Deep learning-driven flood forecasting method based on runoff generation and confluence processes

Abstract

This invention discloses a rainfall-driven deep learning flood forecasting method based on runoff generation and confluence processes. The method includes: calculating the net rainfall sequence for each rain gauge control area based on its preceding rainfall sequence; allocating the net rainfall sequence for each rain gauge control area along the time dimension and calculating its contribution to the runoff at the basin outlet; multiplying the runoff contribution of each rain gauge control area to the basin outlet by the area of ​​the control area and adding the baseflow to obtain the final forecasted runoff. Based on runoff generation and confluence processes, this invention proposes a novel rainfall-driven deep learning framework for runoff generation and confluence, establishing a convolutional weighted flood forecasting model method with a physical mechanism. The effectiveness of the model was verified, and the results show that the established convolutional weighted model has excellent flood forecasting performance and its parameters have practical physical meaning. The runoff generation and confluence coefficients determined by the model can well reflect the runoff generation and confluence characteristics of the basin.

Description

Technical Field

[0001] This invention relates to the field of flood forecasting technology, and more specifically to a deep learning-driven flood forecasting method based on runoff generation and confluence processes. Background Technology

[0002] As one of the most destructive natural disasters, floods have long posed a serious threat to human survival and social development. According to UN statistics, 40% of global economic losses are attributed to floods and their secondary disasters. Between 1995 and 2015, 157,000 people died from floods, and 2.3 billion people were affected, accounting for 56% of all people affected by weather-related natural disasters. Flood prevention and control are divided into engineering and non-engineering measures. Among non-engineering measures, flood forecasting is a crucial method for effectively avoiding casualties and reducing flood damage. To date, an increasing number of flood forecasting models have been developed and applied, including traditional physics-based methods and data-driven methods. Physics-based methods, requiring data from meteorology, surface vegetation, topography, and underground soil, are more difficult to apply to real-world environments. Data-driven methods, on the other hand, mostly require only rainfall and runoff data, making them easier to apply in real-world situations and providing more accurate forecasts. Nowadays, data-driven flood forecasting models based on deep learning technology are receiving increasing attention, and their forecasting effectiveness and capabilities have been unanimously recognized by the industry. Among them, models represented by LSTM and attention have performed the best.

[0003] However, most current deep learning flood forecasting models are black-box models, with unclear internal physical mechanisms and parameters lacking physical meaning, resulting in insufficient reliability and stability when used for flood forecasting. Currently, there is no research directly integrating runoff generation and confluence mechanisms into deep learning models for flood forecasting; building deep learning flood forecasting models based on hydrophysical mechanisms will be a future research trend. Summary of the Invention

[0004] To address the aforementioned technical issues, this invention primarily explores deep learning flood forecasting models that use prior rainfall data as input. It also incorporates the concept of runoff generation and confluence during the modeling process, ultimately establishing a rainfall-driven deep learning flood forecasting method based on runoff generation and confluence processes.

[0005] According to a first aspect of the present invention, the present invention provides a rainfall-driven deep learning flood forecasting method based on runoff generation and confluence processes, wherein a rainfall-driven deep learning runoff generation and confluence model is constructed, the current time is set as t, and the previous rainfall sequence is used as the basis for the forecast. As input to the deep learning runoff generation and confluence model, the runoff at time t+1 in the future is predicted, where m is the length of the rainfall input sequence, j = 1, ..., L represents the control area of ​​the j-th rainfall station, and L represents the total number of control areas of rainfall stations.

[0006] Specifically, the following steps are included:

[0007] S1: Based on the previous rainfall sequence of each rainfall station's control area, calculate the net rainfall sequence of the corresponding rainfall station's control area;

[0008] S2: Distribute the net rainfall sequence of each rainfall station's control area in the time dimension to calculate its contribution to the runoff at the watershed outlet;

[0009] S3: Multiply the runoff contribution of each rainfall station's control area to the watershed outlet by the area of ​​the rainfall station's control area, and then add the baseflow to obtain the final forecast runoff.

[0010] Furthermore, step S1 specifically includes:

[0011] S1.1: Obtain the runoff coefficient w of the control area of ​​the j-th rainfall station. j ;

[0012] S1.2: Obtain the rainfall in the control area of ​​the j-th rainfall station in the rainfall sequence during the time period t-i+1. i = 1, ..., m, j = 1, ..., L, where m represents the length of the rainfall sequence and L represents the total number of rainfall stations;

[0013] S1.3: Obtain the bias term b of the final net rainfall in the control area of ​​the j-th rainfall station. j ;

[0014] S1.4: Based on the production flow coefficient w j Rainfall and bias term b j The net rainfall in the control area of ​​the j-th rainfall station is calculated.

[0015] Furthermore, in step S1.4, the specific calculation formula is as follows:

[0016]

[0017] Among them, the flow generation coefficient w j and bias term b j These are the parameters of a 1×1 convolutional network for the control area of ​​the j-th rain gauge station. For the same rain gauge station control area, w j and b j It does not change over time.

[0018] Furthermore, step S2 specifically includes:

[0019] S2.1: Generate a set of random parameters to be calibrated The flow coefficient is calculated by passing it into the softmax function using the following formula.

[0020]

[0021] S2.2: The net rainfall sequence With confluence coefficient The contribution sequence of net rainfall from time t-m+1 to time t to runoff at time t+1 is obtained by performing element-wise multiplication in reverse order. The calculation formula is as follows:

[0022]

[0023] S2.3: Summing the contribution sequence yields the contribution of rainfall in the controlled area of ​​the rainfall station to runoff at time t+1. The calculation formula is as follows:

[0024]

[0025] Furthermore, the sum of all the confluence coefficients is 1, and each confluence coefficient does not change over time.

[0026] Furthermore, step S3 specifically includes:

[0027] The contribution of different rainfall station control areas to runoff at time t+1 Multiplying (j=1,…,L) by the area of ​​its corresponding control area and adding the baseflow yields the final predicted runoff. The specific calculation formula is as follows:

[0028]

[0029] in, The unit is mm / h, A j This represents the area controlled by the j-th rainfall station, in km². 2 f b Represents the base current, unit m 3 / s,f t+1 This is the final predicted runoff volume, in cubic meters (m³). 3 / s.

[0030] Furthermore, the baseflow is the runoff volume prior to the occurrence of the flood event.

[0031] According to a second aspect of the present invention, the present invention provides a rainfall-driven deep learning flood forecasting device based on runoff generation and confluence processes, used to build a rainfall-driven deep learning runoff generation and confluence model, setting the current time as t, and using the previous rainfall sequence... As input to the deep learning runoff generation and confluence model, the runoff at time t+1 in the future is predicted, where m is the length of the rainfall input sequence, j = 1, ..., L represents the control area of ​​the j-th rainfall station, and L represents the total number of control areas of rainfall stations.

[0032] The deep learning-based flood forecasting device specifically includes the following modules:

[0033] The runoff generation module is used to calculate the net rainfall sequence of the corresponding rainfall station control area based on the previous rainfall sequence of each rainfall station control area;

[0034] The runoff module is used to distribute the net rainfall sequence of each rain gauge control area in the time dimension and calculate its contribution to the runoff at the watershed outlet.

[0035] The runoff calculation module is used to multiply the runoff contribution of each rainfall station's control area to the watershed outlet by the area of ​​the rainfall station's control area, and then add the baseflow to obtain the final forecast runoff.

[0036] According to a third aspect of the present invention, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the deep learning flood forecasting method.

[0037] According to a fourth aspect of the present invention, a storage medium is provided thereon storing a computer program, characterized in that the computer program, when executed by a processor, implements the steps of the deep learning flood forecasting method.

[0038] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0039] This invention provides a novel flood forecasting method based on deep learning. Based on the acquired antecedent rainfall sequences of each rain gauge station's control area, the net rainfall sequence for that control area is calculated. The net rainfall sequences of each control area are distributed over time, and their contribution to the runoff at the watershed outlet is calculated. The result of the net rainfall distribution is multiplied by the area of ​​the control area, and then the baseflow is added to obtain the final forecasted runoff. Based on the watershed runoff generation and confluence process, this invention proposes a novel rainfall-driven deep learning runoff generation and confluence model framework. Based on this framework, a convolutional weighted flood forecasting model with physical mechanisms is established. The effectiveness of the model was verified in the Yutan, Chenda, and Xinqiaozi watersheds. Experimental results show that the established convolutional weighted model has excellent flood forecasting performance, and its parameters have practical physical meaning. The runoff generation and confluence coefficients determined by the model can well reflect the runoff generation and confluence characteristics of the watershed. Attached Figure Description

[0040] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings:

[0041] Figure 1 This is a schematic diagram of the deep learning-based merge model framework in an embodiment of the present invention;

[0042] Figure 2 This is a schematic diagram of a 1D 1×1 convolutional network used in the flow generation process in this embodiment of the invention.

[0043] Figure 3 This is a schematic diagram of softmax weighted merging used in the merging process in this embodiment of the invention;

[0044] Figure 4 This embodiment of the invention studies the location of the watershed and the hydrological station / rain gauge.

[0045] Figure 5 This is a schematic diagram illustrating the division of control areas for rainfall stations in the Yutan and Chendazi watersheds in this embodiment of the invention.

[0046] Figure 6 This is a comparison of the flood flow process predicted and observed by the convolutional weight model at Yutan Station in this embodiment of the invention: (a) flood event 6, (b) flood event 10, and (c) flood event 51.

[0047] Figure 7 This is a comparison of flood flow processes predicted and observed by the convolutional weight model at Chen Dazhan in this embodiment of the invention, (a) flood event 39, (b) flood event 41, and (c) flood event 45;

[0048] Figure 8This is a comparison of the flood flow process predicted and observed by the convolutional weight model at Xinqiao Station in this embodiment of the invention: (a) flood event 28, (b) flood event 29, and (c) flood event 32.

[0049] Figure 9 This is a heat map of the confluence coefficient distribution determined by the Yutanzi watershed model in this embodiment of the invention;

[0050] Figure 10 This is a heat map of the confluence coefficient distribution determined by the Chen Dazi watershed model in this embodiment of the invention;

[0051] Figure 11 This is a heat map of the confluence coefficient distribution determined by the Xinqiaozi watershed model in this embodiment of the invention;

[0052] Figure 12 This describes the distribution characteristics of the confluence coefficients in the Yutan, Chenda, and Xinqiaozi watersheds in this embodiment of the invention. Detailed Implementation

[0053] Various exemplary embodiments of the present invention will now be described in detail with reference to the accompanying drawings. It should be noted that, unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps set forth in these embodiments do not limit the scope of the present invention.

[0054] At the same time, it should be understood that, for ease of description, the dimensions of the various parts shown in the accompanying drawings are not drawn according to actual scale.

[0055] The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the invention or its application or use.

[0056] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0057] Techniques, methods, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and equipment should be considered part of the specification.

[0058] In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values.

[0059] Deep learning confluence framework

[0060] This invention provides a deep learning-driven runoff generation and flow convergence (DL-RGFC) framework that is solely driven by rainfall. Figure 1 As shown, the DL-RGFC framework comprises three modules: a runoff generation module, a runoff collection module, and a runoff calculation module. Figure 1 In this study, L represents the number of rainfall observation stations in the watershed. The runoff generation and runoff collection modules are calculated separately for the control areas of these L rainfall stations. The runoff calculation module, however, aggregates the runoff generation and runoff collection module results from each of the L rainfall stations' control areas using a specific algorithm to calculate the runoff at the sub-watershed outlet. In this invention, the current time is set as t, and the deep learning module for runoff generation and runoff collection only considers the previous rainfall sequence. (j=1,,L) is used as the input to predict the runoff at time t+1 in the future, where m is the length of the rainfall input sequence.

[0061] Specifically, the workflow of each module is explained as follows:

[0062] Runoff Generation Module: During the process of rainfall forming surface runoff, a portion of the rainfall is lost through various means, including vegetation interception, vadose zone filling, depression filling, transpiration and evaporation, and infiltration to replenish groundwater. After removing these losses, the remaining rainfall ultimately forms runoff; this portion of rainfall that contributes to runoff is called net rainfall. In hydrology, the process of calculating net rainfall from rainfall is called watershed runoff generation (Yang et al., 2014). In the runoff generation module of the DL-RGFC framework, deep learning technology will be used to analyze the antecedent rainfall sequences of the control area at each rainfall station. (j=1,,L) are used as inputs to obtain the net rainfall sequence of the corresponding rainfall station control area. It serves as the input to the bus module.

[0063] Runoff Module: Runoff in a watershed includes hillside runoff and river network runoff. Hillside runoff generally includes surface runoff, interflow runoff, and subsurface runoff. Runoff from each runoff enters the river channel, evolves through the channel, and ultimately forms the watershed outlet runoff. Both hillside runoff and river network runoff redistribute net rainfall over time. Net rainfall forms surface runoff, interflow, and subsurface runoff all the way to the watershed outlet. Due to differences in the distance between each runoff-generating unit and the watershed outlet, the time it takes for runoff to reach the outlet also varies. At any given moment, the runoff at the watershed outlet is actually composed of net rainfall from different times. In the DL-RGFC framework, the runoff module, based on the net rainfall obtained from the runoff-generating module, uses a runoff coefficient to reflect the differences in runoff time, and calculates its contribution to the watershed outlet runoff by distributing net rainfall over time.

[0064] Runoff Calculation Module: Based on the production and runoff confluence modules, the runoff calculation module multiplies the net rainfall distribution of each rainfall station's control area by the area controlled by the rainfall station, and then adds the baseflow to obtain the final forecast runoff.

[0065] Convolutional weight model

[0066] Based on a rainfall-driven deep learning-based runoff generation and confluence model framework, this invention constructs a Convolutional Weighted Model (Conv-Weight). The Conv-Weight model comprises three modules: runoff generation, runoff confluence, and runoff calculation. The runoff generation and runoff confluence modules are computed in parallel at the scale of each rainfall station's control area. Without loss of generality, the runoff generation and runoff confluence models are illustrated using a rainfall station control area j (j = 1,,L) as an example.

[0067] Convolutional Network Flow Generation Module

[0068] The main function of the runoff generation module is to convert rainfall into net rainfall. In hydrology, the ratio of runoff depth to the corresponding average rainfall depth in a watershed over a given period is defined as the runoff coefficient, which is also equal to the ratio of net rainfall to total rainfall. Based on this, this invention introduces a runoff coefficient w into the runoff generation module. j The runoff coefficient is used to approximate the watershed runoff coefficient, and the runoff coefficient multiplied by the rainfall amount gives the net rainfall. Therefore, the net rainfall in the control area of ​​the j-th rainfall station is calculated as shown in equation (1).

[0069]

[0070] In the formula w j , and (i = 1, ..., m) represent the runoff generation coefficient of the control area of ​​the j-th rainfall station, the rainfall amount of the j-th rainfall station in the rainfall input sequence, and the net rainfall amount in the time period t-i+1, respectively. Equation (1) can be implemented by a 1-dimensional 1×1 convolutional network (CNN) in deep learning technology. To increase the flexibility of the convolutional network and make the search space of the convolutional network larger, a bias term b close to 0 is added to equation (1). j The final net rainfall calculation formula is shown in equation (2).

[0071]

[0072] Figure 2 This is a schematic diagram of a 1D 1×1 convolutional network computation, where w j and b j These are the parameters of a 1×1 convolutional network for the control area of ​​the j-th rain gauge station. For the same rain gauge station control area, w j and b jThese parameters do not change over time. Their specific values ​​will be determined during model training.

[0073] In the runoff generation module of the convolutional weight model, the runoff generation coefficient does not change over time. However, in actual rainfall runoff generation, the underlying surface characteristics of the watershed may differ at different times, leading to variations in the runoff coefficient. Even under the same underlying surface conditions, the runoff coefficient changes dynamically at different times within the same flood event (Zhang et al., 2001). In the early stages of rainfall, most of the rainfall is lost through vegetation interception and depression filling, resulting in a small proportion of rainfall that is ultimately converted into net rainfall, leading to a low runoff coefficient. Conversely, in the middle and later stages of rainfall, the proportion of lost rainfall decreases significantly, resulting in a higher runoff coefficient (Wu et al., 2006). Therefore, the fact that the runoff generation coefficient in the runoff generation module of this invention does not change over time inevitably introduces certain errors into flood forecasting. The reason why the runoff generation coefficient remains unchanged is mainly due to the following two reasons: (1) Even if the runoff generation coefficient is considered to be dynamic when training the model, it is impossible to determine the actual runoff coefficient within the forecast period when it is actually applied to flood forecasting. Therefore, the average runoff coefficient trained based on historical data is a feasible choice; (2) In the same flood event, the runoff generation ratio is different in different time periods. Using the average runoff coefficient of this flood can basically ensure that the total net rainfall is accurate. The adjustment in the time dimension when estimating the runoff volume based on the total net rainfall can be handled by the subsequent runoff confluence module.

[0074] softmax weighted merging module

[0075] Watershed runoff primarily addresses the redistribution of net rainfall generated by runoff generation within a watershed, specifically determining the net rainfall for the time period t-i+1 (i = 1, ..., m). The study considers the runoff contribution at times m from t-i+2 to t-i+m+1. A set of net rainfall distribution coefficients is introduced. in Let be the contribution coefficient of the net rainfall in the control area of ​​the j-th rainfall station during a certain period to the runoff at time m thereafter. Other coefficients are derived similarly. The net rainfall is obtained according to equation (3). Contribution of runoff to m times from t-i+2 to t-i+m+1

[0076]

[0077] in This represents the contribution of net rainfall during the period t-i+1 within the control area of ​​the j-th rainfall station to runoff at the subsequent time t-i+2, and so on. Net rainfall distribution coefficient. In this invention, these are also referred to as confluence coefficients. Considering the requirements of water balance, the sum of all confluence coefficients should be 1, and each confluence coefficient should not change over time.

[0078] It is not difficult to see from equation (3) that when i equals 1, When i equals 2 And so on, when i equals m, And sequence This is the net rainfall sequence. The contribution of each to the runoff at time t+1 is given. The relationship between the two is shown in equation (4).

[0079]

[0080] The contribution of the preceding rainfall sequence in the control area of ​​the j-th rainfall station at time t to the total runoff at time t+1. It can be obtained from equation (5).

[0081]

[0082] This invention will obtain the weights through the softmax weight merging module. For the detailed calculation process, see Figure 3 First, a set of random parameters to be calibrated is generated. Then, it is passed into the softmax function to calculate the confluence coefficient using equation (6).

[0083]

[0084] Then the net rainfall sequence With confluence coefficient The contribution sequence of net rainfall from time t-m+1 to time t to runoff at time t+1 is obtained by performing element-wise multiplication in reverse order. Then, the contribution sequence is summed to obtain the contribution of rainfall in the region to runoff at time t+1. The key to the flow combination module is to use the softmax function to calculate the flow combination coefficients while ensuring that the sum of the flow combination coefficients is 1, thereby ensuring that the flow combination process meets the water balance requirements. Random parameters to be calibrated. The flow coefficient will be determined during the model training process and then determined by formula (6).

[0085] The above analysis reveals that the confluence coefficient has two meanings. Its forward order represents the proportion of net rainfall over a certain time period allocated to runoff over the next m time periods, while its reverse order represents the contribution of the net rainfall sequence of length m from the current time period to the runoff at the next time period. The confluence module of the convolutional weight model calibrates the confluence coefficient based on its reverse order meaning.

[0086] Runoff Calculation Module

[0087] The runoff generation and confluence modules perform calculations separately for each rainfall station's control area. The main purpose of the runoff calculation module is to sum the calculation results from each rainfall station's control area to obtain the final watershed outlet runoff. Specifically, the calculation process involves calculating the contribution of different rainfall station control areas to the runoff at time t+1. Multiplying (j=1,,L) by the area of ​​the corresponding control area and adding the baseflow yields the final runoff. The calculation process is shown in equation (7), where, The unit is mm / h, A j This represents the area of ​​the j-th control region, in km². 2 f b Represents the base current, unit m 3 / s,f t+1 It is the target runoff volume, in cubic meters (m³). 3 / s. The baseflow is the runoff before the flood event.

[0088]

[0089] In the convolutional weighted model, baseflow data is incorporated to reflect hydrological information such as rainfall and soil moisture prior to flooding, which aids in flood discharge calculation. Furthermore, using baseflow data does not impose limitations on flood forecasting models as it does with using previous runoff sequences. First, baseflow in a watershed generally does not change significantly over a given period, making baseflow data easier to obtain than real-time runoff data. This invention uses runoff values ​​prior to flooding, but multi-year average runoff values ​​could also be used. Second, there is no significant correlation between baseflow and predicted runoff; therefore, incorporating baseflow does not weaken the flood forecasting model's ability to simulate rainfall-runoff relationships.

[0090] The runoff generation, confluence, and runoff calculation modules of the convolutional weight model operate as a whole, satisfying the water balance condition throughout the entire runoff calculation process. If any module breaks the water balance condition, the physical meaning of the parameters in other modules will also change. For example, if the sum of the confluence coefficients is not 1, the confluence module will not be able to satisfy the water balance condition, and consequently, the runoff generation coefficient will not be able to learn the true runoff coefficient of the watershed under study.

[0091] The key technical point of this invention is to utilize deep learning technology to accurately simulate the runoff generation and confluence processes in rainfall-runoff formation, thus giving the parameters used in flood forecasting with deep learning methods real hydrophysical significance. Specifically, firstly, a convolutional neural network is used to simulate the runoff coefficient, enabling the calculation of net rainfall from precipitation amount; secondly, the softmax function is used to simulate the redistribution of net rainfall during the confluence process.

[0092] The beneficial effects of the method of the present invention are verified through three specific embodiments below.

[0093] Study area, data, and model methodology settings

[0094] The study area was selected from three sub-basins of the Shaxi River Basin in Fujian Province, China: Yutan, Chenda, and Xinqiao. Figure 4 The location of the three sub-basins and the locations of the hydrological and rainfall stations within each basin are shown. The Xinqiao sub-basin has only one rainfall station, Xinqiao, so it is considered a rainfall station-controlled area in this study. The Yutan and Chendazi basins contain four and two rainfall stations respectively, and are divided into four and two rainfall station-controlled areas respectively using the Thiessen polygon method. Figure 5 The controlled areas of rainfall stations in the Yutan, Chenda, and Xinqiaozi watersheds are shown in Table 1.

[0095] Table 1. Rainfall monitoring station control area information for the Yutan, Chenda, and Xinqiaozi watersheds.

[0096]

[0097] Through data collection and preprocessing, 55, 49, and 33 flood events were generated for the Yutan, Chenda, and Xinqiao sub-basins, respectively, corresponding to 8022, 8272, and 3704 research data sets, each containing hourly runoff and hourly stationary rainfall data. The training, test, and validation sets for the Yutan, Chenda, and Xinqiao sub-basins were partitioned chronologically in a 3:1:1 ratio. When establishing convolutional weighted flood forecasting models and comparative models for the three sub-basins, considering the relatively long time required for rainfall to pass through interflow and groundwater outflow, the input sequence length for all previous rainfall events was set to 30, even without prior runoff input. Input and output samples were generated using the sliding pane method based on historical observation data. The dataset partitioning and sample details for the three basins are shown in Table 2.

[0098] Table 2 Dataset Division and Sample Details for Yutan, Chenda, and Xinqiao Watersheds

[0099]

[0100] The loss function and optimizer used in the modeling process were MSE and Adam, respectively. The batch size was set to 16. Early stopping was used to prevent overfitting during model training, with a maximum of 500 iterations and a tolerance of 50. When calibrating the model parameters using the training set, all models were run 5 times. Correspondingly, the test models were also run 5 times. Subsequent results analysis focused on the average of the results from the 5 test model runs.

[0101] Analysis of flood forecasting effectiveness

[0102] Table 3 shows the flood forecasting results of the convolutional weight model in the Yutan, Chenda, and Xinqiao sub-basins. As can be seen from Table 3, the established model performs relatively well in the Yutan sub-basin, with a mean NSE of 0.837 when only prior rainfall is used as input. The flood forecasting results in the Chenda and Xinqiao sub-basins are slightly worse, with mean NSE values ​​of 0.662 and 0.605, respectively. This is mainly due to the poor quality of the observational data in the Chenda and Xinqiao sub-basins. Using three flood events from each of the three sub-basins as representatives, a comparison chart of the convolutional weight model's forecasts and actual observed flow processes for the selected three flood events is plotted, as shown below. Figures 6 to 8 As shown in the figure. Overall, the established convolutional weight model's forecasts of peak flow, peak timing, and flood events in the Yutan, Chenda, and Xinqiaozi watersheds are close to the observed values. The above analysis results indicate that the established convolutional weight model can still achieve satisfactory results in flood forecasting in the study area when only prior rainfall data is used as input.

[0103] Table 3. Statistics of prediction results from the convolutional weight model in the Yutan, Chenda, and Xinqiao watersheds.

[0104]

[0105] Table 4 presents the runoff coefficients, mean, maximum, and minimum values ​​of the control areas of each rainfall station determined during the five training iterations of the convolutional weight model for the Yutan, Chenda, and Xinqiao sub-basins. Based on the area of ​​the control area of ​​each rainfall station, the runoff coefficients for the Yutan, Chenda, and Xinqiao sub-basins were weighted to obtain runoff coefficients of 0.41, 0.36, and 0.4, respectively. The mean runoff coefficients for each sub-basin were calculated based on the rainfall and runoff volumes of the observed flood events in the three basins, with mean values ​​of 0.39, 0.4, and 0.43. The differences between the runoff coefficients determined by the model and the runoff coefficients obtained from the statistical analysis of measured data for the three sub-basins are only 0.02, 0.04, and 0.03, respectively. This indicates that the runoff coefficients determined by training the convolutional weight model are very close to the runoff coefficients calculated from actual observations, fully demonstrating that the runoff coefficients of the convolutional weight model based on the physical mechanism of runoff generation and confluence processes have practical physical significance, and that the convolutional weight model has excellent performance in capturing the runoff generation process of the basin.

[0106] Table 4. Runoff coefficients determined by the convolutional weight model during the training period in the Yutan, Chenda, and Xinqiaozi watersheds.

[0107]

[0108] As mentioned earlier, in the established convolutional weight model, the confluence coefficient has two meanings: firstly, it represents the proportion of net rainfall in a certain period of time allocated to runoff in the next m time intervals; secondly, it represents the contribution of the input net rainfall sequence before the current time interval to the runoff in the next time interval. To explore the physical interpretability of the confluence coefficient introduced in the model, the distribution maps of the confluence coefficients of each rainfall station control area determined during the five training runs of the Yutan, Chenda, and Xinqiaozi watersheds were plotted. The results are as follows: Figures 9 to 11 As shown in the figure, the high values ​​of the confluence coefficients in the Yutan, Chenda, and Xinqiaozi watersheds are concentrated in α. 8 -α 14 α 3 -α 6 and α 2 -α 5 This means that the flood discharge in the three sub-basins at time t+1 will mainly come from the rainfall contributions of the preceding 8-14 hours, 3-6 hours, and 2-5 hours. The average times of the peak flood discharge lagging behind the peak rainfall for these three sub-basins are 10.1 hours, 4.6 hours, and 3.6 hours, respectively. The high values ​​of the confluence coefficients determined by the convolutional weight model correspond to time periods that cover the peak flood discharge lagging behind the peak rainfall calculated from historical observation data. This indicates that the confluence coefficients determined by the model accurately reflect the distribution of net rainfall, effectively capture the confluence characteristics of the basins, and have excellent physical interpretability.

[0109] Furthermore, the runoff coefficients of each rainfall station's control area are weighted and summed based on the area of ​​the control area to obtain the runoff coefficient of the entire sub-basin. The result is as follows: Figure 12 As shown, the maximum confluence coefficients of the Yutan, Chenda, and Xinqiao sub-basins occur at points 11, 4, and 3, respectively, which almost perfectly matches the peak lag times of the three sub-basins obtained from statistics. From Figure 12 It can also be seen that the distribution of the runoff coefficient in the Yutanzi watershed is relatively flat with a small peak, while the runoff coefficients in the Chenda and Xinqiaozi watersheds have large peaks and significant differences between them. This is related to the relatively large area of ​​the Yutanzi watershed (769 km²). 2 ), Chen Da (area 167km²) 2 ) and New Bridge (area 61km²) 2 The characteristics of watersheds with smaller sub-basin areas are consistent: large watershed area, long confluence time; small watershed area, short confluence time.

[0110] The above analysis results show that the parameters of the convolutional weight model established in the three sub-basins of Yutan, Chenda and Xinqiao have good physical meaning. The runoff generation coefficient corresponds to the basin runoff coefficient, and the runoff concentration coefficient can capture the runoff characteristics of the basin well. The flood forecast convolutional weight model based on the runoff generation and confluence process has good physical interpretability.

[0111] In addition, as an optional implementation, the present invention provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the deep learning flood forecasting method.

[0112] In addition, as an optional implementation, the present invention provides a storage medium storing a computer program, characterized in that the computer program, when executed by a processor, implements the steps of the deep learning flood forecasting method.

[0113] The embodiments of the present invention provide a novel flood forecasting method, apparatus, device, and storage medium based on deep learning. Based on the runoff generation and confluence processes in a watershed, a novel rainfall-driven deep learning runoff generation and confluence model framework is proposed. Based on this framework, a convolutional weighted flood forecasting model method with physical mechanisms is established. Furthermore, the effectiveness of the model was verified in the Yutan, Chenda, and Xinqiaozi watersheds. Experimental results show that the established convolutional weighted model has excellent flood forecasting performance, and its parameters have practical physical meaning. The runoff generation and confluence coefficients determined by the model can well reflect the runoff generation and confluence characteristics of the watershed.

[0114] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0115] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments. In the unit claims listing several devices, several of these devices may be embodied by the same hardware item. The use of the terms first, second, and third, etc., does not indicate any order and can be interpreted as identifiers.

[0116] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A rainfall-driven deep learning flood forecasting method based on a runoff process, characterized in that, A rainfall-driven deep learning runoff-concentration model is built, and the current time is set as t, and the previous rainfall sequence As the input of the deep learning runoff-concentration model, the runoff at the future time t+1 is predicted, where m is the length of the rainfall input sequence, j=1,…,L represents the jth rainfall station control area, and L represents the total number of rainfall station control areas. Specifically comprising the following steps: S1: According to the antecedent rainfall sequence of each rainfall site control area, the net rainfall sequence corresponding to the rainfall site control area is calculated; S2: The net rainfall sequence of each rainfall site control area is distributed in the time dimension to calculate its runoff contribution to the watershed outlet; S3: The runoff contribution of each rainfall site control area to the watershed outlet is multiplied by the area of the rainfall site control area, and then the base flow is added to obtain the final predicted runoff; Step S2 specifically comprises: S2.1: Generate a set of random parameters to be rated The flow concentration factor is calculated by substituting into the softmax function by the following equation S2.2: The net rainfall sequence is obtained by subtracting the antecedent precipitation index from the rainfall sequence with the confluence coefficient The contribution sequence of the t-m+1 to t period net rainfall sequence of the jth rainfall station control area to the runoff at time t+1 is obtained by element-by-element multiplication of the inverse sequence of the confluence coefficient The calculation formula is as follows: S2.3: Summing up the contribution sequence to obtain the contribution of the rainfall station control area to the runoff at time t+1 The calculation formula is as follows: Step S3 specifically comprises: The contribution of the different rainfall station control areas to the runoff at time t+1 The final predicted runoff is obtained by multiplying the contribution of the different rainfall station control areas to the runoff at time t+1 by the area of the control area corresponding thereto and adding the base flow, and the specific calculation formula is as follows: wherein, unit is mm / h, A j represents the area of the jth rainfall station control area, unit is km 2 , f b represents the base flow, unit m 3 / s, f t+1 is the final forecast runoff, unit m 3 / s.

2. The rainfall-driven deep learning flood forecasting method based on a runoff process according to claim 1, wherein, Step S1 specifically comprises: S1.1: Obtain the runoff coefficient w of the jth rainfall station control area j ; S1.2: Obtain the rainfall of the jth rainfall station control area in the rainfall sequence in the t-i+1 period j = 1,..., L, m represents the length of the rainfall sequence, and L represents the total number of rainfall stations; S1.3: Obtain the bias term b of the final net rainfall of the jth rainfall station control area j ; S1.4: Calculate the net rainfall of the jth rainfall station control area according to the runoff coefficient w j , rainfall and the bias term b j ​ 3. The rainfall-runoff process-based deep learning flood forecasting method according to claim 2, wherein, In step S1.4, the specific calculation formula is: where w j and bias term b j are parameters of the 1 x 1 convolutional network for the jth rainfall site control area, which do not change over time for the same rainfall site control area. j and b j do not change over time for the same rainfall site control area.

4. The rainfall-runoff process-based deep learning flood forecasting method according to claim 1, wherein, The sum of each confluence coefficient is 1, and each confluence coefficient does not change with time.

5. The rainfall-runoff process-based deep learning flood forecasting method according to claim 1, wherein, The base flow uses the runoff before the flood event occurs.

6. A rainfall-driven deep learning flood forecasting device based on a runoff process for implementing the method of any one of claims 1-5, characterized in that, for building a rainfall-driven deep learning runoff-concentration model, let the current time be t, and the previous rainfall sequence be As the input of the deep learning runoff-concentration model, the runoff at future time t+1 is predicted, where m is the length of the rainfall input sequence, j=1,…,L represents the jth rainfall station control area, and L represents the total number of rainfall station control areas. The deep learning flood prediction device specifically comprises: A runoff generation module for calculating the net rainfall sequence corresponding to each rainfall site control area according to the antecedent rainfall sequence of each rainfall site control area; A confluence module for distributing the net rainfall sequence of each rainfall site control area in the time dimension to calculate its runoff contribution to the watershed outlet; A runoff calculation module for multiplying the runoff contribution of each rainfall site control area to the watershed outlet by the area of the rainfall site control area, and then adding the base flow to obtain the final predicted runoff.

7. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the program to realize the steps of the deep learning flood prediction method according to any one of claims 1-5.

8. A storage medium having stored thereon a computer program, characterized in that The computer program is executed by the processor to realize the steps of the deep learning flood prediction method according to any one of claims 1-5.

Images