A strong human activity influence basin flood forecasting method based on a mixture model

Abstract

The application discloses a kind of strong human activity influence basin flood forecasting method based on hybrid model, first, the distributed hydrological model of combination sub-basin division and masking method is constructed and is rated;Again, with the characteristics of hydrological model simulation error and multi-step length basin average precipitation based on river channel propagation time, error correction model is constructed using Bayes optimization LSTM;Finally, real-time prediction is realized by coupling two models.It is proved by Lanxi basin verification that the Nash efficiency coefficient of hybrid model is significantly improved compared with traditional model, the relative error of flood peak is less than or equal to 8%, the time difference of peak appearance is less than or equal to 2 hours, and the accuracy is stable within 12 hours forecast period.The application does not need long series of continuous uninterrupted historical data as support, only needs the data during the period of flood, and is suitable for hour scale flood forecasting in data-deficient strong human activity basin, provides efficient and reliable technical means for flood control and disaster reduction, and has wide engineering application prospect.

Description

Technical Field

[0001] This invention belongs to the field of flood forecasting technology, specifically relating to a flood forecasting method for watersheds with strong human activity impacts based on a hybrid model. Background Technology

[0002] Flood forecasting is a core technical support for flood control, disaster reduction, and water resource allocation. It predicts future hydrological conditions based on known meteorological and hydrological information, and traditional technical approaches mainly rely on hydrological models. However, in recent years, to meet the comprehensive needs of flood control, power generation, and navigation, many river basins have constructed large-scale water conservancy projects (such as reservoirs and dams). The complex scheduling systems formed by these projects have significantly altered the natural runoff processes of the basins. For example, the storage function of reservoirs can delay the occurrence of flood peaks and reduce peak flow, while the scheduling of dams can cause river confluence patterns to deviate from their natural state. This makes it difficult for traditional hydrological models (such as conceptual models and distributed models) to adapt to the flood forecasting needs of basins heavily influenced by human activities.

[0003] To address the aforementioned issues, existing technologies propose two types of improvement schemes: The first is the conceptual hydrological model parameter generalization method, which generalizes the impact of water conservancy projects by adjusting model parameters or introducing virtual reservoirs. However, this method has limited generalization capabilities and cannot uncover the nonlinear runoff response patterns under complex water conservancy project scheduling; it is only applicable to watersheds with relatively weak human interference. The second is the distributed hydrological model method, which simulates water conservancy project scheduling by adding reservoir operation modules to model nodes. However, this method has two major limitations: first, real-time scheduling of complex water conservancy project groups requires a combination of model guidance and human judgment, and pre-set operating rules cannot cover all actual scenarios; second, module operation relies on real-time reservoir water levels, outflow data, etc., but most small and medium-sized reservoirs lack systematic monitoring data, making it difficult to promote this method in areas lacking data.

[0004] Hybrid models (coupling physical mechanism models and data-driven models) offer a new approach to solving the aforementioned problems. "Error correction" is a typical coupling method—the physical model provides mechanistic support, while the data-driven model corrects the forecasting errors of the physical model. However, existing hybrid models have significant shortcomings: ① Their application scenarios are limited to natural watersheds, lacking design for the specific characteristics of watersheds heavily impacted by human activities; ② Their time scales are concentrated on daily runoff simulation, lacking adaptability for hourly-scale real-time flood forecasting; ③ Their interpretability is poor, with the selection of input features and feature step sizes relying on experience or statistical methods, failing to incorporate mechanistic information from hydrological models (such as river confluence time), resulting in weak generalization ability in complex watersheds and an inability to meet the accurate forecasting needs of watersheds heavily impacted by human activities. Summary of the Invention

[0005] The present invention aims to provide a flood forecasting method for watersheds with strong human activity impacts based on a hybrid model, thereby solving the aforementioned technical problems existing in the prior art.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following solution:

[0007] A method for forecasting floods in watersheds with strong human activity impacts based on a hybrid model includes the following steps:

[0008] Step S1: Construct a hydrological model, which includes the following steps:

[0009] S11: Sub-basins based on natural water systems: The original DEM data is processed to generate river links and the forecast basin is divided using watersheds to obtain initial sub-basins. The size of the initial sub-basins is adjusted based on the distribution and density of rain gauges to obtain the target sub-basins. For example, initial sub-basins with an area of ​​less than 50 km² (which can be adjusted according to the actual size of the basin) are merged to avoid interpolation errors in rainfall data caused by sub-basins being too small; initial sub-basins with an area of ​​more than 300 km² are manually divided to ensure that each sub-basin contains at least one rain gauge to guarantee the accuracy of rainfall input for the distributed model.

[0010] S12: Construction of River System Topology and Delineation of Muskingen Sections: Number the target sub-basins and construct the river system topology. Combine river length, river gradient, and river confluence time to delineate the number of Muskingen sections. For example, for river sections with long river lengths and small gradients, delineate more Muskingen sections to refine the confluence process and ensure that the flood wave propagation simulation matches the actual river characteristics.

[0011] S13: Parameter calibration of distributed hydrological model: The parameter calibration method of human-computer interaction is adopted, and the parameters of the distributed hydrological model are optimized with the relative error of runoff depth, the relative error of flood peak, the Nash efficiency coefficient NSE and the peak occurrence time difference as objective functions.

[0012] The specific requirements for each objective function are as follows: during the calibration period, the relative error of runoff depth is ≤15%, the relative error of flood peak is ≤10%, the Nash efficiency coefficient (NSE) is ≥0.8, and the peak time difference is ≤3 hours; during the validation period, the relative error of runoff depth is ≤20%, the relative error of flood peak is ≤15%, the Nash efficiency coefficient (NSE) is ≥0.7, and the peak time difference is ≤3 hours.

[0013] Through iterative optimization of the objective function, a calibrated distributed hydrological model that matches the actual hydrological characteristics of the watershed is obtained.

[0014] Step S2: Construct a deep learning error correction model and train it, which includes the following steps:

[0015] S21: Calculate the simulation error sequence of historical hydrological models: based on historical measured flow rates Q 0 and the historical calculated flow rate output by the distributed hydrological model after calibration QcThe historical hydrological model simulation error sequence was calculated. The simulation error sequence includes influencing factors that were not simulated by the distributed model, such as water conservancy project scheduling and model parameter uncertainties.

[0016] S22: Determine the input features and feature step size of the deep learning model: The input features are the previous multi-step basin-wide average precipitation and the simulation error of the historical hydrological model. The basin-wide average precipitation reflects the driving effect of rainfall on floods, while the simulation error of the historical hydrological model reflects the impact of historical water conservancy projects and model biases. Combining these two factors can cover key influencing factors in flood forecasting. The basin-wide average precipitation is calculated based on the weights of the target sub-basins. Since the impact of complex water conservancy projects is mainly reflected in the river confluence process, the river propagation time of the target sub-basin farthest from the forecast section is used as the feature step size. This river propagation time is calculated using the number of Muskingen sections and the propagation time (KE) of a single Muskingen section.

[0017] S23: Construct a deep learning error correction model: Based on the input features and feature step size in step S22, construct a sample set, divide the sample set into a training set and a validation set, and normalize them respectively.

[0018] A deep learning model network structure is constructed, the hyperparameters are optimized using the Bayesian optimization algorithm, and the model is trained using the training set to obtain a deep learning error correction model.

[0019] Step S3: Real-time flood forecasting: Output the historical calculated flow rate Tc hours before the current time and the initial forecast flow rate for the next 72 hours through the calibrated distributed hydrological model, where Tc > feature step size; calculate the historical hydrological model simulation error for the corresponding period based on the historical calculated flow rate of the previous feature step size in Tc hours before the current time and the historical measured flow rate for the corresponding period; input the basin average precipitation of the previous feature step size and the historical hydrological model simulation error into the deep learning error correction model to calculate the forecast error at the current time and continuously obtain the forecast error for multiple future forecast periods; combine the initial forecast flow rate with the forecast error for the corresponding forecast period to obtain the corrected forecast flow rate.

[0020] Further optimization involves processing the original DEM data in step S11, specifically including: performing mosaicking, depression filling, flow direction extraction, and cumulative flow calculation operations on the original DEM data in sequence.

[0021] The initial sub-basin size is adjusted based on the distribution and density of rain gauges. Specifically, this includes merging initial sub-basins with an area smaller than a first preset threshold, and manually dividing initial sub-basins with an area larger than a second preset threshold, where the second preset threshold is greater than the first preset threshold.

[0022] Further optimization involves dividing the Muskingan river segment into several parts in step S12, taking into account the river length, river gradient, and river confluence time. Specifically, this includes using the river confluence time as the core constraint, and combining the confluence distance corresponding to the river length and the confluence velocity corresponding to the river gradient, to determine the number of Muskingan river segments from each target sub-basin to the forecast section, ensuring that the flood wave propagation process matches the actual river confluence pattern.

[0023] In a further optimization, in step S13, the objective function is calculated based on historical measured flow rates. Q 0 and the historical calculated flow rate output by the distributed hydrological model after calibration Qc The formulas for calculating each objective function are as follows:

[0024] relative error of runoff depth ;in, R 0 represents the flow rate based on historical measured data. Q 0. Calculated measured runoff depth. R C To calculate traffic based on historical data Qc Calculated predicted runoff depth.

[0025] relative error of flood peak ;in, Based on historical measured flow Q The measured peak flow rate extracted from 0 To calculate traffic based on historical data Qc Extracted forecast peak flow.

[0026] Nash efficiency coefficient ;in, Q 0( i ) is the first i Historical measured flow rate at any given time; Q c ( i )for i Historical flow calculation at any given moment; Historical measured flow rate Q The mean of 0; N This represents the length of the flood sequence.

[0027] Peak time difference Δ T = T 0- T c ;in, T 0 represents the flow rate based on historical measured data. Q 0. The exact measured time of the flood peak. T c To calculate traffic based on historical data Qc The predicted time of the flood peak has been determined.

[0028] Further optimization is achieved by using the following formula to calculate the simulation error of the historical hydrological model in step S21:

[0029] Δ Q sim ( i )= Q 0( i )- Q c ( i ), where Δ Q sim ( i ) is the first i Historical hydrological model simulation errors at any given time Q 0( i ) is the first i Historical measured flow rate at any given time Q c ( i ) is the first i Historical flow calculation at any given moment.

[0030] Further optimization is achieved by using the following formula to calculate the average precipitation in the watershed during step S22:

[0031] ;

[0032] in, For the first i The average precipitation in the basin at any given time α j For the first j The weight of each sub-basin P j ( i ) is the first i Time of the first j Precipitation in each sub-basin m This indicates the number of sub-basins.

[0033] The formula for calculating river propagation time is: T L =nKE ,in, T L The river propagation time is the time from the target sub-basin furthest from the forecast section. n This represents the number of Muskingan river segments in the distributed hydrological model. KE This refers to the river propagation time for a single river segment.

[0034] Further optimization involves constructing the sample set in step S23 using historical moments. i Before T L Average precipitation in the basin over a given period And the simulation error of historical hydrological models {Δ Q sim ( iT L ),…,Δ Q sim ( i- 1)} is used as the input sample, with historical moments i The historical hydrological model simulation error is used as the output sample to form sample pairs (Ω, γ ),in,

[0035] ;

[0036] γ= {Δ Q sim ( i )};where Ω is the input sample set, γ is the output sample set, and the historical time i satisfy i ≥TL.

[0037] Further optimization is achieved in step S23, where the deep learning model is an LSTM model. The network structure of the LSTM model includes an input layer, a first LSTM layer, a first Dropout layer, a second LSTM layer, a second Dropout layer, and an output layer. The parameters for each layer are set as follows:

[0038] Input layer: Input dimension is 2× T L ,adaptation" T L Step length of average precipitation in the basin + T L The input feature dimension of "step-size historical hydrological model simulation error";

[0039] First LSTM layer: The number of neurons is set to the first preset value, the activation function is tanh, and the return sequence parameter is set to True, which is used to initially extract short-term time-dependent features from the input sequence;

[0040] First Dropout layer: The output feature matrix of the first LSTM layer is applied. The Dropout rate is set to the first preset rate to suppress the overfitting of the model to abnormal data in the training set.

[0041] The second LSTM layer has the number of neurons set to a second preset value, which is less than the first preset value. The activation function is tanh, and the return sequence parameter is set to False. This layer is used to extract long-term time-dependent features from the input sequence.

[0042] Second Dropout layer: The output feature vector of the second LSTM layer is applied. The Dropout rate is set to the second preset rate to further suppress overfitting of the deep network.

[0043] Output layer: The number of neurons is set to 1, and the activation function is a linear activation function, which is used to output the error prediction value at the corresponding time step.

[0044] Further optimization involves optimizing the hyperparameters of the deep learning model and training it in step S23. Specifically, a Bayesian optimization algorithm is used with the objective function of maximizing the Nash efficiency coefficient (NSE) of the validation set. The hyperparameters include the number of neurons in the first LSTM layer, the number of neurons in the second LSTM layer, the first preset rate, the second preset rate, the learning rate, and the number of training iterations.

[0045] The Adam optimizer is used, with the error value predicted by the LST model and the simulation error Δ from the historical hydrological model as the benchmarks. Q sim ( i The mean squared error of the validation set is used as the loss function. The LSTM model is trained using the training set. When the mean squared error of the validation set does not decrease within a preset number of consecutive rounds or reaches the optimized number of training iterations, the training is terminated, and the trained deep learning error correction model is obtained.

[0046] Further optimization, step S3 specifically includes the following steps:

[0047] S31: Initial Forecast Output: Outputs the current forecast using the calibrated distributed hydrological model. k Flow rate calculated hourly over the previous Tc hours { Q c ( k - T c ),…, Q c ( k -1)} and the initial hourly forecast flow for the next 72 hours Q pre,k+ε , ε Forecast period, ε =1,2,...,72; where Tc > characteristic step size T L ; i For a historic moment, i Earlier than the present moment k .

[0048] S32: Calculation of Historical Simulation Error Sequence: Based on the Current Moment k The first Tc hour T LHistorical flow calculation by step size { Q c ( k - T L ),…, Q c ( k -1)} and the historical measured flow rate for the corresponding time period { Q 0( k - T L ),…, Q 0( k -1)}, calculate the historical hydrological model simulation error subset {Δ Q sim ( k - T L ),…,Δ Q sim ( k -1)}.

[0049] S33: Error Prediction and Rolling Iteration: First, the previous... T L The time-by-time average precipitation of the basin with characteristic step size and the subset of historical hydrological model simulation errors {Δ Q sim ( k - T L ),…,Δ Q sim ( k -1)}, together input into the deep learning error correction model, to calculate the current time step k The forecast error value Δ Q pred ( k ).

[0050] Then, based on the basin average precipitation forecast for the next 72 hours. , with Δ Q pred ( k Using Δ as the initial value, iteratively input the deep learning error correction model: calculate Δ Q pred ( k When +1 is entered, input ; Calculate Δ Q pred ( k When +2), enter ; and so on, successively obtain the forecast error set {Δ} for each forecast period in the next 72 hours. Q pred (k +1),…,Δ Q pred ( k +72)}.

[0051] S34: Final Forecast Flow Calculation: Calculate the initial forecast flow output from the distributed hydrological model. Q pre,k+ε Forecast error Δ with the corresponding lead time Q pred ( k + ε According to the formula Q cor,k+ε = Q pre,k+ε +Δ Q pred ( k + ε The corrected final forecast flow is obtained by time-by-time matching calculation. Q cor,k+ε .

[0052] Compared with the prior art, the beneficial effects of the present invention are:

[0053] 1. The method described in this invention has strong applicability: by calculating the characteristic step size using the number of Muskingen river sections and KE, it avoids dependence on upstream water conservancy project data and solves the forecasting problem of watersheds with scarce data and strong human activity impact.

[0054] 2. The method described in this invention has high accuracy: In 28 floods in the Lanxi River Basin, the NSE of the hybrid model is significantly improved compared with the traditional distributed model, with the NSE improving by 29.1% during the training period and 65.8% during the validation period; and within the 10-hour average forecast period of the basin, the relative error of the flood peak is ≤8% and the peak occurrence time difference is ≤2 hours, which meets the requirements of flood control scheduling.

[0055] 3. The method described in this invention achieves a balance between interpretability and generalization: by incorporating the Muskingan channel confluence time of the hydrological model into the deep learning feature step size design, the "black box" problem of existing hybrid models is solved; at the same time, the nonlinear fitting capability of the deep learning model can accurately capture the complex impact of water conservancy projects, and has strong generalization ability in different watersheds with strong human activity impact.

[0056] 4. The method described in this invention has good time scale adaptability: the model structure is designed for the real-time forecasting needs on an hourly scale, the forecast period covers 1-72 hours, and the forecast accuracy decays relatively slowly within 12 hours, which can meet the time requirements of flood control emergency response. Attached Figure Description

[0057] Figure 1 Flowchart of a flood forecasting method for watersheds with strong human activity impacts based on a hybrid model;

[0058] Figure 2 Map showing the distribution of water conservancy projects in the Lanxi River Basin;

[0059] Figure 3 A schematic diagram illustrating the sub-basin division and numbering of the distributed Xin'anjiang model in the Lanxi River Basin;

[0060] Figure 4 The image shows the effect of applying a hybrid model to 28 flood events in Lanxi.

[0061] Figure 5 This is a diagram illustrating the application effect of a typical flood hybrid model in Lanxi.

[0062] Figure 6 This is a diagram showing the forecast performance of the hybrid model at different lead times. Detailed Implementation

[0063] The technical solution of the present invention will be described in detail below with reference to the embodiments, but the scope of protection of the present invention is not limited to the embodiments described.

[0064] In this embodiment, taking the Lanxi River Basin in Zhejiang Province, a typical watershed affected by strong human activities, as an example, the implementation process of the present invention is described in detail, such as... Figure 1 As shown, the specific steps include the following:

[0065] Step S1: Hydrological model construction, specifically including:

[0066] S11: Data Preparation

[0067] Hydrological and meteorological data for the Lanxi River Basin from January 2020 to June 2024 were compiled, and 28 flood events in Lanxi were selected. Precipitation data consisted of hourly rainfall data from 23 rain gauge stations. Flow data consisted of hourly measured flow data from the Lanxi hydrological station. Geographic data included 30m resolution DEM data, watershed river system vector data, and a map of water conservancy project distribution, such as... Figure 2 As shown, Figure 2 The data shows the spatial distribution of 23 rain gauge stations, 1 hydrological station (Lanxi station), more than 40 reservoirs, and river channels within the basin. Twenty-eight complete flood events were selected from the data, including 20 events from 2020 to 2022 for model calibration and 8 events from 2023 to 2024 for validation.

[0068] S12: Sub-basin division, performing mosaicking, depression filling, flow direction extraction, and cumulative runoff calculation on DEM data to generate river links.

[0069] In this embodiment, the depression-filling threshold is set to 5m, and the runoff threshold is set to 1000 grid cells. The initial sub-basins are segmented using the watershed algorithm, and the sub-basin size is adjusted based on the distribution density of 23 rain gauges. Sub-basins with an area less than 30km² are merged, and sub-basins with an area greater than 250km² are manually divided, ultimately resulting in 41 target sub-basins, such as... Figure 3 As shown.

[0070] S13: River System Topology and Muskingen Section Division: Forty-one sub-basins were numbered 1 to 41, with number 1 being the furthest sub-basin and number 41 being the sub-basin containing the Lanxi section. River system topology was constructed. Based on estimations of channel length, gradient, and confluence time, the river channel from the furthest sub-basin (number 1) to the Lanxi section was divided into 12 Muskingen sections.

[0071] S14: Parameter Calibration: The Xin'anjiang model was adopted as the distributed hydrological model, with the objective functions being "relative error of runoff depth ≤ 15%, relative error of flood peak ≤ 10%, NSE ≥ 0.8, and peak occurrence time difference ≤ 3 hours". Based on data from 20 floods from 2020 to 2022, the model parameters were obtained through human-computer interaction calibration, as shown in Table 1.

[0072] Table 1. Excellent values ​​of parameters for the Xin'anjiang model

[0073]

[0074] The NSE for the eight floods during the validation period was 0.52, the relative error of the flood peak was 18.5%, and the peak occurrence time difference was 4.2 hours, indicating that the traditional model is difficult to adapt to the impact of water conservancy projects, and the forecast accuracy is significantly affected by human activities.

[0075] Step S2: Hybrid model construction, specifically including:

[0076] S21: Calculation of Historical Hydrological Model Simulation Error Sequence: Based on the historical measured flow at Lanxi Station and the historical calculated flow from the Xin'anjiang model, according to the formula Δ Q sim ( i )= Q 0( i )- Q c ( i The simulation error sequence of the historical hydrological model for 28 floods was calculated. The mean simulation error was -85 m³ / s and the standard deviation was 210 m³ / s, reflecting a systematic bias in the Xin'anjiang model.

[0077] S22: Determining Input Features and Feature Step Size:

[0078] Input features: Selecting early stage T LThe hourly average precipitation of the basin is compared with the historical hydrological model simulation error. The average precipitation of the basin is calculated according to the area weight of the sub-basin. For example, the weight of sub-basin 1 is 0.08 and the weight of sub-basin 41 is 0.12.

[0079] Feature step size T L =12×1.3=15.6≈16 hours, so the step size for input features is determined to be 16 hours.

[0080] S23: Construction of the Deep Learning Error Correction Model LSTM:

[0081] Sample set: historical moments i Using the average precipitation over the previous 16 hours and the historical hydrological model simulation error over the previous 16 hours as inputs, and taking historical time as the data... i The historical hydrological model simulation error is used as the output, and 12,000 sample pairs are constructed. They are divided into training set and validation set at an 8:2 ratio, and both are normalized using Z-score.

[0082] Model Training: An LSTM model was constructed, with the following sequence: Input layer → First LSTM layer (64 neurons) → First Dropout layer (Dropout rate set to 0.2) → Second LSTM layer (32 neurons) → Second Dropout layer (Dropout rate set to 0.2) → Output layer. Bayesian optimization was used to obtain the optimal hyperparameters, with a learning rate of 0.005 and 200 iterations. After training, the validation set NSE was 0.89, the mean error decreased to -12 m³ / s, and the standard deviation decreased to 58 m³ / s.

[0083] Step S3: Verification of the prediction performance of the hybrid model, specifically including:

[0084] S31: Historical Flood Simulation: Simulations were performed on 28 floods, and the results are as follows: Figure 4 As shown in the figure, the NSE distributions of the traditional Xin'anjiang model and the hybrid model are compared in the form of scatter plots, showing that the accuracy of the hybrid model is significantly improved. Figure 4 In the legend, XAJ represents the distributed Xin'anjiang hydrological model, a traditional physical mechanism-based hydrological model used to simulate basin flood processes. XAJ-LSTM represents a hybrid model based on the Xin'anjiang model and LSTM, namely the distributed hydrological model + Long Short-Term Memory (LSTM) error correction model described in this invention, used to improve flood forecast accuracy by combining physical mechanisms and data-driven methods.

[0085] pass Figure 4It is evident that the NSE of the 28 floods simulated based on the distributed hydrological model and the Long Short-Term Memory (LSTM) deep learning error correction model described in this invention is almost always close to 1 with minimal fluctuations, demonstrating extremely high accuracy and stability. The NSE improved by 29.1% during the training period and by 65.8% during the validation period. The relative error of the flood peak decreased to 7.2%, and the peak occurrence time difference decreased to 1.8 hours.

[0086] The traditional distributed Xin'anjiang hydrological model exhibits extremely large fluctuations in NSE (Neural Sequence of Errors), with NSE values ​​below 0.6 and even approaching 0.2 for many floods. This indicates that the traditional model's forecast accuracy is unstable and prone to significant deviations in watersheds heavily affected by human activities.

[0087] A typical flood in June 2024 (significantly affected by the joint operation of upstream reservoirs) was selected for analysis, such as... Figure 5 As shown, Figure 5 In the legend, the measured flow at Lanxi (blue line) represents the actual flow data observed at the Lanxi hydrological station, serving as the "true value" reference for flood forecasting. XAJ (black line) represents the flow forecast solely by the distributed Xin'anjiang hydrological model, reflecting the forecast results of the traditional physical model. XAJ-LSTM (red line) represents the flow forecast by the hybrid model described in this invention, reflecting the forecast results after coupling physical mechanisms and data-driven approaches.

[0088] from Figure 5 It is evident that the distributed Xin'anjiang hydrological model alone has a large forecasting error, especially the predicted peak flow is significantly lower than the actual measurement; while the forecasting process of XAJ-LSTM almost overlaps with the actual measurement, and the peak height, peak time and receding process are closer to the actual data, indicating that the forecasting accuracy of the hybrid model described in this invention is very high.

[0089] S32: Forecast Performance Analysis with Different Lead Times: Forecast Accuracy of Hybrid Models with Lead Times of 2-18 Hours as shown in... Figure 6 As shown. From Figure 6 It is evident that as the forecast period lengthens, the NSE (Number of Sequences) during both the training and validation periods decreases, which aligns with the general rule in flood forecasting that "the longer the forecast period, the greater the uncertainty and the lower the accuracy."

[0090] For short forecast periods, such as within 2 hours, both have an NSE close to 0.99, indicating extremely high forecast accuracy.

[0091] For the entire forecast period (0-18 hours), the NSE during the validation period was consistently higher than that during the training period, indicating that the hybrid model has strong generalization ability and good stability for "new flood events" that were not included in the training. Even when the forecast period was extended to 18 hours, the NSE remained above 0.91, with a gradual decay rate, indicating that the hybrid model can still maintain high accuracy over long forecast periods (hourly scales), meeting the real-time flood forecasting needs of watersheds heavily affected by human activities.

[0092] As described above, although the invention has been illustrated and described with reference to specific embodiments, it should not be construed as limiting the invention itself. Various changes in form and detail may be made without departing from the spirit and scope of the invention as defined in the appended claims.

Claims

1. A flood forecasting method for watersheds with strong human activity impacts based on a hybrid model, characterized in that, Includes the following steps: Step S1: Construct a hydrological model, which includes the following steps: S11: Sub-basins based on natural water systems: Process the original DEM data to generate river links and use watersheds to divide the forecast basin to obtain initial sub-basins. Combine the distribution and density of rain gauges to adjust the size of the initial sub-basins to obtain the target sub-basins. S12: Construction of river system topology and division of Mustingen river segments: Number the target sub-basins and construct the river system topology. Combine the river length, river gradient and river confluence time to divide the number of Mustingen river segments. S13: Parameter calibration of distributed hydrological model: The parameter calibration method of human-computer interaction is adopted. The parameters of the distributed hydrological model are optimized by using the relative error of runoff depth, the relative error of flood peak, the Nash efficiency coefficient NSE and the peak occurrence time difference as objective functions to obtain the calibrated distributed hydrological model. Step S2: Construct the hybrid model, which includes the following steps: S21: Calculate the simulation error sequence of historical hydrological models: based on historical measured flow rates Q 0 and the historical calculated flow rate output by the distributed hydrological model after calibration Qc The historical hydrological model simulation error sequence was calculated. S22: Determine the input features and feature step size of the deep learning model: use the previous multi-step average precipitation of the basin and the simulation error of the historical hydrological model as input features; among which, the average precipitation of the basin is calculated based on the weight of the target sub-basin; The river propagation time of the target sub-basin farthest from the forecast section is used as the characteristic step size. T L The river propagation time is calculated by the number of Muskingen sections and the propagation time KE of a single Muskingen section; S23: Construct a deep learning error correction model: Based on the input features and feature step size in step S22, construct a sample set, divide the sample set into a training set and a validation set and normalize them respectively, construct the deep learning model network structure, use the Bayesian optimization algorithm to optimize the hyperparameters, and use the training set for training to obtain the deep learning error correction model. Step S3: Real-time flood forecasting: Output the historical calculated flow rate Tc hours before the current time and the initial forecast flow rate for the next 72 hours through the calibrated distributed hydrological model, where Tc > feature step size; calculate the historical hydrological model simulation error for the corresponding period based on the historical calculated flow rate of the previous feature step size in Tc hours before the current time and the historical measured flow rate for the corresponding period; input the basin average precipitation of the previous feature step size and the historical hydrological model simulation error into the deep learning error correction model to calculate the forecast error at the current time and continuously obtain the forecast error for multiple future forecast periods; combine the initial forecast flow rate with the forecast error for the corresponding forecast period to obtain the corrected forecast flow rate.

2. The flood forecasting method for watersheds with strong human activity impacts based on a hybrid model according to claim 1, characterized in that, In step S11, the original DEM data is processed, specifically including: performing mosaicking, depression filling, flow direction extraction, and confluence accumulation calculation operations on the original DEM data in sequence; The initial sub-basin size is adjusted based on the distribution and density of rain gauges. Specifically, this includes merging initial sub-basins with an area smaller than a first preset threshold, and manually dividing initial sub-basins with an area larger than a second preset threshold, where the second preset threshold is greater than the first preset threshold.

3. The flood forecasting method for watersheds with strong human activity impacts based on a hybrid model according to claim 1, characterized in that, In step S12, the number of Muskingan river segments is determined by combining the river length, river gradient, and river confluence time. Specifically, this includes: taking the river confluence time as the core constraint, and combining the confluence distance corresponding to the river length and the confluence velocity corresponding to the river gradient, determining the number of Muskingan river segments from each target sub-basin to the forecast section, so as to ensure that the flood wave propagation process matches the actual river confluence pattern.

4. A flood forecasting method for watersheds with strong human activity impacts based on a hybrid model according to any one of claims 1-3, characterized in that, In step S13, the objective function is calculated based on historical measured flow rates. Q 0 and the historical calculated flow rate output by the distributed hydrological model after calibration Qc The formulas for calculating each objective function are as follows: relative error of runoff depth ;in, R 0 represents the flow rate based on historical measured data. Q 0. Calculated measured runoff depth. R C To calculate traffic based on historical data Qc Calculated predicted runoff depth; relative error of flood peak ;in, Based on historical measured flow Q The measured peak flow rate extracted from 0 To calculate traffic based on historical data Qc Extracted forecast peak flow; Nash efficiency coefficient ;in, Q 0( i ) is the first i Historical measured flow rate at any given time; Q c ( i ) is the first i Historical flow calculation at any given moment; Historical measured flow rate Q The mean of 0; N The length of the flood sequence; Peak time difference Δ T = T 0- T c ;in, T 0 represents the flow rate based on historical measured data. Q 0. The exact measured time of the flood peak. T c To calculate traffic based on historical data Qc The predicted time of the flood peak has been determined.

5. The flood forecasting method for watersheds with strong human activity impacts based on a hybrid model according to claim 4, characterized in that, In step S21, the formula for calculating the simulation error of the historical hydrological model is: Δ Q sim ( i )= Q 0( i )- Q c ( i ), where Δ Q sim ( i ) is the first i Historical hydrological model simulation errors at any given time Q 0( i ) is the first i Historical measured flow rate at any given time Q c ( i ) is the first i Historical flow calculation at any given moment.

6. The flood forecasting method for watersheds with strong human activity impacts based on a hybrid model according to claim 5, characterized in that, In step S22, the formula for calculating the average precipitation in the watershed is: ; in, For the first i The average precipitation in the basin at any given time α j For the first j The weight of each sub-basin P j ( i ) is the first i Time of the first j Precipitation in each sub-basin m Indicates the number of sub-basins; The formula for calculating river propagation time is: T L =nKE ,in, T L The river propagation time is the time from the target sub-basin furthest from the forecast section. n This represents the number of Muskingan river segments in the distributed hydrological model. KE This refers to the river propagation time for a single river segment.

7. A flood forecasting method for watersheds with strong human activity impacts based on a hybrid model, as described in claim 6, is characterized in that... In step S23, the sample set is constructed as follows: based on historical moments... i Before T L Average precipitation in the basin over a given period And the simulation error of historical hydrological models {Δ Q sim ( iT L ),…,Δ Q sim ( i- 1)} is used as the input sample, with historical moments i The historical hydrological model simulation error is used as the output sample to form sample pairs (Ω, γ ),in, ； γ= {Δ Q sim ( i )};where Ω is the input sample set, γ is the output sample set, and the historical time i satisfy i ≥TL.

8. A flood forecasting method for watersheds with strong human activity impacts based on a hybrid model, as described in claim 7, is characterized in that... In step S23, the deep learning model is an LSTM model. The network structure of the LSTM model includes an input layer, a first LSTM layer, a first Dropout layer, a second LSTM layer, a second Dropout layer, and an output layer. The parameters of each layer are set as follows: Input layer: Input dimension is 2× T L ; First LSTM layer: The number of neurons is set to the first preset value, the activation function is tanh, and the return sequence parameter is set to True; First Dropout layer: The output feature matrix of the first LSTM layer, with the Dropout rate set to the first preset rate; Second LSTM layer: The number of neurons is set to a second preset value, which is less than the first preset value. The activation function is tanh, and the return sequence parameter is set to False. Second Dropout layer: The output feature vector of the second LSTM layer is processed, and the Dropout rate is set to the second preset rate; Output layer: The number of neurons is set to 1, and the activation function is a linear activation function.

9. A flood forecasting method for watersheds with strong human activity impacts based on a hybrid model, as described in claim 8, is characterized in that... In step S23, optimizing the hyperparameters of the deep learning model and training it specifically involves: using a Bayesian optimization algorithm with the objective function of maximizing the Nash efficiency coefficient (NSE) of the validation set, iteratively optimizing the hyperparameters, which include the number of neurons in the first LSTM layer, the number of neurons in the second LSTM layer, the first preset rate, the second preset rate, the learning rate, and the number of training iterations. The Adam optimizer is used, with the error value predicted by the LST model and the simulation error Δ from the historical hydrological model as the benchmarks. Q sim ( i The mean squared error of the model is used as the loss function. The LSTM model is trained using the training set. When the mean squared error of the validation set does not decrease within a preset number of consecutive rounds, or when the optimized number of training iterations is reached, the training is terminated, and the trained deep learning error correction model is obtained.

10. A flood forecasting method for watersheds with strong human activity impacts based on a hybrid model, as described in claim 9, is characterized in that... Step S3 specifically includes the following steps: S31: Initial Forecast Output: Outputs the current forecast using the calibrated distributed hydrological model. k Flow rate calculated hourly over the previous Tc hours { Q c ( k - T c ),…, Q c ( k -1)} and the initial hourly forecast flow for the next 72 hours Q pre,k+ε , ε Forecast period, ε =1,2,...,72; where Tc > characteristic step size T L ; S32: Calculation of Historical Simulation Error Sequence: Based on the Current Moment k The first Tc hour T L Historical flow calculation by step size { Q c ( k - T L ),…, Q c ( k -1)} and the historical measured flow rate for the corresponding time period { Q 0( k - T L ),…, Q 0( k -1)}, calculate the historical hydrological model simulation error subset {Δ Q sim ( k - T L ),…,Δ Q sim ( k -1)}; S33: Error Prediction and Rolling Iteration: First, the previous... T L The time-by-time average precipitation of the basin with characteristic step size and the subset of historical hydrological model simulation errors {Δ Q sim ( k - T L ),…,Δ Q sim ( k -1)}, together input into the deep learning error correction model, to calculate the current time step k The forecast error value Δ Q pred ( k ); Then, based on the basin average precipitation forecast for the next 72 hours. , with Δ Q pred ( k Using Δ as the initial value, iteratively input the deep learning error correction model: calculate Δ Q pred ( k When +1 is entered, input ; Calculate Δ Q pred ( k When +2), enter ; and so on, successively obtain the forecast error set {Δ} for each forecast period in the next 72 hours. Q pred ( k +1),…,Δ Q pred ( k +72)}; S34: Final Forecast Flow Calculation: Calculate the initial forecast flow output from the distributed hydrological model. Q pre,k+ε Forecast error Δ with the corresponding lead time Q pred ( k + ε According to the formula Q cor,k+ε = Q pre,k+ε +Δ Q pred ( k + ε The corrected final forecast flow is obtained by time-by-time matching calculation. Q cor,k+ε .

Images