Multisource hill type watershed flood forecasting method based on bidirectional-multilayer LSTM (Long Short Term Memory) model
By dividing the watershed into sections and combining reservoir scheduling and the Xin'anjiang model, a two-way multi-layer LSTM model is used for flood forecasting in hilly watersheds. This solves the problems of traditional models in characterizing nonlinear processes under steep terrain and not considering reservoir scheduling factors, and achieves high-precision hourly flow forecasting.
Patent Information
- Application Number
- CN202511560761.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-29
- Publication Date
- 2026-02-06
AI Technical Summary
Traditional hydrological models struggle to characterize the nonlinear processes of runoff generation and confluence under steep slopes in hilly watersheds. Distributed hydrological models and coupled meteorological-hydrological models do not adequately consider human activities, particularly reservoir scheduling, which affects forecast accuracy.
A bidirectional-multilayer LSTM model-based approach is adopted to divide the watershed into main stream confluence area, tributary confluence area and other confluence areas. Combining the reservoir scheduling model, the Xin'anjiang three-source model and the simplified confluence formula, time series samples are generated through a sliding window and a flood peak oversampling strategy is introduced to construct a weighted loss function optimization model to achieve hourly flow prediction.
It improves the accuracy and response speed of flood forecasting in hilly watersheds, especially in the prediction of extreme events. By combining partitioned modeling and deep learning techniques, it balances the distribution of the dataset and improves the model's fit to flood peaks.
Smart Images
Figure CN121480262A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the field of hydrology and water resources, and particularly relates to a multi-source hilly watershed flood forecasting method based on a bidirectional-multilayer LSTM model. BACKGROUND
[0002] Hilly watersheds are more difficult to predict than plain areas due to complex terrain, short confluence time, and strong flood suddenness. In recent years, with the increase in extreme weather events, flood disasters in hilly areas have occurred frequently, and related research has made progress in model improvement, data fusion, and early warning technology, but still faces many challenges. Traditional hydrological models (such as SCS and Xin'anjiang) have limited applicability in hilly areas because they cannot accurately depict the nonlinear process of runoff generation and confluence on steep slopes. Distributed hydrological models (such as TOPKAPI and DHSVM) can effectively depict spatial heterogeneity of terrain, soil, and vegetation by using high-resolution terrain data (such as LiDAR) to simulate slope flow. Coupled meteorological-hydrological models (such as WRF-Hydro) can quantify the impact of short-duration heavy rainfall and extend the prediction period. However, these models require high-quality data and modeling structures, and almost do not consider human activities, especially reservoir operation, which affects the prediction accuracy of the model.
[0003] Bi-directional multi-layer LSTM (Bi-ML-LSTM) is an improved recurrent neural network (RNN) that combines bidirectional information flow and deep architecture, and has shown significant advantages in time series data modeling. In recent years, this model has been widely used in natural language processing (NLP), speech recognition, and meteorological and hydrological prediction, and has shown strong ability in handling long sequence dependencies and complex pattern mining. Bi-ML-LSTM improves the robustness of time series modeling through bidirectional deep architecture, but its application still needs to be optimized in specific scenarios to balance performance and efficiency.
[0004] In summary, the existing technology has the following problems: traditional hydrological models (such as SCS and Xin'anjiang) cannot accurately depict the nonlinear process of runoff generation and confluence on steep slopes in hilly watersheds, and have limited applicability in hilly areas; distributed hydrological models (such as TOPKAPI and DHSVM) and coupled meteorological-hydrological models (such as WRF-Hydro) require high-quality data and modeling structures, and almost do not consider human activities, especially reservoir operation, which affects the prediction accuracy of the model. SUMMARY
[0005] The technical problem to be solved by the present application is to provide a multi-source hilly watershed flood forecasting method based on a bidirectional-multilayer LSTM model, which realizes high-precision prediction of hourly flow in multi-source runoff generation and confluence hilly watersheds by coupling hydrological models and deep learning technology in different regions.
[0006] To solve the above technical problems, the technical solution adopted by the present application is: A multi-source hilly watershed flood forecasting method based on a bidirectional-multilayer LSTM model, comprising the following steps: Step 1, divide the to-be-forecasted watershed into a main stream confluence area, a branch stream confluence area and several other confluence areas, and calculate the confluence area runoff of each area using a reservoir scheduling model, a Xin'anjiang three water source model and a simplified confluence formula respectively; Step 2, based on the confluence area runoff of each area, meteorological data and derived features, generate time series samples through a sliding window, and preprocess the sample data using a flood peak oversampling strategy; Step 3, construct a bidirectional-multilayer LSTM model, which includes a bidirectional LSTM layer, a Dropout layer, a unidirectional LSTM layer and a fully connected output layer; Step 4, introduce a weighted loss function to train and optimize the bidirectional-multilayer LSTM model; Step 5, use the trained bidirectional-multilayer LSTM model to predict the hourly flood runoff process of the watershed outlet section.
[0007] Preferably, the confluence area runoff of the main stream confluence area is calculated by a simplified reservoir scheduling model, which comprehensively considers the hourly rainfall, reservoir catchment area, runoff coefficient, water release proportion coefficient converted by reservoir scheduling rules and hydrological response delay duration.
[0008] Preferably, the simplified reservoir scheduling model formula is as follows: ; ; In the formula: is the hourly rainfall; is the reservoir catchment area; is the runoff coefficient; is the water release proportion coefficient converted by the scheduling rule; is the response delay duration.
[0009] Preferably, the branch stream confluence area uses the Xin'anjiang three water source model for runoff calculation; the Xin'anjiang three water source model uses a three-layer evapotranspiration model to calculate evapotranspiration, uses a full storage runoff mode to calculate watershed runoff, divides runoff into surface runoff, interflow and groundwater runoff through a free reservoir structure, and uses a linear reservoir for watershed confluence calculation, and uses the Muskingum method or the lagging calculation method for river confluence calculation.
[0010] Preferably, the runoff yield of the other confluence area is estimated by a simplified confluence formula based on the hourly rainfall, confluence area and runoff coefficient.
[0011] Preferably, the sliding window has a preset time length, the model input is a sequence of hydro-meteorological features in the past continuous hours, and the output is a prediction value of the runoff at the outlet section of the basin in the future one hour; the hydro-meteorological features include the discharge after reservoir regulation, runoff yield of each confluence area, runoff of tributaries, hourly rainfall, cumulative rainfall and time period encoding information. The peak over-sampling strategy includes setting a peak flow determination threshold, and duplicating samples with a true flow greater than or equal to the threshold to increase the proportion of peak samples in the training set.
[0012] Preferably, the structure of the bidirectional-multilayer LSTM model includes: a bidirectional LSTM layer for simultaneously capturing historical information and future trends of the input sequence; a Dropout layer arranged after the bidirectional LSTM layer for preventing overfitting in the model training process; a unidirectional LSTM layer receiving the output of the Dropout layer for extracting deep time sequence features; another Dropout layer arranged after the unidirectional LSTM layer; two fully connected layers, the former fully connected layer for feature compression and nonlinear transformation, and the latter fully connected layer as an output layer with an output dimension of 1 for predicting the future one-hour flow.
[0013] Preferably, the weighted loss function is used to emphasize the importance of peak flow prediction in the model training process, and its calculation method is to assign different weights to different samples; specifically, a higher weight is assigned to the peak flow sample than to the flow sample in the flat water period to balance the gradient contribution of samples with different flow levels in the training process.
[0014] Preferably, the weighted loss function has the following mathematical form: ; wherein: wi is the weight of the i-th sample; is the mean square error; is the true value and the predicted value; ; wherein: is the peak determination threshold, which is set as the 90th percentile of the flow; is the peak sample weight function, which is set as 5.
[0015] The application discloses a multi-source hilly basin flood forecasting system based on a bidirectional-multilayer LSTM model. The basin partition calculation module is used for dividing the basin to be forecasted into a main stream confluence area, a branch stream confluence area and a plurality of other confluence areas, and calling a reservoir regulation model, a Xin'anjiang three water source model and a simplified confluence formula respectively to calculate the confluence area runoff of each area; The data preprocessing module is used for generating time sequence samples through a sliding window based on the confluence area runoff of each area, meteorological data and derived features, and performing oversampling processing on flood peak samples. The model construction and training module is used for constructing a deep learning model comprising a bidirectional LSTM layer, a Dropout layer, a unidirectional LSTM layer and a full connection layer, and training and optimizing the model by introducing a weighted loss function. The flood flow prediction module is used for loading the trained model, inputting the latest hydrological and meteorological data and outputting the future hourly basin outlet section flow prediction result.
[0016] In the bidirectional-multilayer LSTM model constructed by the model construction and training module, the bidirectional LSTM layer is used for extracting bidirectional time sequence features of the input sequence, the unidirectional LSTM layer is used for extracting deep time sequence dependence features, and the full connection layer is used for realizing nonlinear mapping from high-dimensional features to flow prediction values.
[0017] When generating training samples, the data preprocessing module takes the past continuous hydrological and meteorological feature sequence for several hours as input and takes the measured flow of the future one hour as output label; the hydrological and meteorological features include the reservoir regulation treated discharge, the confluence area runoff, the branch stream inflow, the hourly rainfall, the cumulative rainfall and the time period coding information.
[0018] The weighted loss function gives higher weight to the flood peak flow sample in the model training process, so as to improve the prediction accuracy of the model to extreme hydrological events.
[0019] A computer device comprises one or more processors, and one or more executable programs are stored on the processor; when the one or more executable programs are executed by the one or more processors, the one or more executable programs are used to realize the multi-source hilly basin flood forecasting method based on the bidirectional-multilayer LSTM model.
[0020] A storage medium stores one or more executable programs; when the one or more executable programs are executed, the one or more executable programs are used to realize the multi-source hilly basin flood forecasting method based on the bidirectional-multilayer LSTM model.
[0021] The present invention can achieve the following beneficial effects: This invention targets hilly watersheds with multiple runoff sources, integrating reservoir scheduling rules and the Xin'anjiang hydrological model to achieve watershed zoning modeling and computation. Based on the principle of a bidirectional multi-layer LSTM model, through data preprocessing, model building, and model training, a novel hydrological forecasting model is constructed by coupling the zoning hydrological model and deep learning technology. This model can provide decision-making reference and technical support for flood forecasting in hilly watersheds with multiple runoff sources.
[0022] This invention targets hilly watersheds with relatively complex topographical conditions. By integrating reservoir scheduling rules with the Xin'anjiang River model, it achieves regional modeling and calculation of the main stream and tributaries. Based on the LSTM model, a bidirectional LSTM layer is constructed to simultaneously learn historical and future trends, improving the flood peak response speed. The deep LSTM layer addresses the long-term rainfall-runoff lag effect. By combining flood peak oversampling and a weighted loss function, the dataset distribution is balanced, effectively improving the model's fitting effect on flood peaks. Attached Figure Description
[0023] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a flowchart illustrating the technical process of the method of the present invention. Figure 2 Flowchart of LSTM neural unit computation; Figure 3 This is a pseudocode diagram of a bidirectional multilayer LSTM model; Figure 4 Example diagram of flood flow process simulation; Figure 5 This is an example graph showing the results of a flood simulation calculation. Detailed Implementation
[0024] Preferred solutions include Figures 1 to 5 As shown, a flood forecasting method for multi-source hilly watersheds based on a bidirectional-multilayer LSTM model is presented. The specific method is as follows: 1. The basin is divided into the main stream confluence area, the tributary confluence area and several other confluence areas. The runoff and runoff are calculated using the reservoir scheduling model, the Xin'anjiang three-source model and the simplified confluence formula, respectively. 2. Preprocess the model data by using a sliding window to generate samples and a peak oversampling strategy; 3. Design a bidirectional-multi-layer LSTM model, including a bidirectional LSTM layer (64 units), a Dropout layer (0.4), a unidirectional LSTM layer (32 units), and a fully connected layer (16→1 unit). 4. Introduce a weighted loss function to optimize the training of the bidirectional-multilayer LSTM model; 5. Based on the bidirectional-multilayer LSTM model, the flood flow process of the basin is predicted.
[0025] Wherein, the runoff yield calculation of the dry flow confluence area adopts the simplified reservoir regulation model:
[0026]
[0027] In the formula: is the rainfall of this hour, mm; is the reservoir catchment area, km 2 ; is the runoff coefficient; is the discharge proportion coefficient converted from the regulation rule; is the response delay length.
[0028] The tributary confluence area adopts the three-source Xin'anjiang model for runoff yield calculation. The three-source Xin'anjiang model adopts a three-layer evapotranspiration model to calculate evapotranspiration, a full storage runoff yield mode to calculate basin runoff yield, and a method of dividing runoff into surface runoff, interflow and groundwater runoff by using a free reservoir to calculate different runoff. The basin confluence adopts a linear reservoir, and the river confluence adopts the Muskingum algorithm or the lag algorithm.
[0029] A number of confluence areas adopt the confluence formula to estimate runoff yield:
[0030] In the formula, is the rainfall of this hour, mm; is the reservoir catchment area, km 2 ; is the runoff coefficient.
[0031] The data preprocessing adopts a sliding window method to construct time series samples, and the window length is 3 hours. The input is a hydro-meteorological feature sequence of the past 3 consecutive hours, and the output is a prediction value of the basin outlet section flow in the future 1 hour. The features include the hourly discharge after reservoir regulation, runoff yield of each confluence area, tributary inflow, hourly rainfall and its derived indicators (such as 3-hour cumulative rainfall, time coding, etc.). In order to alleviate the model bias caused by the scarcity of flood peak samples, the 90% quantile is used as the flood peak threshold , all window samples of true flow y≥flood peak threshold are oversampled, that is, the flood peak subset is copied several times, and then mixed with the original sample.
[0032] The LSTM (Long Short Term Memory) model is an improved recurrent neural network (RNN). Unlike traditional RNN, the neural unit of LSTM is not a simple neuron with "weighted sum and activation function", but contains multiple internal calculation substructures, and introduces the gating mechanism of forget gate, input gate and output gate. The input of a LSTM unit is the input of the current time step, the hidden state candidate memory unit of the last time step, the memory state update, the output gate and the hidden state update.
[0033] The calculation process of each calculation unit at time t is as follows: let the input be : current time step input vector, : hidden state of the last time step, : memory unit state of the last time step, and , ] will be spliced into an integral input vector.
[0034] (1) Input gate: controls the degree of writing of current input memory.
[0035]
[0036] In the formula, : determines the proportion of new information written into the cell state.
[0037] (2) Forget gate: determines how much old memory to keep,
[0038]
[0039] In the formula, : sigmoid activation function, output range [0, 1] W f : weight matrix of forget gate b f : bias vector of forget gate : represents the degree of forgetting for each unit. If ≈1, it means complete retention, and if ≈0, it means forgetting (3) Candidate memory: generate new content that can be written into memory.
[0040]
[0041] In the formula, tanh: activation function limits the output to [-1, 1], and generates candidate new memory values.
[0042] (4) Update memory state: update long-term memory by integrating forgetting and input.
[0043]
[0044] In the formula, : retain part of the memory, : add part of the new memory.
[0045] (5) Generate hidden state h t : the current updated memory state C t After being processed by tanh and filtered by the output gate, the current hidden state h t is generated, which is used to pass to the next time step or for the final output.
[0046]
[0047] (6) Output gate: determine how much memory to output currently.
[0048]
[0049] The bidirectional-multilayer LSTM model structure includes one layer of bidirectional LSTM layer (64 units) and one layer of unidirectional LSTM layer (32 units), which respectively extract shallow and deep time sequence features; a Dropout layer is interposed in the middle to prevent overfitting; finally, two layers of fully connected (Dense) structure are used to complete the nonlinear mapping of the predicted flow. The overall structure is shown in Table 1.
[0050] Table 1 Bidirectional-multilayer LSTM model structure table
[0051] Among them, the bidirectional LSTM can simultaneously perceive the historical information and future trend of the input sequence, and improve the response ability of the model to sudden flood peaks; the multilayer LSTM enhances the nonlinear modeling and long-term dependence memory ability through deep structure. The model is based on recurrent neural network (RNN), and by introducing the gating mechanism and bidirectional structure, the modeling ability of time sequence dependence and nonlinear characteristics is enhanced, which is especially suitable for the lag response and complex driving factor coupling in the flood process of the basin.
[0052] The present application introduces a balanced weighted loss function, the purpose of which is to give higher weight to the flood peak sample in the loss calculation process, so that the model pays more attention to the prediction accuracy of the flood peak in the training process, and the influence of the minority class (flood peak) in the overall gradient update is closer to that of the majority class (flat water period). Its mathematical form is as follows:
[0053] wherein: wi is the weight of the i-th sample; MSE is the mean square error; yi is the true value and
[0054]
[0055] wherein: Fpeak is the flood peak determination threshold, set as the 90th percentile of the flow; wi is the flood peak sample weight function, set as 5.
[0056] The above-described embodiments are merely preferred technical solutions of the present application, and should not be regarded as a limitation on the present application. The protection scope of the present application should be based on the technical solutions recited in the claims, including equivalent replacement solutions of the technical features recited in the claims. That is, equivalent replacement improvements within this scope are also within the protection scope of the present application.
Claims
1. A flood forecasting method for multi-source hilly watersheds based on a bidirectional-multilayer LSTM model, characterized in that, Includes the following steps: Step 1: Divide the watershed to be forecasted into the main stream confluence area, the tributary confluence area and several other confluence areas, and calculate the confluence area runoff of each area using the reservoir scheduling model, the Xin'anjiang three-source model and the simplified confluence formula respectively; Step 2: Based on the runoff, meteorological data and derived characteristics of each region's catchment area, time series samples are generated through a sliding window, and the sample data are preprocessed using a flood peak oversampling strategy. Step 3: Construct a bidirectional-multi-layer LSTM model, which includes a bidirectional LSTM layer, a Dropout layer, a unidirectional LSTM layer, and a fully connected output layer; Step 4: Introduce a weighted loss function to train and optimize the bidirectional-multilayer LSTM model; Step 5: Using the trained bidirectional-multilayer LSTM model, predict the hourly flood discharge process at the watershed outlet section.
2. The flood forecasting method for multi-source hilly watersheds based on a bidirectional-multilayer LSTM model according to claim 1, characterized in that, The runoff generation of the main stream confluence area is calculated using a simplified reservoir scheduling model. This model comprehensively considers hourly rainfall, reservoir catchment area, runoff generation coefficient, water release ratio coefficient of reservoir scheduling rule conversion, and hydrological response delay time.
3. The flood forecasting method for multi-source hilly watersheds based on a bidirectional-multilayer LSTM model according to claim 1, characterized in that, The simplified reservoir scheduling model formula is as follows: ； ； In the formula: This represents the hourly rainfall. The catchment area of the reservoir; The flow rate coefficient; The water release ratio coefficient is converted from the scheduling rules; To reflect the duration of response delay; This refers to the outflow from the reservoir. The outflow from the reservoir is taken into account in accordance with the scheduling rules.
4. The flood forecasting method for multi-source hilly watersheds based on a bidirectional-multilayer LSTM model according to claim 1, characterized in that, The tributary confluence area is calculated using the Xin'anjiang three-source model; the Xin'anjiang three-source model uses a three-layer evapotranspiration model to calculate evapotranspiration, a full-storage runoff model to calculate watershed runoff, and a free reservoir structure to divide runoff into surface runoff, interflow, and groundwater runoff. A linear reservoir is used for watershed runoff calculation, and river channel runoff is calculated using the Muskingen method or a lag algorithm.
5. The flood forecasting method for multi-source hilly watersheds based on a bidirectional-multilayer LSTM model according to claim 1, characterized in that, The runoff generation of the other runoff areas is estimated by a simplified runoff formula, which is calculated based on hourly rainfall, catchment area, and runoff generation coefficient.
6. The flood forecasting method for multi-source hilly watersheds based on a bidirectional-multilayer LSTM model according to claim 1, characterized in that, The time series length of the sliding window is a preset number of hours. The model input is a sequence of hydrological and meteorological characteristics from the past several consecutive hours, and the output is a predicted value of the watershed outlet section flow for the next hour. The hydro-meteorological characteristics include: outflow after reservoir regulation and treatment, flow rate of each confluence area, inflow of tributaries, hourly rainfall, cumulative rainfall and time period coding information; The flood peak oversampling strategy includes: setting a flood peak flow determination threshold, and copying samples whose actual flow is greater than or equal to the threshold to increase the proportion of flood peak samples in the training set.
7. The flood forecasting method for multi-source hilly watersheds based on a bidirectional-multilayer LSTM model according to claim 1, characterized in that, The structure of the bidirectional multilayer LSTM model includes: A bidirectional LSTM layer is used to simultaneously capture historical information and future trends of the input sequence; A Dropout layer, placed after the bidirectional LSTM layer, is used to prevent overfitting during model training. A unidirectional LSTM layer receives the output of the Dropout layer and is used to extract deep temporal features; Another Dropout layer is set after the unidirectional LSTM layer; Two fully connected layers: the first fully connected layer is used for feature compression and nonlinear transformation, and the second fully connected layer is the output layer with an output dimension of 1, used to predict traffic flow in the next hour.
8. The flood forecasting method for multi-source hilly watersheds based on a bidirectional-multilayer LSTM model according to claim 1, characterized in that, The weighted loss function is used to increase the importance of peak flow prediction during model training. It is calculated by assigning different weights to the mean square error of different samples. Specifically, the peak flow samples are given a higher weight than the normal flow samples to balance the gradient contribution of samples of different flow levels during training.
9. A flood forecasting method for multi-source hilly watersheds based on a bidirectional-multilayer LSTM model according to claim 8, characterized in that, The mathematical form of the weighted loss function is as follows: ； in: Let be the weight of the i-th sample; Mean square error; For actual and predicted values; ； in: The threshold for determining the flood peak is set at the 90th percentile of the flow rate; Let be the peak sample weighting function, set to 5.
10. A flood forecasting system for multi-source hilly watersheds based on a bidirectional-multilayer LSTM model, characterized in that, The system employs a multi-source hilly watershed flood forecasting method based on a bidirectional-multilayer LSTM model, as described in any one of claims 1-9. The system comprises: The watershed zoning calculation module is used to divide the watershed to be predicted into the main stream confluence area, the tributary confluence area and several other confluence areas, and respectively call the reservoir scheduling model, the Xin'anjiang three-source model and the simplified confluence formula to calculate the confluence area runoff of each region; The data preprocessing module is used to generate time-series samples through a sliding window based on the runoff, meteorological data and derived characteristics of each region's runoff area, and to perform oversampling processing on the flood peak samples; The model building and training module is used to build a deep learning model that includes bidirectional LSTM layers, Dropout layers, unidirectional LSTM layers and fully connected layers, and to train and optimize the model by introducing a weighted loss function. The flood flow prediction module is used to load the trained model, input the latest hydrological and meteorological data, and output the hourly flow prediction results of the watershed outlet section.
11. The system according to claim 10, characterized in that, In the bidirectional-multilayer LSTM model constructed by the model building and training module, the bidirectional LSTM layer is used to extract bidirectional temporal features of the input sequence, the unidirectional LSTM layer is used to extract deep temporal dependency features, and the fully connected layer is used to realize the nonlinear mapping from high-dimensional features to traffic prediction values.
12. The system according to claim 10, characterized in that, When generating training samples, the data preprocessing module takes the hydrological and meteorological characteristic sequence of the past several consecutive hours as input and the measured flow rate of the next hour as output label. The hydrological and meteorological characteristics include: outflow after reservoir regulation and treatment, flow rate of each confluence area, inflow of tributaries, hourly rainfall, cumulative rainfall and time period coding information.
13. The system according to claim 10, characterized in that, The weighted loss function assigns higher weights to peak flow samples during model training to improve the model's prediction accuracy for extreme hydrological events.
14. A computer device, characterized in that, The method includes one or more processors, on which one or more executable programs are stored, which, when executed by the one or more processors, are used to implement a method for forecasting floods in multi-source hilly watersheds based on a bidirectional-multilayer LSTM model according to any one of claims 1-9.
15. A storage medium, characterized in that, It stores one or more executable programs, which, when executed, are used to implement a method for forecasting floods in multi-source hilly watersheds based on a bidirectional-multilayer LSTM model, as described in any one of claims 1-9.