Memory cycle perception short-time heavy rainfall prediction method for flood disaster early warning

Abstract

The invention discloses a flood disaster early warning-oriented memory cycle perception short-time heavy rainfall prediction method, and belongs to the technical field of meteorological disaster prevention and control and weather radar echo extrapolation. The method comprises the following steps: acquiring continuous radar echo image data, and constructing an MRP-LSTM cycle unit by fusing a memory cycle sensing module and a context feature storage module to obtain global features of radar echo images; an efficient MRP-Net model is constructed by stacking MRP-LSTM cycle units, after training, the time sequence features and the context relation in meteorological data can be accurately captured, the short-time heavy rainfall prediction efficiency is greatly improved, and a solid guarantee is provided for timeliness and reliability of flood disaster early warning.

Description

Technical Field

[0001] The present invention relates to the technical field of meteorological disaster prevention and control, and particularly relates to a memory cycle perception short-term heavy precipitation prediction method for flood disaster warning. Background Art

[0002] Meteorological disasters occur frequently and wantonly, and flood disasters are particularly common and have far-reaching harms. As a key inducement of them, short-term heavy precipitation has always been the core difficulty focused on in the field of meteorological forecasting. The strong convective characteristics endow short-term heavy precipitation with traits such as strong suddenness, rapid disaster-causing, and a short predictable window. Given the close causal relationship between short-term heavy precipitation and flood disasters, accurately predicting short-term heavy precipitation is the key prerequisite for preventing flood disasters, and short-term and nowcasting precipitation forecasting technology has thus emerged. Short-term and nowcasting precipitation forecasting focuses on the high-resolution prediction of local precipitation within the next 0-2 hours. Compared with traditional short-term precipitation forecasting, it has significant advantages in grasping precipitation intensity and spatial distribution, provides indispensable key data support for flood disaster warning, and effectively reduces potential hazards.

[0003] The rapid development of meteorological radar technology has made the extrapolation method based on radar echo data the mainstream means of current short-term and nowcasting precipitation forecasting. This method realizes precipitation forecasting by deeply mining the internal laws of historical radar echo map sequences, constructing a model to predict future echo sequences. Radar echo extrapolation models are mainly divided into two categories: traditional numerical calculation models, such as cross-correlation tracking method, centroid tracking method, and optical flow method, etc. Although they have a certain foundation, in the prediction of the start and end of heavy rainstorms, due to the disconnection between the pre-set assumptions and the actual weather conditions, the accuracy often goes astray, seriously affecting the judgment of the degree, scope, and development trend of flood disasters. Deep learning models, with their excellent achievements in multiple fields, provide innovative ideas for solving this problem.

[0004] In the application of radar echo extrapolation, deep learning networks include Convolutional Neural Network (CNN), Recurrent Neural Network (RNN), and Hybrid Neural Network (HNN). Classic CNN models such as U-Net and its variants MSLKNet, TempNet, and NLED; representative RNN models include ConvLSTM and its derivatives PredRNN, PredRNN++, MIM, SA-ConvLSTM; as a cutting-edge trend, HNN integrates multiple neural networks to improve spatio-temporal prediction accuracy, and Generative Adversarial Networks (GAN) have been widely used in it. These models can independently summarize the change patterns of historical radar echo data and achieve more effective future echo prediction.

[0005] However, the existing radar echo extrapolation method still has obvious shortcomings in flood disaster warning applications. On the one hand, the model is not universal enough. The climate and terrain in different geographical environments vary greatly. For example, the precipitation convergence patterns in mountainous areas and plains are completely different, and coastal areas are susceptible to extreme weather such as typhoons. It is difficult for existing models to fully take into account these complex factors and cannot accurately adapt to the warning needs in various regions and meteorological conditions, and the application efficiency is limited. On the other hand, the analysis of the radar echo motion trajectory is fuzzy, which makes it difficult to accurately track the dynamic changes in the spatiotemporal distribution of precipitation when predicting short-term heavy precipitation, and the prediction accuracy is greatly reduced.

[0006] Therefore, optimizing the performance of the radar echo extrapolation model and enhancing its adaptability in complex actual scenarios have become key issues that need to be tackled in the field of meteorological disaster warning. Summary of the invention

[0007] In view of the inherent defects of the existing radar echo extrapolation model, the present invention provides a memory cycle perception short-term heavy rainfall prediction method for flood disaster warning, which improves the accuracy and timeliness of flood disaster warning by optimizing the performance of the radar echo extrapolation model.

[0008] The present invention adopts the following technical solution: a memory cycle perception short-term heavy rainfall prediction method for flood disaster warning, which generates a future precipitation distribution map through the MRP-Net framework, specifically including the following steps: Step 1: Collect continuous radar echo image data to monitor the dynamic changes of precipitation in potential flood disaster areas in real time; Step 2, constructing an MRP-LSTM recurrent unit, including: a memory recurrent perception module, an additional memory component, and a context feature fusion module; the MRP-LSTM recurrent unit receives continuous radar echo images, extracts local features, captures the time correlation characteristics between adjacent radar echo images through the memory recurrent perception module, and strengthens the memory of radar echo motion features; through the context feature fusion module, context information at different levels is integrated with the input data state to obtain the global features of the radar echo image; Step 3: A multi-layer stacked MRP-LSTM recurrent unit strategy is used to form an MRP-Net framework, build a flood disaster warning model and perform model training, capture the time series characteristics and contextual relationships in meteorological data, output future high-resolution precipitation forecast results, and trigger flood disaster warning signals based on preset thresholds. The warning signals include risk level, impact range, and recommended emergency measures, providing key decision-making basis for the flood disaster warning system.

[0009] Preferably, in step 2, the MRP-LSTM recurrent unit is a double-layer structure, including an upper branch and a lower branch; The upper branch receives the input state of continuous radar echo images , cell state and hidden state , and transmits them to the encoder for encoding to extract the appearance features of the input image; using the original spatio-temporal LSTM unit for cyclic processing to generate the local historical dependence features of the input sequence , and input them into the context feature fusion module; The original spatio-temporal LSTM unit is used for the dynamic association of information, fusing the temporal and spatial information of the current moment with that of the previous moment and the next moment of information; The lower branch takes the difference sequence of the input sequence of the upper branch as its input sequence , and the difference sequence captures the continuous change information of the current action and imports it into the memory recurrent perception module. By means of the memory vector , it performs a query to obtain the context memory features corresponding to the current difference sequence , embeds them into the context feature fusion module, and performs deep fusion with the local historical dependence features extracted by the upper branch to generate global features for identifying the spatio-temporal evolution law of precipitation in high-risk flood disaster areas.

[0010] Preferably, the memory recurrent perception module includes: a context perception feature storage link and a motion feature recall link, which are used to capture the temporal association characteristics between adjacent radar echo images and strengthen the memory of the echo motion features; In the context perception feature storage link, a context perception encoder is used to extract the perception features of the difference sequence . During the extraction of the appearance features of the difference sequence, a 1×1 convolutional filter is introduced to perform a linear transformation operation on the input image and decompose it into local perception features through channel refinement ; the local perception features are separately stored in the memory component for memory query; In the motion feature recall link, a context motion encoder is used to extract the motion features , and decompose them into local features of the motion features to form a memory query vector for recalling the corresponding echo perception features from the memory component and updating the memory component ​The long-term motion memory vector in it extracts corresponding long-term motion context information from the memory and transports it to the context feature fusion module; the long-term motion context information is used to enhance the prediction accuracy of the critical period for flood disaster triggering.

[0011] Preferably, the context feature fusion module includes: a historical dependence information extraction link and a context memory feature recall link; In the historical dependence information extraction link, the original spatio-temporal LSTM unit extracts historical dependence information, and obtains multi-level context information by adding the hidden state and , and performs deep fusion with the initial input state ; uses a convolutional filter to perform context feature fusion operations on and , and obtains by element-wise multiplication of and , and uses the function to normalize the values in the feature map between 0 and 1; repeat the execution until the updated input state and the hidden state are obtained; In the context memory feature recall link, the context memory feature is recalled through the memory loop perception module, and a 1×1 convolutional filter is introduced to adjust the number of channels to obtain the initial input state of the difference sequence. An addition operation is performed on the hidden state and to obtain multi-level context information , and is deeply fused with ; a convolutional filter is used to perform context feature fusion operations on and , and is element-wise multiplied by to obtain , and the function is used to normalize the values in the feature map between 0 and 1; repeat the execution until the updated input state and the hidden state are obtained.

[0012] Preferably, for the deep fusion, a weighting strategy is used to balance the historical memory and the real-time dynamic features, ensuring that the generation of flood disaster warning signals takes into account both timeliness and accuracy; In the context feature fusion module, the extracted in the historical dependence information extraction link and the extracted in the context memory feature recall linkPerform mapping and fusion, and generate context fusion features after calculating the Hadamard product .

[0013] The MRP-LSTM recurrent unit inputs the context fusion features into the decoder . According to the transposed convolution operation of the decoder , the radar echo sequence for the future time steps is obtained , which is used for short-term heavy precipitation prediction before flood disasters occur.

[0014] Preferably, in step 3, the MRP-Net framework adopts a multi-layer fusion strategy to stack a number of MRP-LSTM units to form a multi-layer recurrent structure. Each column and each row include multiple MRP-LSTM recurrent units. Each column represents a recurrent unit for a single time step, and each row represents the continuous update of the current-level memory unit and state over time; In the multi-layer recurrent structure, the hidden state is continuously transmitted to the next unit along the horizontal and vertical directions, the memory unit is transmitted to the next unit along the horizontal direction, and the output of the memory perception module is transmitted to the next vertical unit along the vertical direction; through cross-time-step context information sharing, the context information at different levels is deeply integrated with the input state, improving the model's ability to continuously track long-term precipitation dynamics.

[0015] Preferably, in the first time step of the MRP-Net framework, for the bottom-layer MRP-LSTM recurrent unit, all memory units and hidden states are initially set to a tensor form with all elements being zero, which is used to ensure that the initial state of the flood disaster warning model is synchronized and updated with real-time data; For the top-layer MRP-LSTM recurrent unit, the information is transmitted to the bottom-layer unit in the next time step; for the non-topmost MRP-LSTM recurrent unit, the hidden state is transmitted to higher-level units at subsequent moments, enabling the flood disaster warning model to dynamically adapt to the mutation characteristics of the spatio-temporal distribution of precipitation in flood disaster warnings, avoiding forgetting echo features, and predicting future echo images.

[0016] The technical solution of the present invention also provides: An electronic device, including: One or more processors; A storage device, on which one or more programs are stored; When the one or more programs are executed by the one or more processors, the one or more processors implement the memory cycle perception short-term heavy precipitation prediction method for flood disaster warning described in any of the above.

[0017] The technical solution of the present invention also provides a computer-readable storage medium, on which a computer program is stored. When the program is executed by a processor, the steps in any of the above memory cycle perception short-term heavy precipitation prediction methods for flood disaster warning are implemented.

[0018] Compared with the prior art, the present invention adopts the above technical solution and has the following technical effects: 1. In the memory cycle perception short-term heavy precipitation prediction method of the present invention, the memory cycle perception module, relying on its unique context perception feature storage mechanism and motion feature recall process, can keenly capture the temporal correlation characteristics between adjacent radar echo maps, significantly enhancing the memory ability of the echo motion features and accurately delivering more valuable context information to the model; the context feature fusion module, on the other hand, uses a sophisticated multi-layer fusion strategy to deeply integrate context information at different levels with the input state, greatly enhancing the model's control ability over long-term dependence features and effectively avoiding the error accumulation problem caused by information forgetting during long-term prediction.

[0019] 2. The memory cycle perception short-term heavy precipitation prediction method of the present invention is an extension based on the traditional ST-LSTM model, innovatively integrating two unique structures, namely, the memory cycle perception module and the context feature fusion module. These two modules complement each other and operate synergistically, endowing the MRP-Net model with excellent performance. When dealing with complex and variable radar echo data, the MRP-Net model demonstrates highly stable extrapolation performance, and the prediction accuracy is significantly improved. Description of the Drawings

[0020] Figure 1 It is the architecture diagram of the MRP-LSTM loop unit of the present invention; Figure 2 It is the module structure diagram of the memory cycle perception module of the present invention; Figure 3 It is the structure diagram of the context feature fusion module of the present invention; Figure 4 It is the network structure MRP-Net composed of 4 layers of MRP-LSTM loop units stacked in the embodiment of the present invention; Figure 5 It is the bar chart of the CSI value comparison of the prediction results of each model in the embodiment of the present invention; Figure 6 It is the bar chart of the HSS value comparison of the prediction results of each model in the embodiment of the present invention; Figure 7This is a comparison and discount graph of the MSE and SSIM of the prediction results of each model in the embodiments of the present invention step by step over time. Detailed implementation manners

[0021] In order to make the objectives, technical solutions and advantages of the present invention clearer, the technical solutions of the application will be further elaborated in detail below with reference to the accompanying drawings. The described embodiments are only a part of the embodiments involved in the present invention. All non-innovative embodiments made by other researchers in the field based on this embodiment fall within the protection scope of the present invention. At the same time, for the step numbers in the embodiments of the present invention, they are only set for the convenience of elaboration and explanation, and no limitation is imposed on the order between the steps. The execution order of each step in the embodiments can be adaptively adjusted according to the understanding of those skilled in the art.

[0022] In view of the inherent defects of the existing radar echo extrapolation model, the present invention proposes a memory recurrent perception short-term heavy precipitation prediction method for flood disaster warning, and innovatively designs a unique recurrent unit - the MRP-LSTM recurrent unit. To enhance the efficacy of the prediction model, the MRP-LSTM recurrent unit integrates a memory recurrent perception module, an additional memory component, and a context feature fusion module, and constructs the MRP-Net framework by adopting the strategy of stacking multiple layers of MRP-LSTM units. It can be widely applied to flood disaster prevention and control scenarios such as urban waterlogging warning, reservoir operation, and basin-wide flood monitoring, significantly improving the accuracy and timeliness of early warning response.

[0023] In an embodiment of the present invention, a memory recurrent perception short-term heavy precipitation prediction method for flood disaster warning generates a future precipitation distribution map through the MRP-Net model, and triggers a flood disaster warning signal based on a preset threshold. The warning signal includes a risk level, an influence range, and recommended emergency measures, and specifically includes the following steps: Step 1: Collect continuous radar echo image data from a Doppler radar of model CINRAD / SA in a certain area of Guangzhou Station. The data is used to monitor the dynamic changes of precipitation in potential flood disaster areas in real time; Step 2: Construct an MRP-LSTM recurrent unit, including: a memory recurrent perception module, an additional memory component, and a context feature fusion module; the MRP-LSTM recurrent unit receives continuous radar echo images, extracts local features, captures the time correlation characteristics between adjacent radar echo images through the memory recurrent perception module, and strengthens the memory of the motion characteristics of the radar echo; through the context feature fusion module, different levels of context information are blended with the input data state to obtain the global features of the radar echo image; Step 3: Adopt a multi-layer stacked MRP-LSTM recurrent unit strategy to form the MRP-Net framework, construct a flood disaster warning model and perform model training, capture the time series features and context relationships in meteorological data, and output high-resolution precipitation prediction results for the next hour, providing a key decision-making basis for the flood disaster warning system.

[0024] In this embodiment, the architecture of the MRP-LSTM recurrent unit is as Figure 1 shown, and the MRP-LSTM recurrent unit architecture presents a double-layer structure from top to bottom.

[0025] The upper branch is responsible for receiving the input states of consecutive radar echo images , cell states and hidden states , and then transmits them to the encoder for encoding operations to extract the appearance features of the images.

[0026] Immediately afterwards, use the original spatio-temporal LSTM unit (Spatio-temporal LSTM, ST-LSTM) for recurrent processing to generate the historical dependence feature representation of the input sequence .

[0027] It should be noted that in the ST-LSTM unit, the temporal and spatial information at a moment cleverly integrates the information of the previous moment and the next moment to achieve dynamic association of information.

[0028] The input sequence of the lower branch is the difference sequence of the input sequence of the upper branch , and this difference sequence accurately captures the continuous change information of the current action.

[0029] By importing this information into the Memory Recurrent Perception (MRP) module and performing query operations with the memory vector , the context memory features corresponding to the current input sequence can be effectively obtained . Subsequently, this feature is deeply fused with the local features extracted by the upper branch to generate global features , and this feature fusion mechanism enhances the representation ability for sudden heavy precipitation processes, can be used to identify the spatio-temporal evolution law of precipitation in high-risk flood disaster areas, and helps to realize the early identification and risk warning of regional flood disasters.

[0030] Furthermore, to accurately capture the time correlation between adjacent radar echo maps, this embodiment designs a Memory Recurrent Perception module, and its structure is as Figure 2As shown. The memory loop perception module mainly covers two key processes: the context perception feature storage process and the motion feature recall process.

[0031] In the context perception feature storage process, with the help of the context perception encoder extracts the perception features of the differential sequence for extraction.

[0032] Specifically, in the process of extracting image appearance features, a 1×1 convolutional filter is innovatively introduced to perform a linear transformation operation on the input image and is finely decomposed into local perception features through channel refinement , and these local perception features are separately stored in the memory component for subsequent memory query.

[0033] In the motion feature recall process, this process is similar to the previous stage. Through the context motion encoder extracts the motion features , and decomposes them into local features of the motion features. These local motion features then form a memory query vector for accurately recalling relevant echo perception features from the memory component .

[0034] This recall process is essentially a memory addressing process. In this embodiment, by cleverly combining echo perception features, the long-term motion memory vector in the memory component is continuously updated, thereby successfully extracting the corresponding long-term motion context information from the memory, improving the recognition ability of severe convective weather processes, and providing more sensitive perception support for the key meteorological signs before the formation of flood disasters.

[0035] To effectively solve the problem of forgetting long-term context-dependent features caused by weak temporal correlation between adjacent radar echo maps, this embodiment proposes a multi-layer context feature fusion module, and the specific structure is as Figure 3 shown. Among them, is the input state, is the cell state, is the hidden state.

[0036] The multi-layer context feature fusion module of this embodiment is divided into left and right parts.

[0037] The right part mainly contains historical dependence information extracted by ST-LSTM. In this part, by adding the hidden state and , multi-level context information is obtained and deeply fused with the initial input state .

[0038] Subsequently, a convolutional filter is applied to and to perform a context feature fusion operation. Specifically, through element-wise multiplication and to obtain . During this process, the function is cleverly used to normalize the values in the feature map between 0 and 1. This process is repeated, and finally, the updated input state and the hidden state are obtained.

[0039] The left part is the context memory feature recalled by the MRP module, and its working process is similar to that of the right part. Finally, the features on both the left and right sides are processed through mapping and fusion to generate the final context fusion feature, providing a richer and more accurate information basis for subsequent model prediction. This feature fusion method is particularly suitable for the fine modeling of the dynamic rainfall process and can provide a decision-making basis with higher spatio-temporal resolution for flood control scheduling and emergency response.

[0040] Furthermore, to improve the modeling accuracy of the radar echo movement and overcome the problem of gradient disappearance, this embodiment constructs an advanced radar echo extrapolation model called MRP-Net. This model forms a unique network architecture by stacking MRP-LSTM units into a 4-layer cyclic structure, as shown in Figure 4 .

[0041] In this network structure, each column represents the recursive unit of a single time step. Among them, and accurately identify the input and predicted radar echo map images at the moment respectively. Each row clearly shows the dynamic change process of the current-level memory unit and state over time, vividly reflecting the continuous update mechanism of the memory unit .

[0042] The hidden state is orderly transmitted to the next unit in the horizontal and vertical directions to ensure the smooth interaction of information in the network. In particular, represents the output result of the memory perception module at the moment and the th layer.

[0043] In the first time step, all memory units and hidden states of the first recursive unit are initialized as all-zero tensors. For the recursive units at the top layer, information will be smoothly transmitted to the bottom layer units in the next time step; for the non-top-layer recursive units, their hidden states will be transmitted to higher-level units at subsequent moments.

[0044] Through such a carefully designed process, the recursive unit can obtain multi-level context information to the greatest extent, thus effectively avoiding the problem of echo feature forgetting, significantly enhancing the timeliness prediction ability of extreme precipitation processes, providing strong technical support for the accurate early warning and risk judgment of flood disasters, and ensuring that the entire stacked structure can accurately predict future echo images.

[0045] To verify the effectiveness and reliability of the memory cycle perception short-term heavy precipitation prediction method of the present invention, a comprehensive experiment was carried out on a carefully collected and sorted radar echo dataset in this embodiment, and the MRP-Net model proposed by the present invention was compared with seven traditional mainstream network models, including: Convolutional Long Short-Term Memory (ConvLSTM), Recurrent Neural Networks for Predictive Learning using Spatiotemporal LSTMs (PredRNN), Towards a Resolution of the Deep-in-Time Dilemma in Spatiotemporal Predictive Learning (PredRNN++), Memory in Memory (MIM), Interaction Dual Attention Long Short-Term Memory (IDA-LSTM), Physical Dynamics Disentangling Network (PhyDNet), Self-Attention Convolutional Long Short - Term Memory (SA-ConvLSTM) in a systematic way.

[0046] During the experiment, the radar echo intensity thresholds were accurately set to 20, 30, and 40, and the predicted radar echo maps were converted into a 0-1 matrix form to ensure the standardization and accuracy of data processing. The experimental results were quantitatively evaluated through two authoritative evaluation indicators, the Critical Success Index (CSI) and the Heidke Skill Score (HSS). The higher the score, the better the model performance.

[0047] Through rigorous experimental comparative analysis, it is fully confirmed that the MRP-Net model has significant advantages compared with traditional models, providing a more efficient and accurate technical solution for the field of flood disaster warning.

[0048] ; ; Among them, respectively represent the number of hits, the number of false alarms, the number of missed alarms, and the number of non-hits.

[0049] The CSI numerical values of the prediction results of each model are compared in Table 1 below. Table 1 shows the comparison of the CSI numerical values of the prediction results of each model when the radar echo intensity threshold is 20, 30, and 40.

[0050] Table 1: Comparison of CSI Numerical Values of Prediction Results of Each Model

[0051] The bar chart of the comparison of the CSI numerical values of the prediction results of each model is as Figure 5 shown, Figure 5 which intuitively shows that the MRP-Net model proposed by the present invention performs best in the CSI evaluation index, while the SA-ConvLSTM and MIM models follow closely and rank second. Specifically, under the CSI-20 index, MRP-Net improves by 1.08% compared with the sub-optimal SA-ConvLSTM; under the CSI-30 index, MRP-Net improves by 2.96% compared with the sub-optimal SA-ConvLSTM; under the CSI-40 index, MRP-Net improves by 5.85% compared with the sub-optimal MIM.

[0052] The HSS numerical values of the prediction results of each model are compared in Table 2 below. Table 2 shows the comparison of the HSS numerical values of the prediction results of each model when the radar echo intensity threshold is 20, 30, and 40.

[0053] Table 2: Comparison of HSS Numerical Values of Prediction Results of Each Model

[0054] The bar chart of the comparison of the HSS numerical values of the prediction results of each model, as Figure 6 shown, Figure 6It is clearly shown that the proposed MRP-Net model of the present invention performs best in terms of the HSS evaluation index, leading the SA-ConvLSTM and MIM models, with the latter ranking second respectively. Specifically, under the HSS-20 index, MRP-Net improves by 4.28% compared to the sub-optimal SA-ConvLSTM; under the HSS-30 index, MRP-Net improves by 6.54% compared to the sub-optimal MIM; under the HSS-40 index, MRP-Net improves by 4.99% compared to the sub-optimal SA-ConvLSTM.

[0055] Furthermore, the present invention uses two indicators, namely the Mean Square Error (MSE) and the Structure Similarity Index Measure (SSIM), to evaluate the performance of the model. MSE is used to calculate the average of the sum of the squares of the differences in pixel values between two images, mainly measuring the difference between the predicted image and the actual image at each pixel point. A lower MSE value means a smaller difference between the predicted image and the real image.

[0056] ; Among them, represents the number of samples, respectively represent the true value and the predicted value of the th sample.

[0057] SSIM is an index used to evaluate the visual quality of images, aiming to measure the similarity between two images. It comprehensively considers multiple factors such as the brightness, contrast, and structure of the images. A higher SSIM value indicates that the predicted image and the real image are more visually similar, and the value range of SSIM is usually between 0 and 1.

[0058] ; Among them, represents the ground truth image, represents the restored image, represents the average value, represents the standard deviation, and are constants, whose function is to avoid errors caused by a denominator of 0.

[0059] The comparison of MSE and SSIM of the prediction results of each model at each time step is shown in Figure 7 as the comparison results of the scores at each time step under the MSE and SSIM evaluation indexes based on the collected radar echo data set.

[0060] Figure 7Among them, (a) is the score comparison step by step over time under MSE. It can be seen that the curve change of the MRP-Net model of the present invention is gentler than that of other models. This indicates that as the time step progresses, the extrapolation of MRP-Net is less affected by other factors and can better meet the needs of long-term prediction.

[0061] Figure 7 Among them, (b) is the score comparison step by step over time under SSIM. This indicates that the curve of the MRP-Net model of the present invention descends slowly and has a higher score, indicating that the model is more stable in extrapolation performance and can maintain better visual quality during the progress of the time step.

[0062] In an embodiment of the present invention, an electronic device is further provided, including: one or more processors; a storage device storing one or more programs thereon; when the one or more programs are executed by the one or more processors, the one or more processors implement the memory loop perception short-term heavy precipitation prediction method for flood disaster warning described in the above embodiment.

[0063] In an embodiment of the present invention, a computer-readable storage medium is further provided, having a computer program stored thereon, and when the program is executed by a processor, the steps in the memory loop perception short-term heavy precipitation prediction method for flood disaster warning in the above embodiment are implemented.

[0064] The above are only the preferred embodiments of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present invention.

Claims

1. A memory cycle perception short-term heavy rainfall prediction method for flood disaster warning, characterized in that: The future precipitation distribution map is generated through the MRP-Net framework, which includes the following steps: Step 1: Collect continuous radar echo image data to monitor the dynamic changes of precipitation in potential flood disaster areas in real time; Step 2, constructing an MRP-LSTM recurrent unit, including: a memory recurrent perception module, an additional memory component, and a context feature fusion module; the MRP-LSTM recurrent unit receives continuous radar echo images, extracts local features, captures the time correlation characteristics between adjacent radar echo images through the memory recurrent perception module, and strengthens the memory of radar echo motion features; through the context feature fusion module, context information at different levels is integrated with the input data state to obtain the global features of the radar echo image; Step 3: A multi-layer stacked MRP-LSTM recurrent unit strategy is used to form an MRP-Net framework, build a flood disaster warning model and perform model training, capture the time series characteristics and contextual relationships in meteorological data, output future high-resolution precipitation forecast results, and trigger flood disaster warning signals based on preset thresholds, including: risk level, impact range and recommended emergency measures, providing key decision-making basis for the flood disaster warning system.

2. The memory cycle sensing short-term heavy rainfall prediction method for flood disaster warning according to claim 1 is characterized in that: In step 2, the MRP-LSTM recurrent unit is a double-layer structure, including an upper branch and a lower branch; The upper branch receives continuous radar echo image input status , cell status and hidden state , transmitted to the encoder Encode and extract the appearance features of the input image; Use the original spatiotemporal LSTM unit for cyclic processing to generate local historical dependency features of the input sequence , input to the context feature fusion module; The original spatiotemporal LSTM unit is used for dynamic association of information, and the temporal and spatial information of the moment are integrated with the previous moment. and the next moment information; The input sequence of the lower branch is the difference sequence of the input sequence of the upper branch , differential sequence Capture the continuous change information of the current action and import it into the memory loop perception module. Perform a query to obtain the context memory features corresponding to the current differential sequence , embedded into the context feature fusion module, and combined with the local historical dependency features extracted by the upper branch Perform deep fusion to generate global features.

3. The memory cycle sensing short-term heavy rainfall prediction method for flood disaster warning according to claim 2 is characterized in that: The memory cycle perception module includes: a context perception feature storage link and a motion feature recall link, which are used to capture the time correlation characteristics between adjacent radar echo images and strengthen the memory of echo motion features; In the context-aware feature storage link, a context-aware encoder is used For the difference sequence Perceptual characteristics In the process of extracting the appearance features of the differential sequence, a 1×1 convolution filter is introduced to perform a linear transformation on the input image and decompose it into local perception features through channel refinement. The local perception feature Stored separately in memory components in, for memory query; In the motion feature recall phase, the context motion encoder is used Extract motion features , decompose the local features of the motion features, form a memory query vector, and extract the vector from the memory component Recall the corresponding echo perception features and update the memory component The long-term motion memory vector in is used to extract the corresponding long-term motion context information from the memory and transmit it to the context feature fusion module.

4. The memory cycle sensing short-term heavy rainfall prediction method for flood disaster warning according to claim 2 is characterized in that: The context feature fusion module includes: extracting historical dependency information and recalling context memory features; In the step of extracting historical dependency information, the original spatiotemporal LSTM unit extracts historical dependency information by transforming the hidden state and Add to get multi-level context information , and with the initial input state Perform deep fusion; use convolution filters to and Implement the context feature fusion operation by element-wise multiplication and get ,use The function normalizes the values ​​in the feature map between 0 and 1; repeat until the updated input state is obtained and hidden state ; In the recall context memory feature link, the context memory feature is recalled through the memory loop perception module, and a 1×1 convolution filter is introduced to adjust the number of channels to obtain the initial input state of the differential sequence , for the hidden state and Perform addition operations to obtain multi-level context information , and and Perform deep fusion; use convolution filters to and Perform context feature fusion operation and and Multiply element by element to get ,use The function normalizes the values ​​in the feature map between 0 and 1; repeat until the updated input state is obtained and hidden state .

5. The memory cycle sensing short-term heavy rainfall prediction method for flood disaster warning according to claim 4 is characterized in that: The deep fusion mentioned above balances historical memory and real-time dynamic characteristics through a weighted strategy to ensure that the generation of flood disaster warning signals takes into account both timeliness and accuracy; In the context feature fusion module, the and extracted from the recall context memory feature link Mapping and fusion are performed, and the context fusion feature is generated after calculating the Hadamard product .

6. The memory cycle sensing short-term heavy rainfall prediction method for flood disaster warning according to claim 5 is characterized in that: The MRP-LSTM recurrent unit fuses the context features Input decoder , according to the decoder Deconvolution operation, get the future time step radar echo sequence , used to predict short-term heavy rainfall before flood disasters occur.

7. The memory cycle sensing short-term heavy rainfall prediction method for flood disaster warning according to claim 1 is characterized in that: In step 3, the MRP-Net framework uses a multi-layer fusion strategy to stack several MRP-LSTM units to form a multi-layer loop structure, each column and each row includes multiple MRP-LSTM loop units, each column represents a recursive unit of a single time step, and each row represents the continuous update of the current level memory unit and state over time; The multi-layer loop structure will hide the state Continuously transfer to the next unit along the horizontal and vertical directions, Transmitted horizontally to the next unit, the output of the memory perception module It is passed to the next vertical unit in the vertical direction; by sharing context information across time steps, context information at different levels is deeply integrated with the input state.

8. The memory cycle sensing short-term heavy rainfall prediction method for flood disaster warning according to claim 7 is characterized in that: In the first time step of the MRP-Net framework, for the underlying MRP-LSTM recurrent unit, all memory cells and hidden state They are all initially set to a tensor form with all elements being zero, to ensure that the initial state of the flood disaster warning model is updated synchronously with the real-time data; For the top-level MRP-LSTM recurrent unit, The information is passed to the bottom unit at the next time step; for the non-top MRP-LSTM recurrent unit, the hidden state It is transmitted to higher-level units at subsequent moments, so that the flood disaster warning model can dynamically adapt to the sudden change characteristics of the spatiotemporal distribution of precipitation in flood disaster warning and predict future echo images.

9. An electronic device, characterized in that: include: one or more processors; a storage device having one or more programs stored thereon; When the one or more programs are executed by the one or more processors, the one or more processors implement the memory cycle perception short-term heavy rainfall prediction method for flood disaster warning as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and when the program is executed by a processor, the steps in the memory cycle perception short-term heavy rainfall prediction method for flood disaster warning described in any one of claims 1 to 8 are implemented.

Images