈[0,1] D represents the forget gate, the forget gate f t Control the internal state c at the previous moment t-1 How much information needs to be forgotten; c t-1 Indicates the state of the memory unit at the previous moment; i t ∈[0,1] D represents the input gate, input gate i t Control the candidate status at the current moment How much information needs to be saved?

[0033] Calculate the output gate, the calculation formula is as follows:

[0034] o t =σ(W o ⊙[h t-1 ,x t ]+b o )

[0035] h t =o t ⊙tanh(c t )

[0036] Among them, the internal state c of the LSTM model t ∈R D It is specialized in linear cyclic information transmission and nonlinearly outputs information to the external state h of the hidden layer. t ∈R D ; At each time t, the internal state c of the LSTM network t Records historical information up to the current moment; W o represents the calculation weight of the output gate; b o Represents the bias parameter of the output gate; o t ∈[0,1] D Represents the output gate, output gate o t Control the internal state c at the current moment t How much information needs to be output to the external state h t .

[0037] Furthermore, step S3 specifically includes:

[0038] S31. Collect meteorological data, digital elevation data, soil type data, and land use data to prepare the input library for the SWAT distributed hydrological model;

[0039] S32, divide sub-basins and hydrological response units through the SWAT distributed hydrological model, and specify watershed nodes and runoff outflow ports;

[0040] S33, inputting a meteorological database, a soil type database, and a land use database, dividing the slope, and merging units according to a set threshold;

[0041] S34, run the SWAT distributed hydrological model, store the results, and record the river channel output file, sub-basin output file, and hydrological response unit output file;

[0042] S35, extracting the output results of the target area target parameters and grouping them according to the experimental design;

[0043] S36, selecting the required features, incorporating them into the runoff matrix, dividing the data set into 70% training set and 30% test set, and performing normalization preprocessing on the training set using the standard deviation normalization method;

[0044] S37, input the training set into the BiLSTM deep learning model, start training, and adjust the hyperparameters according to the change in the loss amount of the BiLSTM deep learning model;

[0045] S38, normalize the test set and input it into the BiLSTM deep learning model for calculation. After the calculation is completed, store the result and perform denormalization;

[0046] S39. Repeat steps S31 to S38 according to the experimental grouping, and calculate the similarity coefficient R and the root mean square error RMSE between each group of experimental observation values ​​and simulation values.

[0047] Compared with the prior art, the present invention has the following advantages:

[0048] The present invention provides a flood runoff forecasting method based on the LSTM-SWAT coupling model. By combining the advantages of the strong interpretability of the SWAT distributed hydrological model and the strong learning ability of the BiLSTM model, its performance is better than that of the traditional distributed hydrological model and the deep learning model. It can better simulate the runoff of medium and high intensity floods, will bring better forecasting effects for flood runoff, and can provide suggestions for subsequent environmental protection, port construction, and waterway maintenance.

[0049] Based on the above reasons, the present invention can be widely promoted in the fields of flood runoff forecasting and the like. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0051] Figure 1 Schematic diagram of the land surface structure of the hydrological cycle.

[0052] Figure 2 This is a diagram of the computing structure of the SWAT distributed hydrological model of the present invention.

[0053] Figure 3 This is the network structure diagram of the LSTM model.

[0054] Figure 4 This is the BiLSTM deep learning model architecture of the present invention.

[0055] Figure 5 This is a structural diagram of the SWAT-BiLSTM coupling model of the present invention. DETAILED DESCRIPTION

[0056] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0057] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0058] The present invention provides a flood runoff forecasting method based on an LSTM-SWAT coupling model, comprising:

[0059] S1. Construct the SWAT distributed hydrological model and analyze the calibration and parameter sensitivity of the SWAT distributed hydrological model;

[0060] S2. Introduce the BiLSTM deep learning model, and couple the constructed SWAT distributed hydrological model with the BiLSTM deep learning model. Based on the physical mechanism calculation of the SWAT distributed hydrological model, complete the expansion of the input meteorological data in space and time;

[0061] S3. The multi-feature data output by the SWAT distributed hydrological model is used as the input data of the BiLSTM deep learning model to predict runoff.

[0062] In specific implementation, as a preferred embodiment of the present invention, in step S1, the constructed SWAT distributed hydrological model includes: a hydrological module, the hydrological module is used for hydrological simulation, and the hydrological simulation includes: the land stage of the hydrological cycle and the river channel calculation stage of the hydrological cycle, wherein:

[0063] The land phase of the hydrological cycle controls the amount of water, sediment, and nutrients entering the river; e.g. Figure 1 shown.

[0064] The river routing stage of the hydrological cycle is defined as the process of water and sediment movement in the river channel to the outlet, and the entire basin system follows the law of water balance.

[0065] like Figure 2 As shown, SWAT (Soil and Water Assessment Tool) is a large-scale, long-time distributed hydrological model, including four sub-modules: hydrological module, soil erosion and sediment transport module, nutrient transport module, and plant growth and management module. In this embodiment, the present invention only uses the hydrological module.

[0066] In the specific implementation, as a preferred embodiment of the present invention, in step S1, SWAT-CUP (SWATCalibration Uncertainty Procedures) software is used to calibrate the SWAT distributed hydrological model and perform parameter sensitivity analysis, and the SUFI-2 (Sequential uncertainty fitting version 2) uncertainty analysis algorithm built into the SWAT-CUP software is used as the optimization method for parameter estimation. The specific process includes:

[0067] Determine the relevant parameters and initial parameter ranges according to the simulation content of the SWAT distributed hydrological model, and determine the return method according to the parameter type, including addition, subtraction, multiplication, and substitution;

[0068] Set the start and end dates and warm-up period parameters for the SWAT distributed hydrological model simulation;

[0069] Run the SWAT-CUP calibration program and save the calculation results;

[0070] Check the calculation results. If the accuracy requirement cannot be met, replace the old range with the new parameter range and repeat the previous step until the accuracy requirement is met.

[0071] After obtaining satisfactory calibration results, save the calibration results and global parameter sensitivity analysis results;

[0072] Replace the initial files in the corresponding path of the SWAT distributed hydrological model with the files in the SWAT-CUP calibration path to complete the calibration.

[0073] In specific implementation, as a preferred embodiment of the present invention, a recurrent neural network (RNN) is a type of recurrent neural network that takes sequence data as input, recursively in the direction of sequence evolution, and all nodes are connected in a chain. The network state is not only affected by the current input conditions, but also takes into account the changes in the input value at the previous moment. Figure 3 As shown in Figure 1, the LSTM long short-term memory model is a variant of the RNN model. It controls the path of information transmission by introducing three gating mechanisms: input gate, forget gate, and output gate. In step S2, the introduced BiLSTM deep learning model considers both the positive and negative bidirectional sequence features of the input sequence. By inputting the bidirectional time series features, it captures more implicit rules in the runoff sequence. Its structure diagram is shown in Figure 1. Figure 4 shown.

[0074] In specific implementation, as a preferred embodiment of the present invention, step S2 also includes a process of calculating the LSTM model, including:

[0075] Calculate the forget gate and select the information to be forgotten. The calculation formula is as follows:

[0076] f t =σ(W f ⊙[h t-1 ,x t ]+b f )

[0077] Among them, σ represents the Sigmoid function, and the output interval is (0, 1); W f represents the weight corresponding to the forget gate; ⊙ represents the product of vector elements; h t-1 Indicates the external state at the previous moment; x t Indicates the input at the current moment; b f represents the bias parameter;

[0078] Calculate the input gate and select the information to be memorized. The calculation formula is as follows:

[0079]

[0080]

[0081] Among them, W i Indicates the weight corresponding to the input gate calculation; W cIndicates the weight corresponding to the calculation of the memory storage unit; b i Indicates the bias parameter corresponding to the input gate calculation; b c Indicates the bias parameter corresponding to the calculation of the memory storage unit; U c represents learnable parameters; Represents the temporary memory unit state obtained by nonlinear function;

[0082] Calculate the cell state at the current moment. The calculation formula is as follows:

[0083]

[0084] Among them, f t ∈[0,1] D represents the forget gate, the forget gate f t Control the internal state c at the previous moment t-1 How much information needs to be forgotten; c t-1 Indicates the state of the memory unit at the previous moment; i t ∈[0,1] D represents the input gate, input gate i t Control the candidate status at the current moment How much information needs to be saved?

[0085] Calculate the output gate, the calculation formula is as follows:

[0086] o t =σ(W o ⊙|[h t-1 ,x t ]+b o )

[0087] h t =o t ⊙tanh(c t )

[0088] Among them, the internal state c of the LSTM model t ∈R D It is specialized in linear cyclic information transmission and nonlinearly outputs information to the external state h of the hidden layer. t ∈R D ; At each time t, the internal state c of the LSTM network t Records historical information up to the current moment; W o represents the calculation weight of the output gate; b o Represents the bias parameter of the output gate; o t ∈[0,1] D Represents the output gate, output gate o t Control the internal state c at the current moment t How much information needs to be output to the external state ht .

[0089] When specifically implemented, as a preferred embodiment of the present invention, Figure 5 As shown, step S3 specifically includes:

[0090] S31. Collect meteorological data, digital elevation data, soil type data, and land use data to prepare the input library for the SWAT distributed hydrological model;

[0091] S32, divide sub-basins and hydrological response units through the SWAT distributed hydrological model, and specify watershed nodes and runoff outflow ports;

[0092] S33, inputting a meteorological database, a soil type database, and a land use database, dividing the slope, and merging units according to a set threshold;

[0093] S34, run the SWAT distributed hydrological model, store the results, and record the river channel output file (Output.rch), sub-basin output file (Output.sub), and hydrological response unit output file (Output.hru);

[0094] S35, extracting the output results of the target area target parameters and grouping them according to the experimental design;

[0095] S36, selecting the required features, incorporating them into the runoff matrix, dividing the data set into 70% training set and 30% test set, and performing normalization preprocessing on the training set using the standard deviation normalization method;

[0096] S37, input the training set into the BiLSTM deep learning model, start training, and adjust the hyperparameters according to the change in the loss amount of the BiLSTM deep learning model;

[0097] S38, normalize the test set and input it into the BiLSTM deep learning model for calculation. After the calculation is completed, store the result and perform denormalization;

[0098] S39. Repeat steps S31 to S38 according to the experimental grouping, and calculate the similarity coefficient R and the root mean square error RMSE between each group of experimental observation values ​​and simulation values.

[0099] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A flood runoff forecasting method based on LSTM-SWAT coupling model, characterized in that: include: S1. Construct the SWAT distributed hydrological model and analyze the calibration and parameter sensitivity of the SWAT distributed hydrological model; S2. Introduce the BiLSTM deep learning model, and couple the constructed SWAT distributed hydrological model with the BiLSTM deep learning model. Based on the physical mechanism calculation of the SWAT distributed hydrological model, complete the expansion of the input meteorological data in space and time; S3. The multi-feature data output by the SWAT distributed hydrological model is used as the input data of the BiLSTM deep learning model to predict runoff.

2. A flood runoff forecasting method based on LSTM-SWAT coupling model according to claim 1, characterized in that: In step S1, the constructed SWAT distributed hydrological model includes: a hydrological module, which is used for hydrological simulation. The hydrological simulation includes: a land phase of the hydrological cycle and a river channel calculation phase of the hydrological cycle, wherein: The terrestrial phase of the hydrological cycle, which controls the amount of water, sediment, and nutrients entering rivers; The river routing stage of the hydrological cycle is defined as the process of water and sediment movement in the river channel to the outlet, and the entire basin system follows the law of water balance.

3. A flood runoff forecasting method based on LSTM-SWAT coupling model according to claim 1, characterized in that: In step S1, the SWAT distributed hydrological model is calibrated and parameter sensitivity analysis is performed using the SWAT-CUP software. The SUFI-2 uncertainty analysis algorithm built into the SWAT-CUP software is used as the optimization method for parameter estimation. The specific process includes: Determine the relevant parameters and initial parameter ranges according to the simulation content of the SWAT distributed hydrological model, and determine the return method according to the parameter type, including addition, subtraction, multiplication, and substitution; Set the start and end dates and warm-up period parameters for the SWAT distributed hydrological model simulation; Run the SWAT-CUP calibration program and save the calculation results; Check the calculation results. If the accuracy requirement cannot be met, replace the old range with the new parameter range and repeat the previous step until the accuracy requirement is met. After obtaining satisfactory calibration results, save the calibration results and global parameter sensitivity analysis results; Replace the initial files in the corresponding path of the SWAT distributed hydrological model with the files in the SWAT-CUP calibration path to complete the calibration.

4. A flood runoff forecasting method based on LSTM-SWAT coupling model according to claim 1, characterized in that: In step S2, the introduced BiLSTM deep learning model considers both the positive and negative bidirectional sequence features of the input sequence, and captures more implicit rules in the runoff sequence by inputting bidirectional time series features.

5. A flood runoff forecasting method based on LSTM-SWAT coupling model according to claim 4, characterized in that: Step S2 also includes the process of calculating the LSTM model, including: Calculate the forget gate and select the information to be forgotten. The calculation formula is as follows: f t =σ(W f ⊙[h t-1 ,x t ]+b f ) Among them, σ represents the Sigmoid function, and the output interval is (0, 1); W f represents the weight corresponding to the forget gate; ⊙ represents the product of vector elements; h t-1 Indicates the external state at the previous moment; x t Indicates the input at the current moment; b f represents the bias parameter; Calculate the input gate and select the information to be memorized. The calculation formula is as follows: i t =σ(W i ⊙[h t-1 ,x t ]+b i ) Among them, W i Indicates the weight corresponding to the input gate calculation; W c Indicates the weight corresponding to the calculation of the memory storage unit; b i Indicates the bias parameter corresponding to the input gate calculation; b c Indicates the bias parameter corresponding to the calculation of the memory storage unit; U c represents learnable parameters; Represents the temporary memory unit state obtained by nonlinear function; Calculate the cell state at the current moment. The calculation formula is as follows: Among them, f t ∈[0,1] D represents the forget gate, the forget gate f t Control the internal state c at the previous moment t-1 How much information needs to be forgotten; c t-1 Indicates the state of the memory unit at the previous moment; i t ∈[0,1] D represents the input gate, input gate i t Control the candidate status at the current moment How much information needs to be saved? Calculate the output gate, the calculation formula is as follows: the t =σ(W o ⊙|[h t-1 ,x t ]+b o ) h t =o t ⊙tanh(c t ) Among them, the internal state c of the LSTM model t ∈R D It is specialized in linear cyclic information transmission and nonlinearly outputs information to the external state h of the hidden layer. t ∈R D ; At each time t, the internal state c of the LSTM network t Records historical information up to the current moment; W o Indicates the weight currently calculated by the output gate; b o Indicates the bias parameter corresponding to the output gate calculation; o t ∈[0,1] D Represents the output gate, output gate o t Control the internal state c at the current moment t How much information needs to be output to the external state h t .

6. A flood runoff forecasting method based on LSTM-SWAT coupling model according to claim 1, characterized in that: Step S3 specifically includes: S31. Collect meteorological data, digital elevation data, soil type data, and land use data to prepare the input library for the SWAT distributed hydrological model; S32, divide sub-basins and hydrological response units through the SWAT distributed hydrological model, and specify watershed nodes and runoff outflow ports; S33, inputting a meteorological database, a soil type database, and a land use database, dividing the slope, and merging units according to a set threshold; S34, run the SWAT distributed hydrological model, store the results, and record the river channel output file, sub-basin output file, and hydrological response unit output file; S35, extracting the output results of the target area target parameters and grouping them according to the experimental design; S36, selecting the required features, incorporating them into the runoff matrix, dividing the data set into 70% training set and 30% test set, and performing normalization preprocessing on the training set using the standard deviation normalization method; S37, input the training set into the BiLSTM deep learning model, start training, and adjust the hyperparameters according to the change in the loss amount of the BiLSTM deep learning model; S38, normalize the test set and input it into the BiLSTM deep learning model for calculation. After the calculation is completed, store the result and perform denormalization; S39. Repeat steps S31 to S38 according to the experimental grouping, and calculate the similarity coefficient R and the root mean square error RMSE between each group of experimental observation values ​​and simulation values.

Images