Radar echo extrapolation method and system based on enhanced recurrent neural network

Abstract

The present application discloses a radar echo extrapolation method and system based on an enhanced recurrent neural network, the method comprising: determining the radar echo data of the target area, preprocessing the radar echo data; using the radar echo surface coverage screening strategy to screen the preprocessed radar echo data to obtain a high-quality radar echo data set; stacking 5 layers of dynamic convolutional spatiotemporal long and short-term memory modules and fusing gradient highway units to construct an enhanced prediction recurrent neural network; inputting the training samples in the radar echo data set into the enhanced prediction recurrent neural network for training and hyperparameter adjustment to obtain the best model; inputting the test samples in the radar echo data set into the best model for image prediction operations to obtain a radar echo extrapolated image. The present invention reduces the false alarm rate, improves the forecast accuracy, and performs excellently in the prediction of heavy precipitation, providing technical support for short-term precipitation forecasting.

Description

Technical Field

[0001] The present application belongs to the field of radar information processing technology, and specifically is a radar echo extrapolation method and system based on an enhanced recurrent neural network. Background Art

[0002] In recent years, affected by global climate change, disastrous weather has occurred frequently, and achieving more accurate, more precise and longer warning short-term precipitation forecasts has become a global concern. Short-term precipitation forecasts refer to the forecast of rainfall or severe convective weather within the next 0-2 hours. Since weather radar detection data has a high temporal and spatial resolution, radar echo extrapolation methods are mainly used in meteorological forecasting to achieve short-term precipitation forecasts.

[0003] In recent years, artificial intelligence technologies represented by deep learning have made breakthrough progress in the fields of image recognition and video, and have been widely used in the field of atmospheric science. In order to solve the problems existing in the above traditional methods, neural networks are introduced into radar echo extrapolation. With the excellent ability of mining massive data and nonlinear modeling, neural network models can fully learn the potential weather evolution laws in massive radar echo data, and have surpassed traditional methods in accuracy, and have achieved good results in short-term forecasting business. Radar extrapolation methods based on deep learning include convolutional long short-term memory networks, trajectory gated recurrent unit networks, predictive recurrent neural networks (PredRNN), and motion recurrent neural networks (MotionRNN).

[0004] The radar echo extrapolation method faces the problem of sample imbalance when dealing with heavy precipitation prediction, that is, heavy precipitation events account for a low proportion of the total samples, which makes it difficult for the model to fully learn the key features related to heavy precipitation, thus affecting the accuracy of precipitation prediction. Moreover, shallow networks are limited to low-level features, and deep networks are susceptible to gradient vanishing problems, which seriously restricts the model from learning complex features and limits the improvement of prediction accuracy. The ordinary convolution operation in the recurrent neural network model limits the model from extracting local features, resulting in insufficient dynamic feature capture capabilities, which makes it difficult to fully capture the complex weather dynamic change characteristics of local features in radar images when processing radar echo images with dynamic characteristics such as rotation, scaling, and generation and disappearance. Over time, the accuracy and clarity of precipitation forecasts will decrease, and there will usually be a problem of excessively high false alarm rates, which will significantly reduce the reliability of the radar system and waste resources. Therefore, it is urgent to overcome this problem through innovative algorithm design and optimization strategies to improve the accuracy and practicality of precipitation prediction. Summary of the invention

[0005] In view of the problems mentioned in the background technology, the purpose of the present invention is to provide a radar echo extrapolation method and system based on an enhanced recurrent neural network, which has the advantages of low false alarm rate and high precipitation forecasting ability.

[0006] An embodiment of the present application provides a radar echo extrapolation method based on an enhanced prediction recurrent neural network, the method comprising:

[0007] Determine the radar echo data of the target area and perform preprocessing;

[0008] The pre-processed radar echo data is screened according to the radar echo surface coverage screening strategy to obtain a high-quality radar echo data set;

[0009] Stack 5 layers of dynamic convolutional spatiotemporal long short-term memory modules and integrate gradient highway units to build an enhanced prediction recurrent neural network;

[0010] The training samples in the radar echo dataset are input into the enhanced prediction recurrent neural network for training and hyperparameter adjustment to obtain the best model;

[0011] The test samples in the radar echo data set are input into the optimal model to perform image prediction operations to obtain a radar echo extrapolated image.

[0012] Optionally, the determining radar echo data of the target area and preprocessing the data includes:

[0013] Traverse the radar echo data of the target area and limit the reflectivity factor of the radar echo data to 0-70dBZ;

[0014] Then normalization is performed, and the normalized radar echo data is scaled using the bicubic interpolation method.

[0015] Optionally, the preprocessed radar echo data is screened according to the radar echo surface coverage screening strategy to obtain a high-quality radar echo data set, including:

[0016] The radar echo data sequence is segmented using the sliding window method with a window size of 20;

[0017] The radar echo surface coverage screening strategy is to keep the echo surface coverage area of ​​at least one frame within the radar effective detection range greater than 10% in the first 10 frames of 20 consecutive frames of images, wherein the basic reflectivity of the radar echo surface coverage is greater than 15dBZ;

[0018] The radar echo data set is divided into a training set, a validation set and a test set.

[0019] Optionally, the stacking of 5 layers of dynamic convolutional spatiotemporal long short-term memory modules and the integration of gradient highway units to build an enhanced prediction recurrent neural network include:

[0020] An enhanced prediction recurrent neural network model for radar echo extrapolation is generated by stacking 5 layers of dynamic convolutional spatiotemporal long short-term memory modules and fusing gradient highway units between specified layers of the dynamic convolutional spatiotemporal long short-term memory modules.

[0021] Optionally, the designated inter-layer fusion gradient highway unit in the dynamic convolutional spatiotemporal long short-term memory module includes:

[0022] A fused gradient highway unit is introduced between the dynamic convolutional spatiotemporal long short-term memory modules of the first and second layers.

[0023] Optionally, the dynamic convolution spatiotemporal long short-term memory module is generated by deeply fusing the spatiotemporal long short-term memory unit with the dynamic convolution, and the convolution parameters are adjusted in real time to capture the instantaneous change characteristics of the radar echo;

[0024] The gradient highway unit is used to promote the learning of deep features of radar echoes and improve the adaptability of long-term and short-term features of radar echoes.

[0025] Optionally, after inputting the training samples in the radar echo data set into the enhanced prediction recurrent neural network for training and hyperparameter adjustment to obtain the best model, the method further includes:

[0026] The radar echo extrapolation method is implemented according to the enhanced prediction recurrent neural network, and the critical success index, hit rate and false alarm rate evaluation indicators are used to quantitatively evaluate the radar echo extrapolation results.

[0027] Another embodiment of the present application provides a radar echo extrapolation system based on an enhanced recurrent neural network, the system comprising:

[0028] An image processing module is used to determine the radar echo data of the target area and perform preprocessing;

[0029] The data set production module is used to filter the pre-processed radar echo data according to the radar echo surface coverage screening strategy to obtain a high-quality radar echo data set;

[0030] Model building module, used to stack 5 layers of dynamic convolutional spatiotemporal long short-term memory modules and integrate gradient highway units to build an enhanced prediction recurrent neural network;

[0031] The model training module is used to input the training samples in the radar echo data set into the enhanced prediction recurrent neural network for training and hyperparameter adjustment to obtain the best model;

[0032] The image prediction module is used to input the test samples in the radar echo data set into the optimal model to perform image prediction operations to obtain a radar echo extrapolated image.

[0033] Yet another embodiment of the present application provides a storage medium, wherein the storage medium stores a computer program, wherein the computer program is configured to execute the above-mentioned method when running.

[0034] In summary, the present invention mainly has the following beneficial effects:

[0035] (1) A high-quality radar echo dataset is constructed using a radar echo surface coverage screening strategy, which effectively reduces the impact of rainless weather on model training. This allows the model to focus more on learning key information related to heavy precipitation and solves the problem of excessive bias in heavy precipitation prediction results caused by sample imbalance, thereby improving the prediction accuracy of heavy precipitation events.

[0036] (2) A dynamic convolutional spatiotemporal long-short-term memory module was designed, and the model structure was deepened and an enhanced prediction recurrent neural network was built by integrating gradient highway units. This effectively improved the utilization rate of radar data, reduced information loss, and enhanced the ability to capture details of radar echoes. The model significantly improved the accuracy of short-term precipitation forecasts and reduced the false alarm rate. It can more accurately predict real severe convective weather events, avoid unnecessary emergency responses and waste of resources triggered by false warnings, and provide more reliable and efficient decision-making support for responding to extreme weather. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 A hardware structure block diagram of a computer terminal for a radar echo extrapolation method based on an enhanced recurrent neural network provided in an embodiment of the present application;

[0038] Figure 2 A schematic diagram of a flow chart of a radar echo extrapolation method based on an enhanced recurrent neural network provided in an embodiment of the present application;

[0039] Figure 3 A schematic diagram of the structure of the enhanced prediction recurrent neural network model provided in the embodiment of the present application;

[0040] Figure 4 A schematic diagram of the average values ​​of various evaluation indicators of the enhanced prediction recurrent neural network provided in an embodiment of the present application at different basic reflectivity thresholds;

[0041] Figure 5 A schematic diagram of radar echo extrapolation results of each model provided in the embodiments of the present application at different basic reflectivity thresholds;

[0042] Figure 6A schematic diagram of the change of the critical success index of each model under different basic reflectivity thresholds provided in the embodiment of the present application;

[0043] Figure 7 A schematic diagram of a flow chart of a radar echo extrapolation method based on an enhanced recurrent neural network provided in an embodiment of the present application;

[0044] Figure 8 A schematic diagram of the structure of a radar echo extrapolation system based on an enhanced predictive recurrent neural network provided in an embodiment of the present application. DETAILED DESCRIPTION

[0045] The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present application, and should not be construed as limiting the present application.

[0046] The embodiment of the present application provides a radar echo extrapolation method based on an enhanced recurrent neural network. The method can be applied to electronic devices, such as computer terminals, specifically ordinary computers, tablets, etc.

[0047] The following describes it in detail by taking running on a computer terminal as an example. Figure 1 The hardware structure block diagram of the computer terminal of the radar echo extrapolation method based on the enhanced recurrent neural network provided in the embodiment of the present application. Figure 1 As shown, the computer terminal may include one or more ( Figure 1 Only one is shown in the figure) a processor 102 (the processor 102 may include but is not limited to a processing device such as a microprocessor MCU or a programmable logic device FPGA) and a memory 104 for storing data. Optionally, the computer terminal may also include a transmission device 106 for communication functions and an input and output device 108. It can be understood by those skilled in the art that Figure 1 The structure shown is only for illustration and does not limit the structure of the above-mentioned computer terminal. Figure 1 More or fewer components as shown, or with Figure 1 Different configurations shown.

[0048] The memory 104 can be used to store software programs and modules of application software, such as the program instructions / modules corresponding to the radar echo extrapolation method based on the enhanced recurrent neural network in the embodiment of the present application. The processor 102 executes various functional applications and data processing by running the software programs and modules stored in the memory 104, that is, to implement the above method. The memory 104 may include a high-speed random access memory, and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 104 may further include a memory remotely arranged relative to the processor 102, and these remote memories may be connected to the computer terminal via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0049] The transmission device 106 is used to receive or send data via a network. The specific example of the above network may include a wireless network provided by a communication provider of a computer terminal. In one example, the transmission device 106 includes a network adapter (Network Interface Controller, NIC), which can be connected to other network devices through a base station so as to communicate with the Internet. In one example, the transmission device 106 can be a radio frequency (RF) module, which is used to communicate with the Internet wirelessly.

[0050] See also Figure 2 In order to solve the problems of insufficient data utilization, simple model structure and low forecast accuracy, this embodiment proposes an enhanced prediction recurrent neural network EN_PredRNN to predict the 6-minute precipitation in Province A in the next hour. Figure 2 The schematic flow chart of the radar echo extrapolation method based on the enhanced recurrent neural network provided in the embodiment of the present application may include the following steps:

[0051] S201: Determine the radar echo data of the target area and perform preprocessing.

[0052] Specifically, the preprocessing of the radar echo data includes:

[0053] Step 1: Traverse the radar echo data of the target area and limit the reflectivity factor of the radar echo data to 0-70dBZ;

[0054] Step 2: Then normalize the data and use the bicubic interpolation method to scale the normalized radar echo data to 256×256.

[0055] For example, the radar echo data of Province A is selected, with a time range from May to September 2016. First, the entire radar echo data is traversed, and all default values ​​and values ​​with basic reflectivity less than or equal to 0dBZ are uniformly replaced with 0. At the same time, for all values ​​greater than 70dBZ, they are set to 70dBZ to ensure that the data values ​​are strictly limited to 0-70dBZ. Then normalization is performed, that is, by dividing each data point by 70, the radar echo data is mapped to the interval of 0 to 1, which is conducive to improving the convergence speed of subsequent models. Finally, the normalized radar echo data is uniformly scaled to 256×256 using bicubic interpolation.

[0056] S202: Filter the preprocessed radar echo data according to the radar echo surface coverage screening strategy to obtain a high-quality radar echo data set.

[0057] Specifically, the preprocessed radar echo data is screened according to the radar echo surface coverage screening strategy to obtain a high-quality radar echo data set, including:

[0058] Step 1: Use the sliding window method to segment the radar echo data sequence, with a window size of 20;

[0059] Step 2: The radar echo surface coverage screening strategy is to keep the echo surface coverage area of ​​at least one frame within the radar effective detection range greater than 10% in the first 10 frames of 20 consecutive frames of images, wherein the basic reflectivity of the radar echo surface coverage is greater than 15dBZ;

[0060] Step 3: Divide the radar echo dataset into a training set, a validation set, and a test set.

[0061] Exemplarily, a sliding window is used to slide on the radar echo image sequence, the sliding window length is 20, and the step size is 1. Therefore, every 20 consecutive radar echo images are a sample, the first 10 radar echo images are used as the input of the model, and the last 10 radar echo images are used as labels, and there is no time overlap between samples. Continuing with the above example, assume that there are 1,836 samples in Province A from May 2016 to September 2016. To ensure independence between data, the data set can be divided into a training set, a validation set, and a test set. Among them, the samples from May to August are the training set, which is used to train subsequent models; the samples in early September are the validation set, which is used for model tuning; the samples in late September are the test set, which is used to evaluate the final performance of the model.

[0062] It should be noted that in the application of radar extrapolation technology, if the proportion of observation data of heavy precipitation events is low, it often leads to the problem of uneven sample distribution. This imbalance will significantly affect the training effect of the subsequent prediction model, which may lead to a decrease in the prediction accuracy of heavy precipitation events and affect the accuracy and reliability of the forecast.

[0063] Therefore, the present invention uses the radar echo surface coverage strategy to screen samples. Screening strategy: Since the reflectivity of precipitation echo is generally above 15dBZ, the first 10 frames of 20 consecutive frames should keep the echo surface coverage (basic reflectivity>15dBZ) of at least 1 frame within the effective detection range of the radar greater than 10%. After sample screening, the above radar echo data is divided into training set, validation set and test set with 551, 103 and 91 samples respectively.

[0064] This application uses radar echo surface coverage screening strategy to optimize the sample data set and alleviate the problem of sample imbalance. This method effectively reduces the impact of rainless weather on subsequent model training, allowing the model training process to focus more on learning key information related to heavy precipitation, thereby improving the model's prediction accuracy for heavy precipitation events.

[0065] S203: Stack 5 layers of dynamic convolutional spatiotemporal long short-term memory modules and integrate gradient highway units to build an enhanced prediction recurrent neural network.

[0066] Specifically, the stacking of 5 layers of dynamic convolutional spatiotemporal long short-term memory modules and the integration of gradient highway units to build an enhanced prediction recurrent neural network may include:

[0067] An enhanced prediction recurrent neural network for radar echo extrapolation is generated by designing and stacking 5 layers of dynamic convolutional spatiotemporal long short-term memory modules and fusing gradient highway units (GHU) between specified layers of the dynamic convolutional spatiotemporal long short-term memory modules.

[0068] Specifically, since radar echo images exhibit complex and changeable characteristics, including rotation, scaling, and generation and disappearance, traditional convolution operations are difficult to flexibly respond to the dynamic changes of local features of radar echo images. To this end, this application designs and builds a dynamic convolutional spatiotemporal long short-term memory module by integrating the adaptive characteristics of spatiotemporal long short-term memory units and dynamic convolution.

[0069] Among them, dynamic convolution can dynamically adjust the convolution kernel parameters according to the real-time changes of the input data, thereby significantly improving the model's expressiveness while keeping the network structure compact (without increasing the depth or width), so as to more flexibly capture the complex features of the radar echo image. The spatiotemporal long short-term memory unit can effectively process the time dependency and spatial correlation in time series data. Therefore, the dynamic convolution spatiotemporal long short-term memory module can significantly enhance the model's ability to extract the time series change characteristics of radar echo data and its adaptability and generalization ability to complex dynamic scene changes, and can more accurately capture the instantaneous change characteristics of radar echoes, providing solid technical support for the accurate prediction of short-term precipitation and heavy precipitation weather events.

[0070] In an optional embodiment, the designated inter-layer fusion gradient highway unit in the dynamic convolutional spatiotemporal long short-term memory module includes: introducing a fusion gradient highway unit between the first layer and the second layer of the dynamic convolutional spatiotemporal long short-term memory module.

[0071] The dynamic convolutional spatiotemporal long short-term memory module is able to effectively capture the spatiotemporal dynamic changes in time series data by combining the advantages of dynamic convolution (Dynamic Convolutional Network) and long short-term memory unit (Long Short-Term Memory, LSTM). In this embodiment, the dynamic convolutional spatiotemporal long short-term memory module is mainly used to capture the instantaneous change characteristics of radar echoes, and expand the model receptive field by deepening the network structure to extract the deep features of radar echoes. The gradient highway unit is used to promote the learning of deep features of radar echoes to improve the adaptability of long-term and short-term features of radar echoes.

[0072] Specifically, the present application first constructs an Enhanced Predictive Recurrent Neural Network (EN_PredRNN) for radar echo extrapolation. The radar data is preprocessed and samples are screened through the above steps S201 and S202 to construct a high-quality radar echo data set; secondly, based on the key architecture of PredRNN, by deeply integrating the spatio-temporal long short-term memory unit (Spatio-Temporal Long Short-Term Memory, ST_LSTM) and dynamic convolution (Dynamic Convolution), a dynamic convolution spatio-temporal long short-term memory module (Dynamic Convolution combined with Spatio-Temporal Long Short-Term Memory, DC_STLSTM) is designed and constructed, and the convolution parameters are adjusted in real time to accurately capture the instantaneous change characteristics of the radar echo, significantly enhancing the model's ability to capture complex dynamic features. Then, by deepening the network structure and expanding the model's receptive field, deeper features of the radar echo are extracted to enhance the accuracy of precipitation prediction and capture details. In order to alleviate the gradient vanishing problem, a gradient highway unit is introduced between the first two layers of the model to adaptively learn the short-term and long-term dynamic change information in the radar echo data, thereby improving the generalization ability and prediction accuracy of the model.

[0073] This model not only significantly reduces the false alarm rate, but also greatly improves the accuracy of precipitation forecasts, especially in the prediction of heavy precipitation events, and can provide strong technical support for precipitation forecasts in target areas.

[0074] It should be noted that by deepening the network structure, expanding the model receptive field, and extracting higher-order features of radar echoes, the accuracy of precipitation prediction and the capture of details can be enhanced. To address the gradient vanishing problem, the gradient highway unit is introduced to promote the learning of deep features, making the model more adaptable in capturing the long-term and short-term features of radar echoes, thereby improving the accuracy and reliability of short-term precipitation forecasts.

[0075] Based on the key architecture of PredRNN, by deeply integrating spatiotemporal long short-term memory units and dynamic convolution, a dynamic convolutional spatiotemporal long short-term memory module is designed and constructed, and the convolution parameters are adjusted in real time to accurately capture the instantaneous change characteristics of radar echoes; by introducing gradient highway units between the first two layers of the dynamic convolutional spatiotemporal long short-term memory module, the short-term and long-term dynamic change information in the radar echo data is adaptively learned, which can improve the generalization ability and prediction accuracy of the model.

[0076] Among them, since shallow networks are limited to low-level features, deeper network structures usually have stronger learning and expression capabilities, and are more suitable for processing more complex data and tasks. Therefore, 5 layers of dynamic convolutional spatiotemporal long short-term memory modules are stacked in the enhanced recurrent neural network constructed in this application to build a deep network architecture. This design aims to significantly improve the network's ability to learn short-term dynamics by increasing the number of cyclic states from input to output. However, deep networks are susceptible to gradient vanishing, and it is difficult for the model to effectively learn complex features. Therefore, a gradient highway unit is introduced between the first and second layers of the model.

[0077] The gradient highway unit provides a "highway" for the gradient during the back-propagation process, allowing the gradient to propagate to the front end of the network more quickly and stably. This not only helps to reduce the attenuation of the gradient in the deep network, but also significantly improves the training efficiency, allowing the model to converge to a better solution more efficiently. The introduction of the gradient highway unit promotes the learning of deep features of radar echoes, making the model more adaptable in capturing the long-term and short-term features of radar echoes. Therefore, while enhancing the performance of the model, the gradient highway unit also improves the stability and reliability of training and reduces the risk of overfitting.

[0078] See also Figure 3 , Figure 3 A schematic diagram of the structure of an enhanced recurrent neural network provided in an embodiment of the present application. In the figure, five layers of dynamic convolutional spatiotemporal long short-term memory modules are stacked, and a gradient highway unit is introduced between the first and second layers of the model. The expression of the gradient highway unit can be:

[0079] ;

[0080] in, Indicates time, yes Time input, yes The hidden state of the moment, yes The hidden state of the moment, It is a gradient highway unit.

[0081] The specific mathematical expression of the enhanced prediction recurrent neural network model is as follows:

[0082] ;

[0083] ;

[0084] ;

[0085] ;

[0086] ;

[0087] ;

[0088] Where t represents time, represents the gradient highway unit, represents the hidden state at time t, represents the hidden state at time t-1, is the input at time t;

[0089] They represent the hidden state, cell state, and spatiotemporal memory flow of the first layer at time t respectively;

[0090] They represent the hidden state, cell state, and spatiotemporal memory flow of the second layer at time t respectively;

[0091] They represent the hidden state, cell state, and spatiotemporal memory flow of the third layer at time t respectively;

[0092] They represent the hidden state, cell state, and spatiotemporal memory flow of the fourth layer at time t respectively;

[0093] They represent the hidden state, cell state, and spatiotemporal memory flow of the fifth layer at time t respectively;

[0094] Respectively represent the hidden state and cell state of the first layer at time t-1;

[0095] They represent the hidden state and cell state of the second layer at time t-1 respectively;

[0096] They represent the hidden state and cell state of the third layer at time t-1 respectively;

[0097] They represent the hidden state and cell state of the fourth layer at time t-1 respectively;

[0098] They represent the hidden state and cell state of the fifth layer at time t-1 respectively;

[0099] Represents the spatiotemporal memory flow of the fifth layer at time t-1;

[0100] represents the first layer of dynamic convolutional spatiotemporal long short-term memory module, where It represents the number of layers of the model. According to the same method and principle, the expressions of the dynamic convolutional spatiotemporal long short-term memory modules in other layers are deduced accordingly and will not be repeated here.

[0101] S204: Input the training samples in the radar echo data set into the enhanced prediction recurrent neural network for training and hyperparameter adjustment to obtain the best model.

[0102] S205: Input the test samples in the radar echo data set into the optimal model to perform image prediction operations to obtain a radar echo extrapolated image.

[0103] Specifically, by executing training and hyperparameter adjustment of the enhanced prediction recurrent neural network model, a trained enhanced recurrent neural network model suitable for radar echo extrapolation is obtained. The model can use radar echo data to accurately predict future precipitation conditions.

[0104] It should be noted that in the training process of the enhanced prediction recurrent neural network model, this application uses a learning rate of 0.001, an optimizer of Adam, a batch size of 4, an iteration number of 10,000 times, and a loss function of mean absolute error + root mean square error. From the experimental results, it can be seen that when five layers of dynamic convolutional spatiotemporal long and short-term memory modules are stacked, the enhanced prediction recurrent neural network model has the best effect, and the enhanced recurrent prediction neural network model can be determined.

[0105] After inputting the training samples in the radar echo data set into the enhanced prediction recurrent neural network for training and hyperparameter adjustment to obtain the best model, the method further includes:

[0106] The radar echo extrapolation method is implemented according to the enhanced prediction recurrent neural network, and the critical success index, hit rate and false alarm rate evaluation indicators are used to quantitatively evaluate the radar echo extrapolation results.

[0107] For example, the basic reflectivity of radar data, 25 dBZ, 35 dBZ and 45 dBZ, can be selected as the corresponding thresholds, and the critical success index, false alarm rate and hit rate can be used to quantitatively evaluate the radar echo test set. Among them, the critical success index, hit rate and false alarm rate indicators are used to quantitatively evaluate the radar echo extrapolation results, and the specific calculation formula is as follows:

[0108] ;

[0109] ;

[0110] ;

[0111] in, is the critical success index, is the hit rate, is the false alarm rate, hits represents the hit point, where the predicted value = 1 and the true observation value = 1; misses represents the missed alarm point, where the predicted value = 0 and the true observation value = 1; false alarms represents the empty alarm point, where the predicted value = 1 and the true observation value = 0.

[0112] It should be noted that the higher the hit rate and critical success index, the more accurate the radar echo prediction; the lower the false alarm rate, the more accurate the model; and the larger the basic reflectivity threshold, the higher the probability of heavy precipitation.

[0113] It should be noted that in order to comprehensively evaluate the comprehensive performance of the enhanced prediction recurrent neural network model En_PredRNN, the prediction recurrent neural network model PredRNN and the motion recurrent neural network model MotionRNN under different precipitation intensities, the application selects basic reflectivity 25dBZ, 35dBZ and 45dBZ as corresponding thresholds under the extrapolation time of 0-1h, and uses the critical success index, false alarm rate and hit rate to quantitatively evaluate the radar echo test set. All models are trained, verified and tested under the same data set and environment. The quantitative evaluation results are shown in Table 1:

[0114] Table 1 Evaluation results of each model at different basic reflectivity thresholds

[0115]

[0116] Combined with Table 1 and see Figure 4 , Figure 4 Mean value schematic diagram of each evaluation index under different basic reflectivity thresholds provided for the embodiment of the present application. Under the same forecast timeliness, the critical success index and false alarm rate of the En_PredRNN model at three reflectivity thresholds are optimal. Under the thresholds of 25dBz, 35dBz, and 45dBz, compared with the benchmark model PredRNN, the critical success index of the En_PredRNN model increased by 19.3%, 17.3%, and 16.5%, respectively, and the false alarm rate decreased by 28.3%, 27.5%, and 26.7%, respectively. The En_PredRNN model is 3.6%, 4.2%, and 3.2% higher than MotionRNN in the critical success index, and the false alarm rate is reduced by 8.3%, 8.5%, and 8.1%, respectively. Therefore, the En_PredRNN model significantly improves the critical success index, reduces the false alarm rate, and effectively improves the accuracy of precipitation forecast.

[0117] See also Figure 5 , Figure 5 Schematic diagram of radar echo extrapolation results of various models provided in the embodiments of the present application at different basic reflectivity thresholds, Figure 5It can be seen that all models made accurate predictions at the first moment. As time goes by, obvious differences emerge between the models. The PredRNN model tends to exaggerate the predicted range and intensity of precipitation, resulting in obvious false alarms, and cannot correctly estimate the intensity distribution inside the echo. The MotionRNN model cannot better capture the long-term state of the radar and tends to underestimate the intensity of convective clouds. The radar echo intensity and position of the En_PredRNN model are relatively consistent with the corresponding radar observations, and can accurately predict the movement trend and distribution range of heavy precipitation in severe convective weather. The En_PredRRNN model significantly enhances the feature extraction capability of the spatiotemporal information of radar echoes, effectively captures the long-term dependencies and important information in radar data, learns the temporal and spatial characteristics of the evolution of severe convection, and accurately predicts the movement position and shape changes of radar echoes.

[0118] To more intuitively observe the forecast time performance of different models, see Figure 6 , Figure 6 This is a schematic diagram of the changes in the critical success index of each model under different basic reflectivity thresholds provided in the embodiment of the present application. The critical success index of different models extrapolated within 1 hour under the basic reflectivity thresholds of 25dBZ, 35dBZ, and 45dBZ is plotted, and the time step interval is 6 minutes. Figure 6 It can be seen that compared with other classic models, the En_PredRNN model has the highest critical success index under the three basic reflectivity thresholds, and with the extension of the forecast time, the critical success index of all models shows a downward trend, and the forecast ability gradually weakens.

[0119] See also Figure 7 , Figure 7 A flow chart of the radar echo extrapolation method based on an enhanced recurrent neural network provided in the embodiment of the present application. First, the radar echo data of the target area is determined, the radar echo data is preprocessed, and the sample is screened using the radar echo surface coverage strategy to obtain a high-quality radar echo sample set; then an enhanced prediction recurrent neural network model for radar echo extrapolation is constructed; based on the high-quality radar echo sample set, the training and hyperparameter adjustment of the enhanced prediction recurrent neural network model are performed, and the model accuracy index is evaluated, and finally, the prediction accuracy of each model is compared by comparing the experimental results of each model, providing strong technical support for precipitation forecasting.

[0120] Another embodiment of the present application provides a radar echo extrapolation system based on an enhanced recurrent neural network, such as Figure 8 The structure diagram of the radar echo extrapolation system based on the enhanced recurrent neural network shown in FIG. 1 includes:

[0121] Image processing module 801, used to determine radar echo data of the target area and perform preprocessing;

[0122] The data set preparation module 802 is used to filter the pre-processed radar echo data according to the radar echo surface coverage screening strategy to obtain a high-quality radar echo data set;

[0123] Model building module 803, used to stack 5 layers of dynamic convolutional spatiotemporal long short-term memory modules and fuse gradient highway units to build an enhanced prediction recurrent neural network;

[0124] The model training module 804 is used to input the training samples in the radar echo data set into the enhanced prediction recurrent neural network for training and hyperparameter adjustment to obtain the best model;

[0125] The image prediction module 805 is used to input the test samples in the radar echo data set into the optimal model to perform image prediction operations to obtain a radar echo extrapolated image.

[0126] An embodiment of the present application further provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute the steps in any one of the above method embodiments.

[0127] Specifically, the electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the processor, and the input / output device is connected to the processor.

[0128] Specifically, in this embodiment, the processor may be configured to perform the following steps through a computer program:

[0129] S201: Determine radar echo data of the target area and perform preprocessing;

[0130] S202: Filter the preprocessed radar echo data according to the radar echo surface coverage screening strategy to obtain a high-quality radar echo data set;

[0131] S203: Stack 5 layers of dynamic convolutional spatiotemporal long short-term memory modules and integrate gradient highway units to build an enhanced prediction recurrent neural network;

[0132] S204: inputting the training samples in the radar echo data set into the enhanced prediction recurrent neural network for training and hyperparameter adjustment to obtain the best model;

[0133] S205: Input the test samples in the radar echo data set into the optimal model to perform image prediction operations to obtain a radar echo extrapolated image.

[0134] Compared with the prior art, the present application first determines the radar echo data of the target area and pre-processes the radar echo data; then, the pre-processed radar echo data is screened according to the radar echo surface coverage screening strategy to obtain a high-quality radar echo data set; then, 5 layers of dynamic convolutional spatiotemporal long and short-term memory modules are stacked and the gradient highway unit is integrated to construct an enhanced prediction recurrent neural network model; then, the training samples in the radar echo data set are input into the enhanced prediction recurrent neural network model for training and hyperparameter adjustment to obtain the best model; finally, the test samples in the radar echo data set are input into the best model for image prediction operation to obtain a radar echo extrapolated image. The present invention can not only significantly reduce the false alarm rate, but also greatly improve the accuracy of precipitation forecasting, especially showing good performance in the prediction of heavy precipitation events, and providing technical support for precipitation forecasting.

[0135] An embodiment of the present application further provides a storage medium, in which a computer program is stored, wherein the computer program is configured to execute the steps of any of the above method embodiments when running.

[0136] Specifically, in this embodiment, the above storage medium may be configured to store a computer program for performing the following steps:

[0137] S201: Determine radar echo data of the target area and perform preprocessing;

[0138] S202: Filter the preprocessed radar echo data according to the radar echo surface coverage screening strategy to obtain a high-quality radar echo data set;

[0139] S203: Stack 5 layers of dynamic convolutional spatiotemporal long short-term memory modules and integrate gradient highway units to build an enhanced prediction recurrent neural network;

[0140] S204: inputting the training samples in the radar echo data set into the enhanced prediction recurrent neural network for training and hyperparameter adjustment to obtain the best model;

[0141] S205: Input the test samples in the radar echo data set into the optimal model to perform image prediction operations to obtain a radar echo extrapolated image.

[0142] Specifically, in this embodiment, the above-mentioned storage medium may include but is not limited to: a USB flash drive, a read-only memory (ROM), a random access memory (RAM), a mobile hard disk, a magnetic disk or an optical disk, and other media that can store computer programs.

[0143] Compared with the prior art, the present application first determines the radar echo data of the target area and pre-processes the radar echo data; then, the pre-processed radar echo data is screened according to the radar echo surface coverage screening strategy to obtain a high-quality radar echo data set; then, 5 layers of dynamic convolutional spatiotemporal long and short-term memory modules are stacked and gradient highway units are integrated to construct an enhanced prediction recurrent neural network; then, the training samples in the radar echo data set are input into the enhanced prediction recurrent neural network for training and hyperparameter adjustment to obtain the best model; finally, the test samples in the radar echo data set are input into the best model for image prediction operation to obtain a radar echo extrapolated image. The present invention can not only significantly reduce the false alarm rate, but also greatly improve the accuracy of precipitation forecasting, especially showing good performance in the prediction of heavy precipitation events, and providing technical support for precipitation forecasting.

[0144] It should be noted that, for the aforementioned method embodiments, for the sake of simplicity, they are all expressed as a series of action combinations, but those skilled in the art should be aware that the present application is not limited by the described order of actions, because according to the present application, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the present application.

[0145] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0146] In the several embodiments provided in the present application, it should be understood that the disclosed device can be implemented in other ways. For example, the device embodiments described above are only schematic, such as the division of the above-mentioned units, which is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, and the indirect coupling or communication connection of devices or units can be electrical or other forms.

[0147] The units described above as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0148] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0149] If the above-mentioned integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable memory. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a memory, including a number of instructions to enable a computer device (which can be a personal computer, server or network device, etc.) to execute all or part of the steps of the above-mentioned methods in each embodiment of the present application. The aforementioned memory includes: U disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), mobile hard disk, disk or optical disk and other media that can store program codes.

[0150] The embodiments of the present application are introduced in detail above. Specific examples are used in this article to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method and core idea of ​​the present application. At the same time, for general technical personnel in this field, according to the idea of ​​the present application, there will be changes in the specific implementation method and application scope. In summary, the content of this specification should not be understood as a limitation on the present application.

Claims

1. A radar echo extrapolation method based on an enhanced recurrent neural network, characterized in that: The method comprises: Determine the radar echo data of the target area and perform preprocessing; The pre-processed radar echo data is screened according to the radar echo surface coverage screening strategy to obtain a high-quality radar echo data set; Five layers of dynamic convolutional spatiotemporal long short-term memory modules are stacked, and a fused gradient highway unit is introduced between the first and second layers of dynamic convolutional spatiotemporal long short-term memory modules to generate an enhanced prediction recurrent neural network model for radar echo extrapolation. The training samples in the radar echo dataset are input into the enhanced prediction recurrent neural network for training and hyperparameter adjustment to obtain the best model; The test samples in the radar echo data set are input into the optimal model to perform image prediction operations to obtain a radar echo extrapolated image.

2. The method according to claim 1, characterized in that The step of determining the radar echo data of the target area and preprocessing the data includes: Traverse the radar echo data of the target area and limit the reflectivity factor of the radar echo data to 0-70dBZ; Then normalization is performed, and the normalized radar echo data is scaled using the bicubic interpolation method.

3. The method according to claim 1, characterized in that The method of screening the pre-processed radar echo data according to the radar echo surface coverage screening strategy to obtain a high-quality radar echo data set includes: The radar echo data sequence is segmented using the sliding window method with a window size of 20; The radar echo surface coverage screening strategy is to keep the echo surface coverage area of ​​at least one frame within the radar effective detection range greater than 10% in the first 10 frames of 20 consecutive frames of images, wherein the basic reflectivity of the radar echo surface coverage is greater than 15dBZ; The radar echo data set is divided into a training set, a validation set and a test set.

4. The method according to claim 1, characterized in that: The dynamic convolution spatiotemporal long short-term memory module is generated by deeply fusing the spatiotemporal long short-term memory unit and the dynamic convolution, and the convolution parameters are adjusted in real time to capture the instantaneous change characteristics of the radar echo; The gradient highway unit is used to promote the learning of deep features of radar echoes and improve the adaptability of long-term and short-term features of radar echoes.

5. The method according to any one of claims 1 to 4, characterized in that: After inputting the training samples in the radar echo data set into the enhanced prediction recurrent neural network for training and hyperparameter adjustment to obtain the best model, the method further includes: The radar echo extrapolation method is implemented according to the enhanced prediction recurrent neural network, and the critical success index, hit rate and false alarm rate evaluation indicators are used to quantitatively evaluate the radar echo extrapolation results.

6. A radar echo extrapolation system based on an enhanced recurrent neural network, characterized in that: The system comprises: An image processing module is used to determine the radar echo data of the target area and perform preprocessing; The data set production module is used to filter the pre-processed radar echo data according to the radar echo surface coverage screening strategy to obtain a high-quality radar echo data set; A model building module is used to stack 5 layers of dynamic convolutional spatiotemporal long short-term memory modules and introduce a fused gradient highway unit between the first and second layers of dynamic convolutional spatiotemporal long short-term memory modules to generate an enhanced prediction recurrent neural network model for radar echo extrapolation; The model training module is used to input the training samples in the radar echo data set into the enhanced prediction recurrent neural network for training and hyperparameter adjustment to obtain the best model; The image prediction module is used to input the test samples in the radar echo data set into the optimal model to perform image prediction operations to obtain a radar echo extrapolated image.

Images