Small reservoir flood early warning method based on water conservancy rain measuring radar

Abstract

The invention discloses a small reservoir flood early warning method based on a water conservancy rainfall radar, and relates to the technical field of flood early warning, and the method comprises the following steps: building a radar rainfall calculation model; prediction radar data are obtained through prediction by adopting a short-term rainfall prediction model based on a deep learning model and are input into the radar rainfall calculation model, so that the near extrapolation predicted rainfall is obtained; combining the short-term numerical forecast rainfall with the adjacent extrapolation forecast rainfall to obtain short-term and imminent forecast rainfall; adopting a fusion model based on a long-short-term neural network and a three-water-source Xinanjiang model to obtain a predicted reservoir flood volume; and according to the predicted reservoir flood volume, calculating the positively calculated rain accepting capability of the small reservoir in a rolling manner and carrying out flood early warning. According to the method, the rainfall prediction accuracy and the reservoir flood volume prediction accuracy are improved by combining the short-time numerical forecast rainfall with the near extrapolation forecast rainfall and fusing the long-short-term neural network and the three-water-source Xinanjiang model, so that the reliability of flood early warning is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of flood warning, in particular to a small reservoir flood warning method based on a water conservancy rain measurement radar. BACKGROUND

[0002] Small reservoirs are prone to be damaged due to small engineering scale, low flood control standard, and insufficient flood discharge capacity. In recent years, influenced by global climate change, extreme events such as heavy rain in China have increased and intensified, and the level of climate risk has tended to rise. As a result, the probability and frequency of floods upstream of small reservoirs have also shown a gradually increasing trend. How to improve the warning accuracy and efficiency of small reservoirs and effectively protect the flood control safety of small reservoirs is a problem to be studied.

[0003] In the prior art, Chinese patent CN120028888A discloses a small watershed flood warning method based on an X-band dual-polarization phased array radar, including the following steps: obtaining an X-band dual-polarization phased array radar data set and a rain gauge station measured rainfall data set, establishing a radar rainfall calculation model through linear fitting; obtaining an X-band dual-polarization phased array radar image set, using a short-term rainfall prediction model for prediction to obtain a predicted radar image and corresponding predicted radar data; inputting the predicted radar data into the radar rainfall calculation model to obtain predicted rainfall; and using a flood prediction model for small watershed flood warning.

[0004] However, the rainfall prediction accuracy of the above-mentioned prior art is poor, and the inflow flood volume is not considered for prediction, and the reliability of the flood warning is poor. SUMMARY

[0005] The present application provides a small reservoir flood warning method based on a water conservancy rain measurement radar, to solve the problem of poor rainfall prediction accuracy of the existing flood warning technology, not considering the prediction of inflow flood volume, and poor reliability of flood warning.

[0006] In one aspect, the present application provides a small reservoir flood warning method based on a water conservancy rain measurement radar, including the following steps: Step one, obtaining an X-band dual-polarization phased array radar data set and a rain gauge station measured rainfall data set, and establishing a radar rainfall calculation model through linear fitting.

[0007] Step two, using a short-term rainfall prediction model based on a deep learning model to predict the real-time obtained X-band dual-polarization phased array radar data, obtaining predicted radar data and inputting the radar rainfall calculation model to obtain near extrapolation predicted rainfall.

[0008] Step three, combining the short-time numerical prediction rainfall obtained by the short-time numerical prediction technology with the nowcasting extrapolation prediction rainfall to obtain a short-term and nowcasting prediction rainfall.

[0009] Step four, using a fusion model based on a long short-term neural network and a three-source Xin'anjiang model to obtain a predicted reservoir inflow volume according to the short-term and nowcasting prediction rainfall.

[0010] Step five, rolling calculating the positive calculation rainfall capacity of the small reservoir according to the predicted reservoir inflow volume and performing flood warning.

[0011] In one possible implementation, in step one, data quality control and mixed elevation particle value are performed on the data in the X-band dual-polarization phased array radar data set.

[0012] The data quality control includes: abnormal point removal, secondary echo filtering, electromagnetic interference filtering, ground object echo removal and attenuation correction.

[0013] In one possible implementation, in step two, the short-term and nowcasting rainfall prediction model uses a convolutional neural network model, and the network framework uses a UNet network framework, which includes down-sampling and up-sampling structures.

[0014] In one possible implementation, the UNet network framework is equipped with a convolutional attention module and a depthwise separable convolution.

[0015] In one possible implementation, a generative adversarial network is introduced on the basis of the UNet network framework.

[0016] In one possible implementation, in step three, the short-time numerical prediction rainfall obtained by the short-time numerical prediction technology is combined with the nowcasting extrapolation prediction rainfall in a weighted fusion manner. With the increase of time, the weight of the nowcasting extrapolation prediction rainfall gradually decreases, and the weight of the short-time numerical prediction rainfall gradually increases.

[0017] In one possible implementation, in step four, the fusion process of the fusion model includes: Calculating a preliminary reservoir inflow volume according to the short-term and nowcasting prediction rainfall by using a three-source Xin'anjiang model.

[0018] Learning the dynamic correlation characteristics of the preliminary reservoir inflow volume and the measured reservoir inflow volume in the time dimension by using a long short-term neural network to generate a dynamic error correction model for the preliminary reservoir inflow volume.

[0019] Performing dynamic correction on the preliminary reservoir inflow volume by using the dynamic error correction model to obtain a predicted reservoir inflow volume.

[0020] In a possible implementation, in step five, the positive calculation of the rain storage capacity of the small reservoir is calculated according to the predicted inflow flood volume by using the reservoir rain storage capacity calculation method without considering flood discharge and the reservoir rain storage capacity calculation method with storage and discharge.

[0021] The reservoir rain storage capacity calculation method without considering flood discharge corresponds to the case that the reservoir water level is below the preset water level, and the reservoir rain storage capacity calculation method with storage and discharge corresponds to the case that the reservoir water level is above the preset water level.

[0022] In a possible implementation, in step five, the backstepping rain storage capacity is obtained by using the backstepping flood regulation method in advance, and a rain storage capacity curve is drawn, and then the rain storage capacity curve is corrected according to the positive calculation of the rain storage capacity.

[0023] When the predicted inflow flood volume is greater than the corrected rain storage capacity, a flood warning is performed.

[0024] The small reservoir flood warning method based on the water conservancy rain measurement radar in the application has the following advantages: By combining the short-term numerical prediction rainfall and the near extrapolation prediction rainfall, and fusing the long-short term neural network and the three-source Xin'anjiang model, the rainfall prediction accuracy and the inflow flood volume prediction accuracy are improved, and the reliability of the flood warning is improved.

[0025] The short-term numerical prediction rainfall obtained by using the short-term numerical prediction technology and the near extrapolation prediction rainfall are combined in a weighted fusion manner. With the increase of time, the weight of the near extrapolation prediction rainfall gradually decreases, and the weight of the short-term numerical prediction rainfall gradually increases, realizing the smooth transition of the near extrapolation prediction and the short-term numerical prediction, and improving the seamless prediction level of high spatio-temporal resolution quantitative precipitation.

[0026] The three-source Xin'anjiang model is used to calculate the preliminary inflow flood volume according to the short-term and near-term prediction rainfall. The long-short term neural network is used to learn the dynamic correlation characteristics of the preliminary inflow flood volume and the measured inflow flood volume in the time dimension, and a dynamic error correction model for the preliminary inflow flood volume is generated. The dynamic error correction model is used to dynamically correct the preliminary inflow flood volume, and the predicted inflow flood volume is obtained. The advantages of the physical mechanism driven three-source Xin'anjiang model and the data driven long-short term neural network are deeply complementary, and the prediction accuracy and reliability of the runoff concentration process and the inflow flood volume in the catchment area of the small reservoir are significantly improved.

[0027] The backstepping rain storage capacity is obtained by using the backstepping flood regulation method in advance, and a rain storage capacity curve is drawn, and then the rain storage capacity curve is corrected according to the positive calculation of the rain storage capacity, which improves the reliability of the rain storage capacity curve, and further improves the reliability of the flood warning. BRIEF DESCRIPTION OF DRAWINGS

[0028] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the accompanying drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, the accompanying drawings in the following description only represent some embodiments of the present application, and for those skilled in the art, other drawings can be obtained without creative labor.

[0029] Figure 1 A flowchart of a small reservoir flood early warning method based on a water conservancy rain measurement radar is provided for the embodiments of the present application. DETAILED DESCRIPTION

[0030] The technical solutions in the embodiments of the present application will be described clearly and completely in the following with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments only represent some embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0031] As Figure 1 shown, the embodiments of the present application provide a small reservoir flood early warning method based on a water conservancy rain measurement radar, including the following steps: Step one, obtain an X-band dual-polarization phased array radar data set and a rain gauge station measured rainfall data set, and establish a radar rainfall calculation model through linear fitting.

[0032] Step two, use a short-impending rainfall prediction model based on a deep learning model to predict the real-time obtained X-band dual-polarization phased array radar data, obtain predicted radar data and input the radar rainfall calculation model, and obtain near extrapolation predicted rainfall.

[0033] Step three, combine the short-term numerical prediction rainfall obtained by a short-term numerical prediction technology with the near extrapolation predicted rainfall to obtain short-impending predicted rainfall.

[0034] Step four, use a fusion model based on a long-short term neural network and a three-source Xin'anjiang model to obtain predicted reservoir inflow according to the short-impending predicted rainfall.

[0035] Step five, according to the predicted reservoir inflow, calculate the positive calculation rainfall capacity of the small reservoir and perform flood early warning.

[0036] Exemplarily, in step one, data quality control and mixed elevation angle particle value are performed on the data in the X-band dual-polarization phased array radar data set.

[0037] The data quality control includes: abnormal point removal, secondary echo filtering, electromagnetic interference filtering, ground object echo removal and attenuation correction.

[0038] Specifically, data quality control of radar is the key to ensure the accuracy and reliability of rainfall estimation. When performing data quality control, the accuracy, completeness and consistency of radar reflectivity factor data need to be considered. Common data quality control methods include removing outliers, secondary echo filtering, electromagnetic interference filtering, removing ground echo and attenuation correction, etc. The purpose of these methods is to eliminate noise and errors in the data and improve the reliability and accuracy of the data.

[0039] Mixed elevation particle value is one of the key steps in radar quantitative precipitation estimation process. When performing mixed elevation particle value, the radar elevation, particle size and particle shape need to be considered. Different elevation and different particle size data need to be processed differently to obtain accurate precipitation estimation results. At the same time, the mixed elevation particle value results are verified and corrected, which can improve the reliability and accuracy of the results.

[0040] The radar rainfall estimation model is a process of predicting and monitoring the rainfall in a region by using radar reflectivity factor data obtained from X-band dual-polarization phased array radar (i.e. data in the X-band dual-polarization phased array radar dataset) and rainfall station measured rainfall data (i.e. data in the rainfall station measured rainfall dataset), and through linear fitting analysis and modeling. Z, ZDR, KDP are important parameters of radar reflectivity factor, which represent the intensity, particle size and morphology of precipitation, respectively. Before model establishment, data set needs to be preprocessed, including data cleaning, removing outliers, data segmentation, etc. For 5-minute level rainfall data, its characteristics are large data volume, high acquisition frequency, fast change, etc., which need effective feature extraction and processing to improve the accuracy and reliability of the model.

[0041] Exemplarily, in step two, the short-term rainfall forecast model adopts a convolutional neural network model, and the network framework adopts a UNet network framework, which includes down-sampling and up-sampling structures.

[0042] Specifically, in a macroscopic view, we usually take the radar image of a certain moment of the research area as a two-dimensional image, and the whole rainfall process is composed of a sequence of two-dimensional radar rainfall images, which have strong spatio-temporal correlation between adjacent frames, and can be processed by spatio-temporal sequence prediction method. For short-term rainfall prediction, the observation data at each moment is a two-dimensional MxN radar rainfall image. If the radar rainfall image is divided according to the P-dimensional scale, it can be regarded as a many-to-many spatio-temporal sequence prediction problem. According to such modeling method, i.e. end-to-end processing method, it can be trained and predicted by using deep learning method. This modeling method does not need to select features, but performs semantic segmentation on the whole features to automatically obtain image features, but requires very high computing power, which is suitable for deep learning method simulation and prediction.

[0043] Compared with recurrent neural network, convolutional neural network is computationally efficient, can predict the extrapolated results at one time, does not need to iterate the prediction, and takes less time. Convolutional neural network usually includes four parts of input layer, alternately appearing convolutional layer and pooling layer, and finally output layer, and there may be normalization layer and dropout layer to avoid overfitting in the middle according to needs. A deep and complete convolutional neural network is formed by repeatedly stacking different layers. The working process is as follows: the original image enters the network through the input layer, and the size of the input vector is determined by the size of the original image. The convolutional neural network extracts different features of the image through different convolution kernels, then reduces the data dimension through the pooling layer, and finally the output layer is connected with the fully connected layer to obtain the final output of the network.

[0044] In this embodiment, the convolutional neural network model adopts UNet network framework, down-sampling is used to gradually reveal environmental information, and up-sampling is used to restore detailed information by combining the information of each layer of down-sampling and the input information of up-sampling, and gradually restore the image accuracy. This model is a lightweight but very efficient model.

[0045] Exemplarily, the UNet network framework is equipped with a convolutional attention module and a depth separable convolution.

[0046] Specifically, the convolutional block attention module applies attention mechanism to the channel and spatial dimensions in sequence; attention is a mechanism that amplifies useful signals and suppresses secondary signals, which can guide the network to pay more attention to features that are more important to the task. In this embodiment, the convolutional block attention module is inserted in the skip connection between the encoder and the decoder of the UNet network framework, and a loss function with attention mechanism is used to optimize the network parameters in the training process.

[0047] Specifically, the depth separable convolution divides the traditional convolution operation into a depth convolution and a point convolution, which can reduce the parameters of the model without significantly sacrificing performance. In the present embodiment, the traditional standard convolution is replaced by the depth separable convolution in the encoder and decoder of the UNet network framework. This includes the convolution operation after the convolution layer, the pooling layer and the up-sampling layer. Since the output channel number of the depth separable convolution is the same as the input channel number (in the depth convolution stage), the output channel number is adjusted to match the subsequent layers of the UNet in the point convolution stage, ensuring that the replaced network structure still maintains the U-shaped structure and the skip connection of the UNet.

[0048] Illustratively, a generative adversarial network is introduced on the basis of the UNet network framework.

[0049] Specifically, in the present embodiment, the UNet network framework is taken as the main structure of the generator of the generative adversarial network, and its U-shaped structure and skip connection are kept unchanged. The discriminator of the generative adversarial network adopts a convolutional neural network structure, which is used to distinguish between the generated predicted image and the real image. An adversarial learning strategy is adopted to force the predicted image to be closer to the real image.

[0050] Illustratively, in step three, the short-time numerical prediction rainfall obtained by the short-time numerical prediction technology is combined with the near extrapolation prediction rainfall by using a weighted fusion method. As time increases, the weight of the near extrapolation prediction rainfall gradually decreases, and the weight of the short-time numerical prediction rainfall gradually increases.

[0051] Specifically, the prediction period of the nowcasting extrapolation rainfall provided by the radar rainfall estimation model is 0-3 hours. Due to the limited radar coverage area, the prediction period of the nowcasting extrapolation is short. It is found that the nowcasting extrapolation rainfall in 0-1 hours is relatively reliable, and the nowcasting extrapolation rainfall in 1-3 hours can be used as a reference. The prediction period of the short-term numerical prediction technology is 0-168 hours (7 days). Due to the limitation of model start-up, the prediction effect of the short-term numerical prediction technology in the first period of the prediction period (about 0-3 hours) is not good. In order to solve the defects of the above two kinds of prediction, the two kinds of prediction are fused, and the fusion process rule is: from the starting point, the rainfall prediction in the next 1 hour is all the nowcasting extrapolation rainfall in 0-1 hours provided by the radar rainfall estimation model; the rainfall prediction in the second hour is the nowcasting extrapolation rainfall in 1-2 hours provided by the radar rainfall estimation model and the short-term numerical prediction rainfall each accounting for 50%; the rainfall prediction in the third hour is the nowcasting extrapolation rainfall in 2-3 hours provided by the radar rainfall estimation model accounting for 20%, and the short-term numerical prediction rainfall accounting for 80%; the rainfall prediction after 4 hours is all the short-term numerical prediction rainfall. In other possible embodiments, the weights of the nowcasting extrapolation rainfall and the short-term numerical prediction rainfall can also be adjusted to other values, as long as the weight of the nowcasting extrapolation rainfall gradually decreases and the weight of the short-term numerical prediction rainfall gradually increases with the increase of time.

[0052] Exemplarily, in step four, the fusion process of the fusion model comprises: A three-source Xinanjiang model is used to calculate the preliminary storage flood according to the short-term and nowcasting rainfall.

[0053] A long short-term neural network is used to learn the dynamic correlation characteristics of the preliminary storage flood and the measured storage flood in the time dimension, and generate a dynamic error correction model for the preliminary storage flood.

[0054] The dynamic error correction model is used to dynamically correct the preliminary storage flood to obtain the predicted storage flood.

[0055] Specifically, the Xin'anjiang Three-Source Model, as a lumped hydrological model, calculates watershed evapotranspiration according to a three-layer evapotranspiration model, calculates total runoff generated by rainfall based on the concept of full-storage runoff generation, and uses watershed storage curves to account for the impact of uneven underlying surface on runoff generation area changes. In the runoff calculation stage, total runoff is divided into saturated surface runoff, intercalary water runoff, and groundwater runoff. In the runoff concentration calculation stage, the unit hydrograph method is generally used for surface runoff concentration per unit area, while the linear reservoir method is used for intercalary water runoff and groundwater runoff concentration. River network confluence is generally calculated using the Muskingum method or time-delay method with piecewise continuous calculation. The Xin'anjiang Three-Source Model has certain limitations in hydrological forecasting of small and medium-sized watersheds. The key to the accuracy of the Xin'anjiang model of the Three-Source Water Source lies in the combination of model parameters corresponding to the topography, soil and other characteristics of each small and medium-sized watershed. Therefore, the Xin'anjiang model of the Three-Source Water Source needs to be calibrated to improve the accuracy and reliability of the prediction. This means that in actual operation during the flood season, the hydrological forecasting system usually needs to be updated quickly and automatically, and the calibration operation needs to be performed frequently to ensure its forecast accuracy.

[0056] Long Short-Term Neural Networks (LSNs) are a variant of Recurrent Neural Networks (RNNs) specifically designed to address the difficulties traditional RNNs face in handling long-term dependencies. Compared to conventional RNNs, LSNs are more effective at learning and remembering data patterns over long time spans. This is achieved through three key components: a forget gate, an input gate, and an output gate. The forget gate determines which information is no longer needed and should be forgotten; the input gate determines which new information is worth preserving in the cell structure state; and the output gate helps extract information useful for the current task. These components work together to enable LSNs to continue learning new data without losing earlier signals, offering significant advantages in natural language processing, time series forecasting, and other domains requiring long-term memory capabilities. Due to its optimized internal structure, LSNs also effectively avoid the vanishing or exploding gradient problems during training, making it a popular choice for processing sequential data in deep learning. Compared to traditional physical hydrological models, data-driven and machine learning methods like LSNs differ in their input and output. The main inputs to the model are rainfall data from rain gauge radar, river level / flow information from the target hydrological station and its upstream stations, while the output is the future river level / flow at the target station. This data-driven approach can capture, record, and analyze complex nonlinear relationships, leading to more accurate predictions. However, since the accuracy of the model largely depends on the quality of the training data, if the target watershed lacks hydrological or rainfall data, or if there are significant human factors affecting the flow within the watershed, and the long-term and short-term neural networks lack sufficient learning depth, the accuracy of AI forecasts will be affected.

[0057] Therefore, this embodiment integrates the Xin'anjiang River model with three water sources with a long short-term neural network, specifically as follows: First, the Xin'anjiang River model with three water sources is used to calculate the preliminary inflow of floodwater based on short-term forecast rainfall. This is achieved by using real-time monitoring and forecast rainfall data from hydrological rainfall radar, historical hydrological observation data (such as water level and flow rate), and detailed watershed geographic information (such as topography, soil, and vegetation cover) to calibrate the key physical parameters of the Xin'anjiang River model with high precision, ensuring that the model accurately represents the runoff generation and confluence physical mechanisms of the target watershed. Using fused high-precision rainfall data (integrating measured rainfall from rainfall radar, nowcast extrapolation, and short-term numerical forecasts, i.e., short-term forecast rainfall) as the core input, and employing a 10-minute time step, refined flood forecast calculations are performed on key hydrological stations within the watershed. The output includes key elements such as flood occurrence time, water level, and flow rate (including the preliminary inflow of floodwater). The forecast calculations are automatically updated every 5 minutes to ensure real-time capture of the impact of the latest rainfall information on the flood process.

[0058] Secondly, a long short-term neural network (LSN) is employed to learn the dynamic correlation characteristics between the initial and measured inflow volumes over time, generating a dynamic error correction model for the initial inflow volume. The LSN systematically analyzes and learns the complex nonlinear mapping relationship and dynamic error patterns between the historical forecast sequences output by the Xin'anjiang model from the three water sources and the corresponding measured data sequences from hydrological stations before the forecast start time. This LSN, through deep learning algorithms, deeply mines the dynamic correlation characteristics between forecast and measured values ​​over time (i.e., the dynamic correlation characteristics between the initial and measured inflow volumes), constructing a prediction model capable of accurately capturing the evolution of model errors. Based on these learning results, the LSN generates a dynamic error correction model for the initial inflow volume, providing data-driven intelligent support for subsequent forecast optimization.

[0059] Finally, a dynamic error correction model is used to dynamically correct the initial inflow flood volume to obtain the predicted inflow flood volume. This also includes intelligent real-time correction and optimization of forecast results (especially flood occurrence time, water level, and flow rate).

[0060] For example, in step five, the positive calculation of the rain-carrying capacity of a small reservoir is calculated on a rolling basis according to the predicted inflow flood volume using a method that does not consider flood discharge and a method that calculates the rain-carrying capacity of a reservoir based on storage and discharge.

[0061] The method for calculating the rain-carrying capacity of a reservoir without considering flood discharge corresponds to the case where the reservoir water level is below the preset water level, while the method for calculating the rain-carrying capacity of a reservoir based on storage and discharge corresponds to the case where the reservoir water level is above the preset water level.

[0062] Specifically, the calculation method for the rain-carrying capacity of a reservoir without considering flood discharge is as follows: When the reservoir is at a low water level and has a large rain-carrying capacity, the inflow of water generated by rainfall is within the reservoir's rain-carrying capacity range. The maximum inflow is the rain-carrying capacity ∆V below the reservoir's design flood level. The real-time water level of the reservoir is taken as Z. n The corresponding reservoir capacity, determined from the reservoir capacity curve, is V. n Then ∆V is the reservoir capacity V0 corresponding to the normal water level minus V. n The maximum rainfall carrying capacity (area average rainfall) of the reservoir-controlled watershed P n The calculation method is as follows: .

[0063] .

[0064] Among them, P n ∆V represents the maximum rainfall carrying capacity of the reservoir's controlled watershed, in mm; ∆V represents the maximum inflow, in m³; k represents the runoff coefficient, which is roughly divided into four levels based on the current water content of the underlying surface of the controlled watershed: saturated (0.7), humid (0.6), semi-humid (0.5), and relatively arid (0.4); F represents the area of ​​the reservoir's controlled watershed, in km²; V0 represents the reservoir's capacity corresponding to its normal water level, in m³; V n This represents the reservoir capacity corresponding to the real-time water level, in m³. The method for calculating the reservoir's rain-carrying capacity without considering flood discharge assumes that the reservoir's rain-carrying capacity at the current water level is greater than the cumulative rainfall during this rainfall event. Therefore, the premise for using this method is that there is a certain margin between the current reservoir water level and the flood discharge level.

[0065] Specifically, the calculation method for the rain-receiving capacity of a reservoir based on the principle of "storage and release" is as follows: When the reservoir is at a high water level and its rain-receiving capacity is insufficient, part of the incoming rainfall is discharged through the spillway. If the reservoir's spillway has no gate control (e.g., a fixed overflow weir), or if the gate opening remains constant during the rain-receiving and discharge period, then the reservoir water level and overflow are interrelated—when the inflow is greater than the overflow, the reservoir water level gradually rises, and the overflow increases accordingly; conversely, the overflow decreases. Furthermore, when the reservoir only has the spillway as its sole discharge method, a portion of the inflow is the rain-receiving capacity ∆V below the top of the spillway. a This refers to the reservoir capacity V corresponding to the top elevation of the spillway. d Subtract the real-time water level Z n Corresponding storage capacity V n Part of the total flood discharge, W, is through the reservoir's spillway. p Another portion is the flood detention capacity ∆V, which exceeds the top of the spillway but has not yet overflowed. e That is, the reservoir capacity V0 corresponding to the design flood level minus the reservoir capacity V corresponding to the spillway crest elevation. d Under this condition, the maximum rainfall carrying capacity P of the reservoir-controlled watershed is...n The calculation method is as follows: .

[0066] .

[0067] .

[0068] .

[0069] Where ∆V represents the maximum inflow volume, in m³; ∆V a Indicates the reservoir's stormwater storage capacity below the top of the spillway, in m³; ∆V e V represents the flood detention capacity, in m³; V0 represents the reservoir capacity corresponding to the design flood level, in m³; V d W represents the reservoir capacity corresponding to the spillway crest elevation, in m³. p This represents the total flood discharge of the reservoir, in m³. The method for calculating the rainwater carrying capacity of a reservoir based on storage and discharge incorporates reservoir overflow into the calculation. That is, it is assumed that the reservoir overflow equipment has no gate control (such as a fixed overflow weir), or that the reservoir gate opening remains unchanged under the current dispatch mode, then the relationship between the outflow and the reservoir water level is clear.

[0070] Small reservoirs mostly use fixed overflow weirs for flood discharge. Based on the characteristics of the research object, when the current water level of the research object is low (the reservoir water level is below the preset water level), the reservoir's rainwater carrying capacity can be calculated using a method that does not consider flood discharge. When the current water level of the research object is high (the reservoir water level is above the preset water level), when the water level exceeds the normal storage level during the rainwater collection and flood regulation process, the inflow is the flood discharge, and the maximum discharge flow is the design flood discharge flow. This conforms to the characteristics of reservoir rainwater carrying capacity calculation based on storage, and the reservoir rainwater carrying capacity calculation method based on storage can be used to calculate the rainwater carrying capacity.

[0071] For example, in step five, the reverse calculation method for flood control is used in advance to obtain the reverse rainfall capacity and plot the rainfall capacity curve, and then the rainfall capacity curve is corrected according to the forward calculation rainfall capacity.

[0072] A flood warning is issued when the predicted inflow of floodwater exceeds the corrected rainfall capacity.

[0073] Specifically, firstly, a warning water level is set (based on the current engineering status and flood control needs of the reservoir, a warning water level is set; in this embodiment, the design flood level is selected as the warning water level) and a starting regulation water level are set (based on the reservoir's capacity and operation, different gradient starting regulation water levels are set; in this embodiment, the starting regulation water level for each reservoir is 0-4 meters below the normal storage level, with a gradient of 0.5 meters). The flood regulation calculation back-calculation method is used to calculate the corresponding rainfall carrying capacity (i.e., back-calculate the rainfall carrying capacity) using the warning water level and the starting regulation water level, and a rainfall carrying capacity curve is plotted. Then, the rainfall carrying capacity curve is corrected based on the forward-calculated rainfall carrying capacity.

[0074] Specifically, the flood control calculation back-inference method includes: designing typical flood processes (calculating typical flood processes of the reservoir at various frequencies based on the reservoir's geographical location and catchment area), setting different starting water levels (setting water levels at different gradients below the normal storage level as the starting water levels for flood control calculation based on the reservoir's operation), flood control calculation (using different starting water levels and typical flood processes of different frequencies, obtaining the corresponding highest water level and flood control process through flood control calculation), calculating the flood control process and discharge volume (using the warning water level to calculate the flood control process and corresponding discharge volume corresponding to different starting water levels), and obtaining the back-inferenced rainfall carrying capacity.

[0075] This application's embodiments improve the accuracy of rainfall prediction and reservoir inflow prediction by combining short-term numerical forecast rainfall with near-term extrapolated forecast rainfall, and by integrating long-term and short-term neural networks with the Xin'anjiang Three-Source Water Model, thereby enhancing the reliability of flood warnings.

[0076] The proposed method combines short-term numerical weather prediction (SFM) rainfall with nowcast rainfall using a weighted fusion approach. As time progresses, the weight of nowcast rainfall gradually decreases while the weight of short-term numerical weather prediction gradually increases, achieving a smooth transition between nowcast and short-term numerical weather prediction. This approach can improve the level of seamless quantitative precipitation forecasting with high spatiotemporal resolution.

[0077] The proposed method uses the Xin'anjiang Three-Source Model to calculate the initial inflow flood volume based on short-term forecast rainfall. A Long Short-Term Neural Network (LSTN) is employed to learn the dynamic correlation between the initial and measured inflow flood volumes over time, generating a dynamic error correction model for the initial inflow flood volume. This model is then used to dynamically correct the initial inflow flood volume, yielding the predicted inflow flood volume. This approach achieves a deep synergy between the physically driven Xin'anjiang Three-Source Model and the data-driven LSTN, significantly improving the accuracy and reliability of forecasts for runoff generation and inflow flood volumes within the catchment area of ​​small reservoirs.

[0078] The proposed method involves using a flood control calculation back-calculation method to obtain the back-calculated rainfall capacity and plot the rainfall capacity curve. Then, the rainfall capacity curve is corrected based on the forward-calculated rainfall capacity, which improves the reliability of the rainfall capacity curve and thus enhances the reliability of flood warnings.

[0079] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0080] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A small reservoir flood warning method based on a water conservancy rain measurement radar, characterized in that, The method comprises the following steps: Step one, obtaining an X-band dual-polarization phased array radar data set and a rainfall station measured rainfall data set, and establishing a radar rainfall calculation model through linear fitting; Step two, predicting the real-time obtained X-band dual-polarization phased array radar data by using a short-term rainfall prediction model based on a deep learning model, obtaining predicted radar data and inputting the radar rainfall calculation model to obtain a near extrapolation prediction rainfall; Step three, combining the short-term numerical prediction rainfall obtained by a short-term numerical prediction technology with the near extrapolation prediction rainfall to obtain a short-term and near-term prediction rainfall; Step four, obtaining a predicted reservoir inflow flood volume according to the short-term and near-term prediction rainfall by using a fusion model based on a long short-term neural network and a three-source Xin'anjiang model; Step five, rolling calculating the positive calculation rainfall capacity of a small reservoir according to the predicted reservoir inflow flood volume and performing flood warning.

2. The small reservoir flood warning method based on the water-measuring rain radar according to claim 1, characterized in that, In step one, data quality control and mixed elevation particle value selection are performed on the data in the X-band dual-polarization phased array radar data set. The data quality control comprises: abnormal point removal, secondary echo filtering, electromagnetic interference filtering, ground object echo removal and attenuation correction.

3. The small reservoir flood warning method based on the water-measuring rain radar according to claim 1, characterized in that, In step two, the short-term and near-term rainfall prediction model adopts a convolutional neural network model, and the network framework adopts a UNet network framework, which comprises down-sampling and up-sampling structures.

4. The small reservoir flood warning method based on the water-measuring rain radar according to claim 3, characterized in that, The UNet network framework is equipped with a convolutional attention module and a depth separable convolution.

5. The small reservoir flood warning method based on the water-measuring rain radar according to claim 3, characterized in that, A generative adversarial network is introduced on the basis of the UNet network framework.

6. The small reservoir flood warning method based on the water-measuring rain radar according to claim 1, characterized in that, In step three, the short-term numerical prediction rainfall obtained by the short-term numerical prediction technology is combined with the near extrapolation prediction rainfall by using a weighted fusion method, and with the increase of time, the weight of the near extrapolation prediction rainfall gradually decreases, and the weight of the short-term numerical prediction rainfall gradually increases.

7. The small reservoir flood warning method based on the rainfall radar of water conservancy measurement according to claim 1, characterized in that, In step four, the fusion process of the fusion model comprises: calculating a preliminary reservoir inflow flood volume according to the short-term and near-term prediction rainfall by using a three-source Xin'anjiang model; learning the dynamic correlation characteristics of the preliminary reservoir inflow flood volume and the measured reservoir inflow flood volume in the time dimension by using a long short-term neural network to generate a dynamic error correction model for the preliminary reservoir inflow flood volume; performing dynamic correction on the preliminary reservoir inflow flood volume by using the dynamic error correction model to obtain a predicted reservoir inflow flood volume.

8. The small reservoir flood warning method based on the water-measuring rain radar according to claim 1, characterized in that, In step five, the positive calculation rainfall capacity of a small reservoir is rolling calculated according to the predicted reservoir inflow flood volume by using a reservoir rainfall capacity calculation method without considering flood discharge and a reservoir rainfall capacity calculation method with storage and discharge; the reservoir rainfall capacity calculation method without considering flood discharge corresponds to the case that the reservoir water level is below a preset water level, and the reservoir rainfall capacity calculation method with storage and discharge corresponds to the case that the reservoir water level is above the preset water level.

9. The small reservoir flood warning method based on the water-measuring rain radar according to claim 1, characterized in that, In step five, the inverse rainfall capacity is obtained by using a flood regulation calculation backstepping method in advance, and a rainfall capacity curve is drawn, and then the rainfall capacity curve is corrected according to the positive calculation rainfall capacity; when the predicted reservoir inflow flood volume is greater than the corrected rainfall capacity, flood warning is performed.

Images