Mountain torrent disaster monitoring data source fusion processing method and system based on deep learning

Abstract

The invention belongs to the technical field of mountain torrent disaster monitoring, and discloses a mountain torrent disaster monitoring data source fusion processing method and system based on deep learning. The method comprises the steps of multi-source rainfall data collection, multi-source data fusion and fusion processing of multiple monitoring data sources, including data integration, data alignment, data fusion algorithm and other operations, and through fusion processing of the multiple monitoring data sources, a high-precision rainfall monitoring result is obtained. Analyzing rainfall data, and predicting the probability and range of mountain torrent disasters possibly occurring in the future according to an analysis result; and releasing mountain torrent early warning information. Based on a multi-source data fusion technology, the rainfall condition can be monitored more accurately, and the accuracy and reliability of early warning are improved. Early warning information is sent out 2 hours in advance, enough time can be provided for relevant mountain torrent defense departments to take countermeasures, and casualties and property losses are reduced.

Description

Technical Field

[0001] The present invention belongs to the technical field of flash flood disaster monitoring, and in particular to a flash flood disaster monitoring data source fusion processing method and system based on deep learning. Background Art

[0002] Flash floods are one of the typical natural disasters that threaten human survival and development. Flash floods occur randomly and affect a wide range, causing great losses and casualties, especially in southern and southwestern my country. They have become a top priority for defense, and it is very necessary to issue early warnings for areas threatened by flash floods.

[0003] The causes of flash floods involve meteorology, hydrology, topography, etc. It is not realistic to study and seek solutions to flash floods from the root. The research work on flash flood disaster prevention and control at home and abroad focuses on the forecast and warning of flash floods to effectively reduce casualties and property losses. Rainfall is one of the main factors that induce geological disasters. Continuous rainfall and local rainstorms have a greater impact on geological disasters. That is, rainfall intensity and rainfall are often used as warning indicators to achieve early warning of flash floods. At present, the main precipitation monitoring method in my country is mainly based on the ground rain gauge monitoring method, which is often insufficient to meet the requirements of heavy rainfall analysis. For non-uniform rainfall processes, it is easy to miss the center of heavy rainfall, and it cannot accurately reflect the spatial and temporal characteristics of rainfall. Moreover, it is ground rain, and it is impossible to quantitatively predict the occurrence of heavy rainfall in advance. When heavy rainfall is monitored, when a flash flood warning is issued, it is too late to warn and evacuate residents in the area affected by the flash flood. It is very necessary to improve the accurate forecast period of heavy rainfall, issue accurate warning information in advance, and notify relevant flash flood defense departments to evacuate the affected groups. Summary of the invention

[0004] In order to overcome the problems existing in the related technologies, the disclosed embodiments of the present invention provide a flash flood disaster monitoring data source fusion processing method and system based on deep learning. Specifically, it relates to a flash flood disaster warning method based on multi-source data fusion technology of rain measuring radar. The purpose of the present invention is to achieve high-precision precipitation monitoring and early warning 2 hours in advance based on multi-source data fusion technology of rain measuring radar, and solve the technical problem that the delay in ground rain monitoring cannot meet the flash flood disaster early warning service. It aims to monitor the precipitation situation with high precision 2 hours in advance and issue early warning information so that the relevant flash flood defense departments can evacuate the affected people in time to reduce casualties and property losses.

[0005] The technical solution is as follows: A flash flood disaster monitoring data source fusion processing method based on deep learning, comprising:

[0006] S1, multi-source rainfall data collection: using rainfall radar and ground rain gauge to obtain rainfall data, and preprocessing and screening the obtained rainfall data;

[0007] S2, multi-source data fusion: multiple monitoring data sources are fused to obtain high-precision precipitation monitoring results;

[0008] S3, rainfall data analysis: according to the monitoring results, analyze the rainfall intensity and rainfall indicators; according to the analysis results, predict the probability and scope of future flash flood disasters;

[0009] S4, Flash flood warning information release: Issue warning information in advance based on forecast results.

[0010] In step S1, the acquired rainfall data is preprocessed and screened, including data cleaning and format conversion.

[0011] In step S2, multiple monitoring data sources include: colorful cloud weather forecast data, short-term precipitation forecast data;

[0012] The fusion processing includes: data integration, data alignment, and data fusion.

[0013] In step S2, multiple monitoring data sources are fused and processed, including:

[0014] S201, by means of auxiliary manual labeling, marking non-precipitation echoes in the historical radar image, and constructing a radar denoising training data set; inputting the radar images of the previous and next two frames of each sample in the radar denoising training data set into a deep neural network, performing denoising using the radar denoising model constructed by the deep neural network, and outputting the denoised radar image;

[0015] S202, taking 2 frames of input and 25 frames of forecast radar images as one sample, constructing a radar forecast data set based on multiple sample sequences, and inputting the radar forecast data set into a precipitation forecast model to achieve precipitation estimation for a certain period in the future;

[0016] S203, based on the radar denoising model, precipitation forecast model, satellite inversion, and Himawari8 and radar observation puzzles, a data set is created. The current and previous 30-minute observations are used as input, and the radar puzzle is used as output. Resnet and ConvNeXt are used to invert the radar echo intensity, and GAN is used to guide the training to make the inverted radar map clear in texture, supplement the radar real-time conditions in terrain occlusion, mountainous areas, and ocean areas, and obtain the inverted radar and the next 2-hour precipitation forecast model;

[0017] S204, build a short-term precipitation forecast platform, update the obtained radar data and forecast data, and realize the pattern correction of precipitation, temperature, humidity, air pressure, wind speed and dew point temperature through the partial differential equation neural network model, realize the prediction of short-term precipitation forecast data 2 hours in advance, and issue early warning information.

[0018] In step S201, a radar denoising training data set is constructed, including: the high-resolution network HRnet gradually joins the low-resolution feature map sub-network in parallel through the high-resolution feature map main network, multi-scale fusion and feature extraction are realized through different networks, and SEBlock is used to add an attention mechanism to the multi-scale feature fusion to construct a radar denoising model;

[0019] The use of SEBlock to add an attention mechanism to multi-scale feature fusion includes: the channel attention mechanism submodule in the SEBlock attention mechanism module is G c , the spatial attention mechanism submodule is G s ;

[0020] The channel attention mechanism submodule uses global average pooling and maximum pooling to highlight the feature information of multi-scale feature areas. The expression is:

[0021]

[0022] Where x( ) is the sigmoid nonlinear activation function, SAG( ) is the shared network in the SEBlock attention mechanism module, AvgPool() is the average pooling function, MaxPool() is the maximum pooling function, D is the feature information of the multi-scale feature region, E0, E1 are the hidden layer weights and output layer weights in the multi-layer perceptron in SAG, respectively, and E0∈H Clr×B ,E1∈H B ×Clr , are the global average pooling features and the maximum pooling features respectively; H is the weight value, Clr is the number of hidden layers, and B is the number of output layers;

[0023] Under the action of the channel attention mechanism, the spatial attention mechanism submodule is introduced, and the expression is:

[0024]

[0025] In the formula, e 7×7 is a 7×7 convolution operation, They are the global average pooling features and the maximum pooling features in space, F is the feature information of the multi-scale feature area in space, A is the width, and E is the height.

[0026] In step S202, a radar forecast data set is constructed based on a plurality of sample sequences, and the radar forecast data set is input into a precipitation forecast model, including:

[0027] Based on Unet and SEBlock, the deep rainfall generation model DGMR of DeepMind is integrated with the multimodal pre-training model overview. A discriminator is added to the DGMR model, and a loss of image authenticity is added to the loss function to build a precipitation forecast model. Unet integrated with SEBlock is integrated with DGMR to build a general network structure fusion model, and the nodes in the network are divided into two subsets Vl and Vunl according to whether there are forecast frame radar map labels in the radar forecast data set. Vl is the node set with forecast frame radar map labels, and Vunl is the node set without forecast frame radar map labels.

[0028] The nodes in Vl are divided into k subsets according to their categories, where k represents the total number of node features in different frames, denoted as: Vl = {L1, L2…L k}, and record the label corresponding to the node v0∈Vl in Vl as The set of first-order neighbor nodes of node v0 with the forecast frame radar chart label is recorded as

[0029] The network structure is fused by adding edges between nodes with forecast frame radar chart labels and belonging to the same frame features. o Calculate the number of edges added at one end The expression is:

[0030]

[0031] In the formula, For node v o The first-order neighboring nodes with the forecast frame radar chart label are followed by v o The number of nodes with the same label as v, γ is o The number of edges established when none of the first-order neighbors of the forecast frame radar chart label, v j is the jth node, is the label corresponding to the jth node, μ is v o First-order labeled neighbor nodes with v o The preset proportion of nodes with the same label; σ is the equivalence judgment function. If but Returns 1 if yes, otherwise returns 0.

[0032] In step S204, a short-term precipitation forecast platform is constructed, and the acquired radar data and forecast data are updated, including: learning personalized abnormal parameters and guessed parameters from the feature vectors of the included radar data and forecast data, integrating the abnormal and guessed parameters, obtaining a new feature vector, and predicting the performance of the short-term precipitation forecast platform in a specific environment;

[0033] First, learn personalized anomaly parameters and guess parameters from the feature vector, expressed as:

[0034] q′ t =tanh(I1L t +b1)

[0035] y′ t =tanh(I2L t +b2)

[0036] Where, L″ t is the label vector at time t after the update, F″ t ' is the feature information of the multi-scale feature region in the space at time t after the update, L' t To integrate the label vector at time t based on the short-term precipitation forecast sample to be predicted, F' t In order to integrate the characteristic information of multi-scale characteristic regions in space at time t based on the short-term precipitation forecast samples to be predicted, To incorporate the operator, q' t To learn personalized anomaly parameters from feature vectors, y' t To learn personalized guess parameters from feature vectors;

[0037] Secondly, the personalized abnormal parameters q' and guessed parameters y' based on the short-term precipitation forecast samples to be predicted and the abnormal parameters q and guessed parameters y based on the cognitive diagnosis model are integrated to obtain a new feature vector:

[0038]

[0039]

[0040] In the formula, q' t To learn personalized anomaly parameters from feature vectors, y' t To learn personalized guess parameters from feature vectors, I1 is the radar data parameter, I2 is the forecast data parameter, and L t is the label vector at time t, b1 is the feature constant vector in radar data learning, and b2 is the feature constant vector in forecast data learning;

[0041] Finally, the new feature vector obtained is activated through the sigmoid function to predict the performance of the short-term precipitation forecast platform in a specific environment; the performance of the short-term precipitation forecast platform V t The formula is as follows:

[0042]

[0043] Where V tis the performance value of the short-term precipitation forecast platform, β is the performance coefficient of the short-term precipitation forecast platform, I3 is the short-term precipitation forecast parameter, T is the order, and b3 is the constant value in the performance of the short-term precipitation forecast platform.

[0044] In step S204, the pattern correction of precipitation, temperature, humidity, air pressure, wind speed and dew point temperature is realized by using the partial differential equation neural network model, including:

[0045] Regional model correction: A radar mosaic composed of multiple radars, hourly ground observation data from multiple control stations, and regional numerical model forecast data are used to construct a numerical model forecast correction dataset, and a deep learning model is used to implement post-model correction of precipitation, 2m temperature, humidity, air pressure, wind speed, and dew point temperature;

[0046] Regional model forecast: real-time precipitation regression, 500pha geopotential height extrapolation, and use of -3-hour and 0-hour regional model upper-air variables to forecast precipitation for the next 12 hours, and learn the movement trend of precipitation;

[0047] The error of precipitation forecast data is controlled by using the partial differential equation neural network model PINNs with neural network automatic differentiation.

[0048] Extreme precipitation forecasts are also provided through NowcastNet.

[0049] Another object of the present invention is to provide a flash flood disaster monitoring data source fusion processing system based on deep learning, the system implements the flash flood disaster monitoring data source fusion processing method based on deep learning, the system comprises:

[0050] Multi-source rainfall data collection module, used to obtain rainfall data using rainfall radar and ground rain gauge, and to pre-process and filter the acquired rainfall data;

[0051] The multi-source data fusion module is used to fuse multiple monitoring data sources to obtain high-precision precipitation monitoring results;

[0052] The rainfall data analysis module is used to analyze rainfall intensity and rainfall indicators based on monitoring results; and to predict the probability and scope of future flash flood disasters based on the analysis results;

[0053] The flash flood warning information release module is used to issue warning information in advance based on the prediction results.

[0054] In combination with all the above technical solutions, the beneficial effects of the present invention are as follows: the present invention includes the processing of rain radar monitoring data and rain gauge monitoring data, the fusion processing of rain radar data and rain gauge data, and the association of rain radar coverage area with flash flood warning objects, and realizes visual warning in the form of flashing layer objects. Warning objects include warnings of objects affected by flash flood areas. The present invention provides a method for realizing the key steps of multi-source rainfall data collection, multi-source rainfall data fusion, rainfall data analysis, and flash flood warning information release.

[0055] The present invention is based on multi-source data fusion technology, which can monitor precipitation more accurately and improve the accuracy and reliability of early warning. The present invention issues early warning information 2 hours in advance, which can give relevant mountain flood defense departments enough time to take countermeasures and reduce casualties and property losses. The present invention can monitor rainfall conditions in real time, avoid missing the center of heavy rainfall, and improve the timeliness of early warning. The present invention can perform zoned monitoring of the monitoring area, and carry out different early warning levels and countermeasures for different areas, thereby improving the pertinence of early warning. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] The accompanying drawings herein are incorporated in and constitute a part of the specification, illustrate embodiments consistent with the present disclosure, and together with the description, serve to explain the principles of the present disclosure;

[0057] Figure 1 It is a flow chart of a flash flood disaster monitoring data source fusion processing method based on deep learning provided by an embodiment of the present invention;

[0058] Figure 2 is a schematic diagram of a radar denoising model provided by an embodiment of the present invention;

[0059] Figure 3 is a radar denoising flow chart provided by an embodiment of the present invention;

[0060] Figure 4 is a schematic diagram of a radar denoising network framework provided by an embodiment of the present invention;

[0061] Figure 5 Comparison chart of ground noise denoising effect;

[0062] Figure 6 Comparison chart of ray noise denoising effect;

[0063] Figure 7 Comparison chart of clear sky echo denoising effect;

[0064] Figure 8 This is a schematic diagram of radar quality control accuracy;

[0065] Fig. 9 This is a schematic diagram of the precipitation forecast model;

[0066] Fig.10 This is the forecast principle diagram of the precipitation forecast model;

[0067] Fig.11 A forecast effect map of a city using a precipitation forecast model;

[0068] Fig.12 This is a schematic diagram of precipitation estimation;

[0069] Fig.13 MAE plot for precipitation estimation;

[0070] Fig.14 This is a schematic diagram of TS scoring for Global Satellite Inversion Radar.mov;

[0071] Fig.15 This is the radar map inversion result;

[0072] Fig.16 It is the inversion + real observation map;

[0073] Fig.17 Schematic diagram for using Reduced-LSTM and GNN graph neural network models to achieve hourly PM2.5 air quality forecast for the next 15 days;

[0074] Fig.18 This is the 24-hour small, medium and heavy rainstorm forecast accuracy chart;

[0075] Fig.19 Figure 1 is a partial differential equation neural network model (PINNs (Physics-informed Neural Networks) diagram;

[0076] Fig. 20 It is the signal strength-precipitation distribution diagram;

[0077] Fig.21 It is the signal strength-clear sky distribution map;

[0078] Fig. 22 This is the signal strength-rainy day distribution diagram. DETAILED DESCRIPTION

[0079] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below in conjunction with the accompanying drawings. In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without violating the connotation of the present invention, so the present invention is not limited by the specific implementation disclosed below.

[0080] Embodiment 1, as Figure 1As shown, the flash flood disaster monitoring data source fusion processing method based on deep learning provided by the embodiment of the present invention includes:

[0081] S1, multi-source rainfall data collection: using rainfall radar and ground rain gauge to obtain rainfall data, and preprocessing and screening the obtained rainfall data;

[0082] Including data cleaning, format conversion and other operations to ensure data accuracy and availability;

[0083] S2, multi-source data fusion: multiple monitoring data sources are fused to obtain high-precision precipitation monitoring results;

[0084] Fusion processing includes operations such as data integration, data alignment, and data fusion algorithms;

[0085] S3, rainfall data analysis: according to the monitoring results, analyze the rainfall intensity and rainfall indicators; according to the analysis results, predict the probability and scope of future flash flood disasters;

[0086] S4, Flash Flood Warning Information Release: Issue warning information in advance based on the forecast results;

[0087] The warning information will be promptly notified to relevant flash flood prevention departments, including by SMS, phone calls, etc.

[0088] The present invention also provides a flash flood disaster monitoring data source fusion processing system based on deep learning, the system comprising:

[0089] Multi-source rainfall data collection module, used to obtain rainfall data using rainfall radar and ground rain gauge, and to pre-process and filter the acquired rainfall data;

[0090] The multi-source data fusion module is used to fuse multiple monitoring data sources to obtain high-precision precipitation monitoring results;

[0091] The rainfall data analysis module is used to analyze rainfall intensity and rainfall indicators based on monitoring results; and to predict the probability and scope of future flash flood disasters based on the analysis results;

[0092] The flash flood warning information release module is used to issue warning information in advance based on the prediction results.

[0093] Example 2, illustrative, one implementation of the present invention, in step S2, fusing multiple monitoring data sources includes:

[0094] S201, in a manner assisting manual labeling, non-precipitation echoes in the historical radar image are marked to construct a radar denoising training data set; the radar images of the two frames before and after each sample in the radar denoising training data set are input into a deep neural network, and the radar denoising model constructed by the deep neural network is used for denoising, and the denoised radar image is output;

[0095] S202, based on the radar image after radar denoising, taking 2 frames as input and 25 frames of radar image forecast as one sample, constructing a radar forecast data set based on multiple sample sequences, and inputting the radar forecast data set into a precipitation forecast model to realize precipitation estimation for a certain period in the future;

[0096] S203, according to the radar denoising model and precipitation forecast model, combined with satellite inversion, a dataset is produced based on Himawari8 and the radar observation puzzle of a certain area, with the current and previous 30 minutes of observations as input and the radar puzzle as output, using Resnet and ConvNeXt to invert the radar echo intensity, and using GAN to guide the training, so that the texture of the inverted radar map is clear, supplementing the radar real situation in terrain occlusion, mountainous and oceanic areas, and obtaining the radar based on inversion and the precipitation forecast model for the next 2 hours;

[0097] S204, construct a short-term precipitation forecast platform to update the radar data and forecast data obtained in steps S201-S203, and realize the pattern correction of precipitation, 2m temperature, humidity, air pressure, wind speed and dew point temperature through the partial differential equation neural network model, realize the prediction of short-term precipitation forecast data 2 hours in advance, and issue early warning information.

[0098] Exemplarily, in step S201, constructing a radar denoising model includes: the high-resolution network HRnet gradually adds a low-resolution feature map sub-network in parallel to the high-resolution feature map main network, different networks realize multi-scale fusion and feature extraction, and finally use SEBlock to add an attention mechanism to the multi-scale feature fusion to construct a radar denoising model;

[0099] The method of adding attention mechanism to multi-scale feature fusion using SEBlock includes: the channel attention mechanism submodule in the SEBlock attention mechanism module is G c , the spatial attention mechanism submodule is G s ;

[0100] The channel attention mechanism submodule uses global average pooling and maximum pooling to highlight the feature information of multi-scale feature areas. The expression is:

[0101]

[0102] Where x( ) is the sigmoid nonlinear activation function, SAG( ) is the shared network in the SEBlock attention mechanism module, AvgPool( ) is the average pooling function, MaxPool( ) is the maximum pooling function, D is the feature information of the multi-scale feature region, E0, E1 are the hidden layer weights and output layer weights in the multi-layer perceptron in SAG, respectively, and E0∈H Clr×B ,E1∈H B ×Clr , are the global average pooling features and the maximum pooling features respectively; H is the weight value, Clr is the number of hidden layers, and B is the number of output layers;

[0103] Under the action of the channel attention mechanism, the spatial attention mechanism submodule is introduced, and the expression is:

[0104]

[0105] In the formula, e 7×7 is a 7×7 convolution operation, They are the global average pooling features and the maximum pooling features in space, F is the feature information of the multi-scale feature area in space, A is the width, and E is the height.

[0106] Exemplarily, in step S202, the method for constructing a precipitation forecast model includes:

[0107] Based on Unet and SEBlock, the deep rainfall generation model DGMR of DeepMind is integrated with the multimodal pre-training model overview. A discriminator is added to the DGMR model, and a loss of image authenticity is added to the loss function to build a precipitation forecast model. Unet integrated with SEBlock is integrated with DGMR to build a general network structure fusion model, and the nodes in the network are divided into two subsets Vl and Vunl according to whether there are forecast frame radar map labels in the radar forecast data set. Vl is the node set with forecast frame radar map labels, and Vunl is the node set without forecast frame radar map labels.

[0108] The nodes in Vl are divided into k subsets according to their categories, where k represents the total number of node features in different frames, denoted as: Vl = {L1, L2…L k}, and record the label corresponding to the node v0∈Vl in Vl as The set of first-order neighbor nodes of node v0 with the forecast frame radar chart label is recorded as

[0109] The network structure is fused by adding edges between nodes with forecast frame radar chart labels and belonging to the same frame features. lMiddle node v o Calculate the number of edges added at one end The expression is:

[0110]

[0111] In the formula, For node v o The first-order neighboring nodes with the forecast frame radar chart label are followed by v o The number of nodes with the same label as v, γ is o The number of edges established when none of the first-order neighbors of the forecast frame radar chart label, v j is the jth node, is the label corresponding to the jth node, μ is v o First-order labeled neighbor nodes with v o The preset proportion of nodes with the same label; σ is the equivalence judgment function. If but Returns 1 if yes, otherwise returns 0.

[0112] From v o Select nodes belonging to the same category nodes to join the edge; select the root node v o Nodes that are farther away are added to the edge to speed up the radar chart prediction speed between feature nodes in the same frame; after learning the node vector representation in the first stage, the node vector is embedded with the proximity between nodes, and v is used to represent the node vector representation. o and The similarity between the vectors of the nodes is represented by v o and The distance between midpoints; Select the one with the smallest similarity nodes and v o Establish edges between them to integrate v o The closeness between nodes belonging to the same category can speed up the prediction of related label information between nodes.

[0113] Exemplarily, in step S204, constructing a short-term precipitation forecast platform to update the radar data and forecast data obtained in steps S201 to S203 includes:

[0114] The personalized abnormal parameters and guessed parameters are learned from the feature vector containing radar data and forecast data, and the abnormal and guessed parameters are integrated to obtain a new feature vector to predict the performance of the short-term precipitation forecast platform in a specific environment; specifically:

[0115] First, learn personalized anomaly parameters and guess parameters from the feature vector, expressed as:

[0116] q' t =tanh(I1L t +b1)

[0117] y′ t =tanh(I2L t +b2)

[0118] In the formula, q' t To learn personalized anomaly parameters from feature vectors, y' t To learn personalized guess parameters from feature vectors, I1 is the radar data parameter, I2 is the forecast data parameter, and L t is the label vector at time t, b1 is the feature constant vector in radar data learning, and b2 is the feature constant vector in forecast data learning;

[0119] Secondly, the personalized abnormal parameters q' and guessed parameters y' based on the short-term precipitation forecast samples to be predicted and the abnormal parameters q and guessed parameters y based on the cognitive diagnosis model are integrated to obtain a new feature vector:

[0120]

[0121] Where L' t ' is the label vector at time t after the update, F' t ' is the feature information of the multi-scale feature region in the space at time t after the update, L' t To integrate the label vector at time t based on the short-term precipitation forecast sample to be predicted, F' t In order to integrate the characteristic information of multi-scale characteristic regions in space at time t based on the short-term precipitation forecast samples to be predicted, To incorporate the operator, q' t To learn personalized anomaly parameters from feature vectors, y' t To learn personalized guess parameters from feature vectors;

[0122] Finally, the new feature vector obtained is activated through the sigmoid function to predict the performance of the short-term precipitation forecast platform in a specific environment; the performance of the short-term precipitation forecast platform V t The formula is as follows:

[0123]

[0124] Where V t is the performance value of the short-term precipitation forecast platform, β is the performance coefficient of the short-term precipitation forecast platform, I3 is the short-term precipitation forecast parameter, T is the order, and b3 is the constant value in the performance of the short-term precipitation forecast platform.

[0125] In step S204, the pattern correction of precipitation, 2m temperature, humidity, air pressure, wind speed and dew point temperature is implemented by the partial differential equation neural network model, including:

[0126] Regional model revisions:

[0127] A radar mosaic composed of multiple radars, hourly ground observation data from multiple control stations, and regional numerical model forecast data were used to construct a numerical model forecast correction dataset, and a deep learning model was used to implement post-model correction of precipitation, 2m temperature, humidity, air pressure, wind speed, and dew point temperature;

[0128] Regional model forecast: real-time precipitation regression, 500pha geopotential height extrapolation, and use of -3-hour and 0-hour regional model upper-air variables to forecast precipitation for the next 12 hours, and learn the movement trend of precipitation;

[0129] The error of precipitation forecast data is controlled by using the partial differential equation neural network model PINNs with neural network automatic differentiation.

[0130] In step S204, extreme precipitation forecast is also performed through NowcastNet.

[0131] Example 3, as another implementation of the present invention, in step S2, the multi-source data includes colorful cloud weather forecast data and short-term precipitation forecast data;

[0132] Exemplary, for Caiyun weather forecast data.

[0133] On September 30, 2012, the Top-5 classification error rate of AlexNet developed by Hinton's group on ImageNet dropped to 15.3%.

[0134] Weather forecasts are closely related to people's daily lives and social production, and have an important impact on all areas of society. Among them, short-term precipitation forecasts have long been a pain point in weather forecasts. Traditional meteorological numerical models use supercomputers to achieve long-term weather forecasts using atmospheric dynamics equations based on various observation data such as sounding data, satellites, and automatic meteorological stations. However, due to computing power limitations, traditional meteorological numerical models also have corresponding limitations on the temporal and spatial resolution of forecasts and the update frequency of forecasts, making it difficult to achieve kilometer-by-kilometer and minute-by-minute forecasts.

[0135] Meteorological radar can detect precipitation near the ground by emitting electromagnetic waves. The stronger the radar echo, the stronger the precipitation. Most radars detect once every 6 minutes, and can detect precipitation information within a range of 230km each time, with a resolution of up to 1km, which lays a solid foundation for observing the actual precipitation situation at the kilometer and minute levels. In recent years, machine learning technology has developed rapidly in image, speech, and natural language processing. It has been widely used in many fields such as video object detection, autonomous driving, and speech synthesis, and has generated huge social value. It has also established a good theoretical and engineering foundation for the application of deep learning in the field of meteorology.

[0136] Based on the above existing situation, the present invention makes the following improvements to the colorful cloud weather forecast data.

[0137] 1. Since 2014, the Caiyun Weather team has taken advantage of deep learning and meteorological big data to improve the accuracy of precipitation forecasts based on observation data from ground stations, weather radars, satellites, and traditional meteorological numerical model forecasts, achieving:

[0138] (1) Short-term precipitation forecast at the minute and kilometer level for the next 2 hours, realizing precipitation estimation;

[0139] (3) Global satellite inversion radar;

[0140] (4) Air pollutant forecast;

[0141] (5) Neural network forecast based on meteorological numerical model for the next 24 hours.

[0142] The short-term precipitation forecasting technology of the present invention is combined with the latest deep learning technology and improved step by step. In radar denoising and forecasting, the present invention evolves from traditional radar denoising algorithms and optical flow forecasting to deep neural network models. As the forecast time increases, the role of radar in forecasting gradually decreases. Therefore, in the precipitation forecast after two hours, the present invention uses a deep neural network to integrate ground observation stations, radar and numerical model forecasts to further improve short-term precipitation forecasts.

[0143] 2. Radar quality control.

[0144] In the process of radar detecting precipitation, interference echoes caused by flocks of birds, ground objects, air refraction, etc. will also be received. How to separate the precipitation echo from the original radar echo is a difficult task. It is difficult to solve it with traditional methods, but the present invention uses a deep neural network method to achieve good results in radar denoising.

[0145] 2.1 Radar quality control data set; Caiyun Weather is a member of the Meteorological Social Observation Alliance, and has access to user feedback data sent back by terminal users such as the Caiyun Weather App, 360, and mobile phones. Based on years of accumulated radar data, and by combining user feedback and precipitation observations from meteorological observation stations, the present invention uses auxiliary manual labeling to mark out non-precipitation echoes in historical radar maps and constructs a radar denoising training data set. Each sample takes the radar map of the previous and next two frames as input, and the labeled denoised radar map as output. Figure 2 Schematic diagram of radar denoising model.

[0146] 2.2 Denoising: Radar denoising is a standard image semantic segmentation task. In order to achieve more sophisticated denoising, the present invention tests models such as GCN and HRnet models and integrates SEBlock. HRnet gradually adds low-resolution feature map sub-networks to the high-resolution feature map main network in parallel, and different networks achieve multi-scale fusion and feature extraction. Finally, SEBlock is used to add an attention mechanism to the multi-scale feature fusion, making the model perform better. Figure 3 , Figure 4 shown.

[0147] In the radar rainfall situation, the radar real-time model can be well represented, and has a good denoising effect on ground noise, ray noise and clear sky echo. Figure 5 Comparison of ground noise denoising effects. Figure 6 Comparison chart of ray noise denoising effect. Figure 7 Comparison chart of clear sky echo denoising effect;

[0148] Take the heavy rain in a certain area on the afternoon of May 21, 2020 as an example. Figure 5 , Figure 6 , Figure 7 The middle left part is a comparison of radar images of a city before and after denoising at 15:30 in the afternoon. The radar image before denoising contains light-colored non-precipitation echoes and ray clutter in the southeast direction. In the radar image after denoising, it can be clearly seen that the denoising model accurately removes those non-precipitation echoes. Figure 5 , Figure 6 , Figure 7 The middle right part shows the accurate removal of ground noise. Figure 8 shown.

[0149] 3 Radar forecast.

[0150] 3.1 Dataset; After radar denoising, accurate precipitation information can be provided for rainfall forecast. Based on the quality-controlled radar data, the present invention takes 2 frames as input and 25 frames of radar images as a sample, and constructs a radar forecast dataset based on tens of millions of radar image sequences. The radar image has one frame every 6 minutes, and 25 frames of radar images are forecasted, which can achieve a forecast for the next 2.5 hours, providing radar delay tolerance time for business operations. Fig. 9 Schematic diagram of the precipitation forecast model.

[0151] 3.2 Precipitation forecast model; In the task of radar forecasting, the present invention gradually evolves from the optical flow method and Tracknn single-point forecast to radar forecast mapping. In the latest precipitation forecast model, the present invention is based on Unet, incorporates SEBlock, ResBlock, etc., and combines the two on the basis of DeepMind's DGMR model, which can maintain a high forecast accuracy. In the DGMR model, the idea of ​​GAN is included, a discriminator model is added, and a Loss of picture authenticity is added to the loss function, making the radar forecast map more realistic. On the test set, the CSI of the radar forecast for the next 2.5 hours changes from about 0.8 to about 0.4. Fig.10 Diagram of the forecasting principle of the precipitation forecast model. Fig.11 Forecast results of precipitation forecast model for a certain city.

[0152] 4Precipitation Estimate (QPE).

[0153] There is a strong correlation between radar intensity and precipitation, but it is difficult to accurately calculate precipitation using empirical formulas. The present invention constructs a dataset for precipitation estimation using radar images and 1-hour cumulative rainfall data from meteorological observation stations, aligns the location of the station with the location of the radar, inputs 10 radar images in one hour, and outputs one-hour precipitation. The present invention trains the ResUnet model so that the model can use radar dbz data to deduce precipitation with a resolution of 1km. The schematic diagram of precipitation estimation is shown in the figure below. Fig.12 shown.

[0154] On a dataset of more than 2,000 observation stations, the MAE of colorful cloud precipitation estimation is 0.977 mm / h, which is 35% higher than the commonly used ZR relationship in the industry. Fig.13 The MAE plot of precipitation estimation is shown.

[0155] 5 satellite inversion radar.

[0156] Real radar observations are difficult to cover mountainous areas, oceans and other areas, while observations from geostationary meteorological satellites can cover the entire world. Therefore, the present invention produces a dataset based on Himawari8 and a radar observation puzzle of a certain area, with the current and previous 30-minute observations as input and the radar puzzle as output. The radar echo intensity is inverted using Resnet and ConvNeXt, and GAN is used to guide the training, making the texture of the inverted radar map clearer, with a TS score of 0.4 at 15 dBZ and a TS score of 0.2 at 35 dBZ, effectively supplementing the radar real-time situation in areas such as terrain occlusion, mountains, and oceans, and extending this application migration to the global scope. In addition, based on the inverted radar and the future 2-hour forecast model, a global radar forecast with a resolution of 4 km and 10 minutes is produced.

[0157] in, Fig.14 This is a schematic diagram of TS scoring for Global Satellite Inversion Radar.mov; Fig.15 This is the radar map inversion result; Fig.16 It is the inversion + real observation map.

[0158] 6. Air pollutant forecast.

[0159] Caiyun uses GFS weather numerical forecasts and real-time PM2.5 data observed by air quality monitoring stations, and uses Reduced-LSTM and GNN graph neural network models to achieve hourly PM2.5 air quality forecasts for the next 15 days. Fig.17 shown.

[0160] 7. Stability of precipitation forecast platform.

[0161] In order to provide a weather API for short-term precipitation forecast, the present invention has designed and developed a short-term precipitation forecast platform, and some services use LAIN and K8S platforms to achieve automated operation and maintenance. The time from receiving radar data to updating precipitation forecast is controlled within 10s, making full use of the timely observation of radar. The present invention uses kafka, redis, mongodb, etc. to transmit and store radar data and forecast data. The short-term precipitation forecast platform processes 300 to 400 million calls every day, with an average service stability of 99.94%.

[0162] 8Neural network numerical model forecasting.

[0163] 8.1 Regional model revision.

[0164] A numerical model forecast correction dataset was constructed using a radar puzzle consisting of more than 200 radars, hourly ground observation data from more than 2,000 national control stations, and regional numerical model forecast data. A deep learning model was used to implement post-model corrections for precipitation, 2m temperature, humidity, air pressure, wind speed, and dew point temperature. In the 24-hour rainfall forecast correction, overall, Caiyun Weather's 24-hour precipitation forecast has improved significantly compared to the regional model precipitation forecast for light, medium, and heavy rain; the TS score for the first 4 hours of heavy rain has also improved, but the TS score after 4 hours is not as high as the original forecast of the model. The 24-hour rainfall forecast accuracy rate for light, medium, and heavy rain is shown in the figure below: Fig.18 As shown. Fig.18 Figure (a) shows light rain. Fig.18 Figure (b) shows moderate rain. Fig.18 Figure (c) shows heavy rain. Fig.18 Figure (d) shows heavy rain.

[0165] In the 24-hour forecast correction, compared with the traditional model forecast, the MAE and MSE of temperature, dew point temperature, relative humidity, U wind, V wind and surface pressure are also significantly reduced on the test set.

[0166] 8.2 Regional Mode.

[0167] 1) Real-time precipitation regression: Using the regional model high-altitude variables at -3 hours and 0 hours to regress precipitation, precipitation in a large area can be regressed, but the rainfall intensity will be smaller.

[0168] 2) 500pha potential height extrapolation: In the problem of GPH potential height extrapolation, the loss value of PredRNN v2 (loss = 5.3e-7) is better than that of PhyDNet (loss = 6.9e-7), and the model is lighter. The prediction of PhyDNet can better maintain the morphology, but for the simulation of the potential field and ridge position, the prediction of PredRNN v2 is more accurate.

[0169] 3) Using the upper air variables of the regional model at -3 hours and 0 hours to predict the precipitation in the next 12 hours, we can learn the movement trend of precipitation. However, as the forecast time increases, the number of false positives will increase. Comparison of precipitation forecast maps: From the original results of the model forecast, the results obtained by extrapolating the upper air variables, and the results of the precipitation extrapolation, it can be clearly seen that using the upper air variables extrapolation, the model can better restore the structure of this frontal precipitation process and obtain results closer to the model forecast.

[0170] 8.3 Partial differential equation neural network model (PINNs (Physics-informed Neural Networks). Fig.19 shown.

[0171] When solving differential equations, traditional finite difference methods and finite element methods have the disadvantage of stage errors. Using Physics-informed Neural Networks (PINNs), the present invention can make full use of the characteristics of automatic differentiation (autodiff) of neural networks for implicit solution, avoiding the stage errors caused by gridding methods. The DeepXDE architecture built on this basis can be used to conveniently carry out PINNs experiments, which provides an interface for setting initial boundary conditions and the geometry of the test area.

[0172] The PINNs method does not require grid generation, and because of the use of autodiff, it is relatively insensitive to noise in the data, and can be extended to more types of differential equations, such as integral differential equations, inverse problems of differential equations, etc.

[0173] 9API call data analysis.

[0174] 9.1 Sunny and rainy detection based on API call volume.

[0175] Over a period of time, user calls usually increase with the increase of precipitation and precipitation frequency. Analysis of the relationship between call volume and precipitation and experiments show that it is difficult to directly determine precipitation through call volume, but it is possible to detect whether it is sunny or rainy. Based on user call volume, a 10km, 1-hour resolution dataset was created. The call volume and time code for the past 6 hours were input, and the sunny or rainy classification was output.

[0176] 9.2 Electromagnetic rainfall measurement.

[0177] Affected by weather, especially precipitation, communication signals will attenuate during transmission. Generally, the higher the communication frequency (above 4GHz), the greater the attenuation. The precipitation can be inverted using the signal strength between base stations and base stations, and between base stations and terminals. Based on the signal strength received by the terminal (mobile phone) and the corresponding precipitation, the present invention explores the relationship between signal strength and precipitation by analyzing correlation, viewing data distribution, fitting, etc. Fig. 20 is the signal strength-precipitation distribution, such as Fig.21 is the signal strength-clear sky distribution, such as Fig. 22 is the signal strength-rainy day distribution.

[0178] For example, for short-term precipitation forecast data. The short-term precipitation forecast, that is, the prediction of small and medium-scale precipitation within 6 hours, is currently mainly achieved by using the continuous change characteristics of weather radar echoes combined with the coupled equations of various physical quantities to build a forecast model. In addition to traditional numerical forecast models, some deep learning models have also begun to be widely used in the field of short-term precipitation forecasts in recent years. The principle is to use a large amount of weather radar echo observation data to build an end-to-end neural network prediction model to output the precipitation intensity of each grid point. Its advantages have been verified by common evaluation indicators such as RMSE and CSI.

[0179] ConvLSTM (Shi et al., 2015) combines the characteristics of convolutional neural networks and long short-term memory networks. It uses convolutional networks to extract features from radar echo images of consecutive times, and then enters the obtained time series into a temporal network with a gated structure. The forecast effect of this model on short-term precipitation is improved on FC-LSTM. PredRNN (Wang et al., 2022) also builds a model based on the principle of combining convolutional networks and temporal networks in time and space. The difference is that the memory units in the model are decoupled from each other, trained independently and then integrated, and the extraction of bidirectional "memory flow" in the sequence is added to obtain the dynamic changes of different scales of the image. Compared with ConvLSTM, the MSE of this model on the mobile version of MINST dataset is reduced by nearly half. Similarly, Franch et al. (2020) integrated RNN, convolutional network and downscaling network using TrajGRU (trajectory gated recurrent unit, Shi et al. 2017) structure, reducing the forecast error of the sub-model for extreme precipitation to half. MetNet (Espeholt et al. 2022) starts from the perspective of probabilistic forecasting, uses ConvLSTM to encode the time series of downsampled multi-channel images, then uses self-attention for decoding, and finally outputs the probability of precipitation. Compared with the numerical forecast model, MetNet has a lower precipitation forecast error within 7 to 8 hours in the continental United States. In addition, Chen et al. (2020) constructed a downscaled precipitation dataset RainNet, which provides downscaled annotated data for precipitation prediction models.

[0180] Among the many neural network models, the DGMR (deep generative model of radar, Ravuri et al., 2021) model proposed by DeepMind has made a more outstanding contribution. This model generates continuous precipitation intensity in space and time for the next 90 minutes based on GAN (generative adversarial network, Goodfellow et al., 2014), while being able to capture the characteristics of the convective system and characterize the uncertainty of the ensemble forecast. This model outperforms Unet (Ronneberger et al. 2015) in 89% of cases, but for extreme precipitation forecasts, this model still has unnatural phenomena such as erroneous movement of the convective system and dissipation over time. NowcastNet (Zhang et al., 2023) has made corrections to this problem, introducing physical constraints in the framework of deep learning to reduce errors. The model adds an evolution network to the main model to decode the movement and intensity of the convective system respectively, and then introduces a two-dimensional continuity equation as a constraint in this part. This improvement enabled Nowcast's forecast performance to surpass the optimal model at the time in 71% of cases, while still maintaining good accuracy in forecasting extreme precipitation.

[0181] It can be seen from the above embodiments that the present invention can monitor precipitation more accurately based on multi-source data fusion technology, thereby improving the accuracy and reliability of early warning. The present invention issues early warning information 2 hours in advance, which can give relevant mountain flood defense departments enough time to take countermeasures and reduce casualties and property losses. The present invention can monitor rainfall conditions in real time, avoid missing the center of heavy rainfall, and improve the timeliness of early warning. The present invention can perform zoning monitoring of the monitoring area, and carry out different early warning levels and countermeasures for different areas, thereby improving the pertinence of the early warning.

[0182] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with the technical field within the technical scope disclosed by the present invention and within the spirit and principles of the present invention should be covered within the protection scope of the present invention.

Claims

1. A method for fusion processing of flash flood disaster monitoring data sources based on deep learning, characterized in that: The method includes: S1, multi-source rainfall data collection: using rainfall radar and ground rain gauge to obtain rainfall data, and preprocessing and screening the obtained rainfall data; S2, multi-source data fusion: multiple monitoring data sources are fused to obtain high-precision precipitation monitoring results; S3, rainfall data analysis: according to the monitoring results, analyze the rainfall intensity and rainfall indicators; according to the analysis results, predict the probability and scope of future flash flood disasters; S4, Flash flood warning information release: Issue warning information in advance based on forecast results.

2. The method for fusion processing of flash flood disaster monitoring data sources based on deep learning according to claim 1 is characterized in that: In step S1, the acquired rainfall data is preprocessed and screened, including data cleaning and format conversion.

3. The method for fusion processing of flash flood disaster monitoring data sources based on deep learning according to claim 1 is characterized in that: In step S2, multiple monitoring data sources include: colorful cloud weather forecast data, short-term precipitation forecast data; The fusion processing includes: data integration, data alignment, and data fusion.

4. The method for fusion processing of flash flood disaster monitoring data sources based on deep learning according to claim 1 is characterized in that: In step S2, multiple monitoring data sources are fused and processed, including: S201, by means of auxiliary manual labeling, marking non-precipitation echoes in the historical radar image, and constructing a radar denoising training data set; inputting the radar images of the previous and next two frames of each sample in the radar denoising training data set into a deep neural network, performing denoising using the radar denoising model constructed by the deep neural network, and outputting the denoised radar image; S202, taking 2 frames of input and 25 frames of forecast radar images as one sample, constructing a radar forecast data set based on multiple sample sequences, and inputting the radar forecast data set into a precipitation forecast model to achieve precipitation estimation for a certain period in the future; S203, based on the radar denoising model, precipitation forecast model, satellite inversion, and Himawari8 and radar observation puzzles, a data set is created. The current and previous 30-minute observations are used as input, and the radar puzzle is used as output. Resnet and ConvNeXt are used to invert the radar echo intensity, and GAN is used to guide the training to make the inverted radar map clear in texture, supplement the radar real-time conditions in terrain occlusion, mountainous areas, and ocean areas, and obtain the inverted radar and the next 2-hour precipitation forecast model; S204, build a short-term precipitation forecast platform, update the obtained radar data and forecast data, and realize the pattern correction of precipitation, temperature, humidity, air pressure, wind speed and dew point temperature through the partial differential equation neural network model, realize the prediction of short-term precipitation forecast data 2 hours in advance, and issue early warning information.

5. The method for fusion processing of flash flood disaster monitoring data sources based on deep learning according to claim 4 is characterized in that: In step S201, a radar denoising training data set is constructed, including: the high-resolution network HRnet gradually joins the low-resolution feature map sub-network in parallel through the high-resolution feature map main network, multi-scale fusion and feature extraction are realized through different networks, and SEBlock is used to add an attention mechanism to the multi-scale feature fusion to construct a radar denoising model; The use of SEBlock to add an attention mechanism to multi-scale feature fusion includes: the channel attention mechanism submodule in the SEBlock attention mechanism module is G c , the spatial attention mechanism submodule is G s ; The channel attention mechanism submodule uses global average pooling and maximum pooling to highlight the feature information of multi-scale feature areas. The expression is: Where x() is the sigmoid nonlinear activation function, SAG( ) is the shared network in the SEBlock attention mechanism module, AvgPool( ) is the average pooling function, MaxPool( ) is the maximum pooling function, D is the feature information of the multi-scale feature region, E0, E1 are the hidden layer weights and output layer weights in the multi-layer perceptron in SAG, respectively, and E0∈H Clr×B ,E1∈H B×Clr , are the global average pooling features and the maximum pooling features respectively; H is the weight value, Clr is the number of hidden layers, and B is the number of output layers; Under the action of the channel attention mechanism, the spatial attention mechanism submodule is introduced, and the expression is: In the formula, e 7×7 is a 7×7 convolution operation, They are the global average pooling features and the maximum pooling features in space, F is the feature information of the multi-scale feature area in space, A is the width, and E is the height.

6. The method for fusion processing of flash flood disaster monitoring data sources based on deep learning according to claim 4 is characterized in that: In step S202, a radar forecast data set is constructed based on a plurality of sample sequences, and the radar forecast data set is input into a precipitation forecast model, including: Based on Unet and SEBlock, the deep rainfall generation model DGMR of DeepMind is integrated with the multimodal pre-training model overview. A discriminator is added to the DGMR model, and a loss of image authenticity is added to the loss function to build a precipitation forecast model. Unet integrated with SEBlock is integrated with DGMR to build a general network structure fusion model, and the nodes in the network are divided into two subsets Vl and Vunl according to whether there are forecast frame radar map labels in the radar forecast data set. Vl is the node set with forecast frame radar map labels, and Vunl is the node set without forecast frame radar map labels. The nodes in Vl are divided into k subsets according to their categories, where k represents the total number of node features in different frames, denoted as: Vl = {L1, L2…L k }, and record the label corresponding to the node v0∈Vl in Vl as The set of first-order neighbor nodes of node v0 with the forecast frame radar chart label is recorded as The network structure is fused by adding edges between nodes with forecast frame radar chart labels and belonging to the same frame features. o Calculate the number of edges added at one end The expression is: In the formula, For node v o The first-order neighboring nodes with the forecast frame radar chart label are followed by v o The number of nodes with the same label as v, γ is o The number of edges established when the first-order neighbors of have no forecast frame radar chart labels, v j is the jth node, is the label corresponding to the jth node, μ is v o First-order labeled neighbor nodes with v o The preset proportion of nodes with the same label; σ is the equivalence judgment function. If but Returns 1 if yes, otherwise returns 0.

7. The method for fusion processing of flash flood disaster monitoring data sources based on deep learning according to claim 4 is characterized in that: In step S204, a short-term precipitation forecast platform is constructed, and the acquired radar data and forecast data are updated, including: learning personalized abnormal parameters and guessed parameters from the feature vectors of the included radar data and forecast data, integrating the abnormal and guessed parameters, obtaining a new feature vector, and predicting the performance of the short-term precipitation forecast platform in a specific environment; First, learn personalized anomaly parameters and guess parameters from the feature vector, expressed as: q' t =tanh(I1Lt+b1) <h2 style=";text-align:left;direction:ltr">y'<h2 style=";text-align:left;direction:ltr"> t <h2 style=";text-align:left;direction:ltr"> (2L)<h2 style=";text-align:left;direction:ltr"> t <h2 style=";text-align:left;direction:ltr"> +b2) In the formula, q' t To learn personalized anomaly parameters from feature vectors, y' t To learn personalized guess parameters from feature vectors, I1 is the radar data parameter, I2 is the forecast data parameter, and L t is the label vector at time t, b1 is the feature constant vector in radar data learning, and b2 is the feature constant vector in forecast data learning; Secondly, the personalized abnormal parameters q' and guessed parameters y' based on the short-term precipitation forecast samples to be predicted and the abnormal parameters q and guessed parameters y based on the cognitive diagnosis model are integrated to obtain a new feature vector: Where L' t ' is the label vector at time t after the update, F' t ' is the feature information of the multi-scale feature region in the space at time t after the update, L' t To integrate the label vector at time t based on the short-term precipitation forecast sample to be predicted, F' t In order to integrate the characteristic information of multi-scale characteristic regions in space at time t based on the short-term precipitation forecast samples to be predicted, To incorporate the operator, q' t To learn personalized anomaly parameters from feature vectors, y' t To learn personalized guess parameters from feature vectors; Finally, the new feature vector obtained is activated through the sigmoid function to predict the performance of the short-term precipitation forecast platform in a specific environment; the performance of the short-term precipitation forecast platform V t The formula is as follows: Where V t is the performance value of the short-term precipitation forecast platform, β is the performance coefficient of the short-term precipitation forecast platform, I3 is the short-term precipitation forecast parameter, T is the order, and b3 is the constant value in the performance of the short-term precipitation forecast platform.

8. The method for fusion processing of flash flood disaster monitoring data sources based on deep learning according to claim 4 is characterized in that: In step S204, the pattern correction of precipitation, temperature, humidity, air pressure, wind speed and dew point temperature is realized by using the partial differential equation neural network model, including: Regional model correction: A radar mosaic composed of multiple radars, hourly ground observation data from multiple control stations, and regional numerical model forecast data are used to construct a numerical model forecast correction dataset, and a deep learning model is used to implement post-model correction of precipitation, 2m temperature, humidity, air pressure, wind speed, and dew point temperature; Regional model forecast: real-time precipitation regression, 500pha geopotential height extrapolation, and use of -3-hour and 0-hour regional model upper-air variables to forecast precipitation for the next 12 hours, and learn the movement trend of precipitation; The error of precipitation forecast data is controlled by using the partial differential equation neural network model PINNs with neural network automatic differentiation.

9. The method for fusion processing of flash flood disaster monitoring data sources based on deep learning according to claim 8 is characterized in that: Extreme precipitation forecasts are also provided through NowcastNet.

10. A flash flood disaster monitoring data source fusion processing system based on deep learning, characterized in that: The system implements the deep learning-based flash flood disaster monitoring data source fusion processing method according to any one of claims 1 to 9, and the system comprises: Multi-source rainfall data collection module, used to obtain rainfall data using rainfall radar and ground rain gauge, and to pre-process and filter the acquired rainfall data; The multi-source data fusion module is used to fuse multiple monitoring data sources to obtain high-precision precipitation monitoring results; The rainfall data analysis module is used to analyze rainfall intensity and rainfall indicators based on monitoring results; and to predict the probability and scope of future flash flood disasters based on the analysis results; The flash flood warning information release module is used to issue warning information in advance based on the prediction results.

Images