Rainfall prediction method and rainstorm monitoring and early warning device
Through the improved MIM model and channel attention unit, a radar echo and precipitation inversion model for precipitation prediction is constructed, which solves the limitations of recurrent neural networks in spatiotemporal prediction and the difficulty in determining parameters and low accuracy in precipitation prediction, achieving higher accuracy prediction.
Patent Information
- Application Number
- CN202411645606.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-18
- Publication Date
- 2025-03-25
AI Technical Summary
Existing recurrent neural networks for spatiotemporal prediction cannot learn too complex changes, and the parameters in precipitation forecasts are difficult to determine and have low accuracy.
The improved MIM model is adopted, combined with the channel attention unit, a radar echo prediction model and precipitation inversion model are constructed. By normalizing and downsampling the radar echo data, the automatic learning fitting of machine learning is used for prediction.
It improves the accuracy of radar echo prediction, can better learn spatial correlation and time dependence, solves the problems of difficult parameters and low accuracy, and provides important support for meteorological forecasting.
Smart Images
Figure CN119689610A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the field of meteorological forecasting, and in particular to a precipitation forecasting method and a rainstorm monitoring and early warning device. Background Art
[0002] Traditional extrapolation prediction methods based on radar echoes include linear extrapolation, cross-correlation and single-cell centroid methods. Since these methods are all translated at a constant speed, the extrapolation error will continue to accumulate and increase. At the same time, it is easy to lose track of strong convective cells and fail to predict the formation, dissolution and merging of convective cells. Based on the machine learning neural network algorithm, the optimal prediction model generated after training with historical data can not only learn the time evolution characteristics of the radar, but also perform nonlinear extrapolation to identify the formation, dissolution and merging trends of single cells.
[0003] Natural spatiotemporal processes are very unstable in many aspects, such as low-level non-stationarity, spatial correlation, temporal dependence of local pixel values, and high-level changes such as accumulation, deformation or dissipation of radar echoes. However, the relatively simple state transition functions in most recurrent neural networks (a type of machine learning neural network) used for spatiotemporal prediction do not support them to learn too complex changes in space and time.
[0004] At present, precipitation forecasts are usually fitted based on the radar echo extrapolation results using the ZR relationship, which has the problems of difficult parameter determination and low accuracy. Summary of the invention
[0005] The present invention provides a precipitation prediction method and a rainstorm monitoring and early warning device, which are used to solve the problems that the existing recursive neural network used for spatiotemporal prediction cannot learn too complex changes and the parameters existing in the existing precipitation forecast are difficult to determine and have low accuracy.
[0006] In a first aspect, the present invention provides a precipitation prediction method, comprising:
[0007] Normalizing and downsampling the radar echo data within T1 hours before the current moment to obtain first prediction input data, wherein the downsampling is achieved by maximum pooling and average pooling;
[0008] Inputting the first prediction input data into a trained radar echo prediction model to obtain radar echo prediction data within the next T2 hours, wherein the radar echo prediction model includes a MIM model and a channel attention unit, and the channel attention unit is used to calculate an attention weight matrix according to a hidden state in the MIM model and fuse the attention weight matrix with the hidden state;
[0009] Splicing the radar echo data within T1 hours before the current moment and the radar echo prediction data within T2 hours in the future in chronological order as the second prediction input data, slidingly selecting data within T1 hours from the second prediction input data at intervals of a preset step length to input into the trained precipitation inversion model, and obtaining precipitation prediction data within T2 hours in the future, wherein the precipitation inversion model is the MIM model;
[0010] The precipitation forecast data within the next T2 hours is spatially interpolated and numerically amplified to obtain the precipitation forecast result within the next T2 hours.
[0011] In the above scheme, the MIM model is improved, and the improved radar echo prediction model with a channel attention unit is added. It can not only autonomously learn the spatial correlation and time dependence of local pixel values from the image, but also have a better prediction of the accumulation, deformation or dissipation of radar echoes, effectively improving the prediction accuracy; a precipitation inversion model is constructed based on the MIM model, and based on the radar echo data within the previous T1 hour of the current moment and the radar echo prediction data within the next T2 hours, the precipitation inversion model is used for cyclic prediction to obtain the precipitation prediction data within the next T2 hours. The fusion forecast is realized through the automatic learning fitting of machine learning, which solves the problems of difficult parameter determination and low precision in fitting using the ZR relationship, and provides important support and guarantee for the fields of meteorological forecasting, disaster prevention and mitigation.
[0012] Optionally, the training method of the radar echo prediction model includes:
[0013] Selecting a first number of radar echo data within T3 consecutive hours from the historical radar echo data, where T3 = T1 + T2;
[0014] The radar echo data within the first T1 hour of the continuous radar echo data within T3 hours is used as the first input data, and the radar echo data within the last T2 hours is used as the first label data;
[0015] Dividing the first number of radar echo data within T3 consecutive hours into a first training set and a first test set, performing the normalization process on all data in the first training set and the first test set, and then performing the downsampling process on all first input data in the first training set and the first test set;
[0016] The radar echo prediction model is trained multiple times based on the processed first training set, and the next training is performed on the radar echo prediction model after the last training;
[0017] Based on the radar echo data in each continuous T3 hours in the processed first test set, the first input data is input into the radar echo prediction model after each training for testing to obtain the first test data, and the first TS score, the first hit rate POD and the first false alarm rate FAR are calculated according to the first label data and the first test data;
[0018] An optimal radar echo prediction model is selected from radar echo prediction models obtained through multiple trainings based on the first TS score, the first hit rate POD and the first null alarm rate FAR.
[0019] Optionally, respectively calculating a first TS score, a first hit rate POD, and a first false alarm rate FAR for the first test data according to the first label data includes:
[0020] Calculating the TP, FN and FP of the corresponding radar echo data in the first test data according to each radar echo data in the first label data, where TP, FN and FP are respectively the number of correctly predicted echo data points, the number of incorrectly predicted echo data points and the number of incorrectly predicted non-echo data points, and each radar echo data contains a plurality of said data points;
[0021] Calculate the TP sum, FN sum and FP sum of all radar echo data in the first test data;
[0022] The first TS score, the first hit rate POD and the first false alarm rate FAR are calculated respectively according to the TP sum, the FN sum and the FP sum.
[0023] Optionally, the selecting an optimal radar echo prediction model from radar echo prediction models obtained through multiple trainings based on the first TS score, the first hit rate POD, and the first null alarm rate FAR includes:
[0024] Selecting the radar echo prediction model with the highest first TS score from the radar echo prediction models obtained through the multiple trainings as the optimal radar echo prediction model;
[0025] If the first TS scores are the same, selecting the radar echo prediction model with the highest first hit rate POD from the radar echo prediction models obtained through the multiple trainings as the optimal radar echo prediction model;
[0026] If the first hit rates POD are the same, the radar echo prediction model with the lowest first null alarm rate FAR is selected from the radar echo prediction models obtained through the multiple trainings as the optimal radar echo prediction model.
[0027] Optionally, the training method of the precipitation inversion model includes:
[0028] Selecting a second quantity of second input data from the historical radar echo data, wherein the second input data is radar echo data within a continuous period of T1 hours;
[0029] Selecting a second number of second label data from the historical accumulated precipitation real-time data, where the second label data is accumulated precipitation real-time data for the period to which the second input data belongs, and the second input data corresponds to the second label data one by one;
[0030] Dividing the second number of second input data and the corresponding second label data into a second training set and a second test set, performing the normalization process on all data in the second training set and the second test set, and then performing the downsampling process on all second input data in the second training set and the second test set;
[0031] The precipitation inversion model is trained multiple times based on the processed second training set, and the next training is performed on the precipitation inversion model after the last training;
[0032] Based on each second input data in the processed second test set, input it into the precipitation inversion model after each training for testing to obtain second test data, and calculate the second TS score, the second hit rate POD and the second false alarm rate FAR according to the corresponding second label data and the second test data;
[0033] An optimal precipitation inversion model is selected from precipitation inversion models obtained through multiple trainings based on the second TS score, the second hit rate POD and the second false alarm rate FAR.
[0034] The calculating the second TS score, the second hit rate POD and the second false alarm rate FAR respectively according to the corresponding second label data and the second test data includes:
[0035] Calculating the TP, FN and FP of the corresponding actual cumulative precipitation data in the second test data according to each actual cumulative precipitation data in the second label data, TP, FN and FP are respectively the number of correctly predicted cumulative precipitation data points, the number of incorrectly predicted cumulative precipitation data points and the number of incorrectly predicted non-cumulative precipitation data points, and each actual cumulative precipitation data contains a plurality of the data points;
[0036] Calculate the TP sum, FN sum and FP sum of all accumulated precipitation actual data in the second test data;
[0037] The second TS score, the second hit rate POD and the second false alarm rate FAR are calculated respectively according to the TP sum, the FN sum and the FP sum.
[0038] Optionally, the selecting the optimal precipitation inversion model from precipitation inversion models obtained through multiple trainings based on the second TS score, the second hit rate POD, and the second false alarm rate FAR includes:
[0039] Selecting the precipitation inversion model with the highest second TS score from the precipitation inversion models obtained through the multiple trainings as the optimal precipitation inversion model;
[0040] If the second TS scores are the same, selecting the precipitation inversion model with the highest second hit rate POD from the precipitation inversion models obtained through the multiple trainings as the optimal precipitation inversion model;
[0041] If the second hit rates POD are the same, the precipitation inversion model with the lowest second false alarm rate FAR is selected from the precipitation inversion models obtained through the multiple trainings as the optimal precipitation inversion model.
[0042] In a second aspect, the present invention provides a rainstorm monitoring and early warning device, including a data processing module, a radar prediction module, a radar display module, a precipitation prediction module, a result processing module, a precipitation display module and an early warning push module, wherein:
[0043] The data processing module is used to perform normalization and downsampling processing on the radar echo data within T1 hours before the current moment to obtain first prediction input data, and the downsampling processing is implemented by maximum pooling and average pooling;
[0044] The radar prediction module is used to input the first prediction input data into a trained radar echo prediction model to obtain radar echo prediction data within the next T2 hours, wherein the radar echo prediction model includes a MIM model and a channel attention unit, and the channel attention unit is used to calculate an attention weight matrix according to a hidden state in the MIM model and fuse the attention weight matrix with the hidden state;
[0045] The radar display module is used to display the radar echo prediction data within the next T2 hours at a township level;
[0046] The precipitation prediction module is used to splice the radar echo data within the previous T1 hour of the current moment and the radar echo prediction data within the future T2 hours in chronological order as the second prediction input data, and slide and select the data within the T1 hour from the second prediction input data at intervals of a preset step length to input the trained precipitation inversion model to obtain the precipitation prediction data within the future T2 hours, and the precipitation inversion model is the MIM model;
[0047] The result processing module is used to perform spatial interpolation and numerical amplification on the precipitation prediction data within the next T2 hours to obtain the precipitation prediction result within the next T2 hours;
[0048] The precipitation display module is used to display the precipitation forecast results within the next T2 hours at the township level;
[0049] The warning push module is used to determine whether the maximum value in the future T2 hour precipitation forecast result exceeds a preset threshold. If it exceeds, a warning push is performed at the township level granularity.
[0050] The above solution realizes artificial intelligence monitoring and early warning of rainstorm weather based on machine learning. The monitoring and early warning push are refined to the township level, avoiding multi-level response from the city level to the township level, improving the response speed, and ensuring the timeliness of the front-line response.
[0051] In a third aspect, the present invention provides a computing device, comprising:
[0052] A memory for storing program instructions;
[0053] The processor is used to call the program instructions stored in the memory and execute any of the above methods according to the obtained program.
[0054] In a fourth aspect, the present invention provides a computer-readable non-volatile storage medium, comprising computer-readable instructions, which enable a computer to execute any of the above methods when the computer reads and executes the computer-readable instructions. BRIEF DESCRIPTION OF THE DRAWINGS
[0055] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0056] Figure 1 A schematic diagram of a flow chart of a precipitation prediction method provided by an embodiment of the present invention;
[0057] Figure 2 A schematic diagram of a MIM model and a radar echo prediction model provided by an embodiment of the present invention, wherein:
[0058] A is the MIM model, and B is the radar echo prediction model;
[0059] Figure 3 A schematic diagram of the structure of a channel attention unit provided in an embodiment of the present invention;
[0060] Figure 4A schematic structural diagram of a rainstorm monitoring and early warning device provided in an embodiment of the present invention. DETAILED DESCRIPTION
[0061] In order to make the purpose, technical scheme and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0062] It should be noted that the terms "first", "second" etc. involved in the documents of the present application are used to distinguish similar objects, and need not be used to describe a specific order or sequential order. It should be understood that the data used in this way can be interchangeable in appropriate circumstances, so that the embodiments of the present application described herein can be implemented in a sequence other than the content illustrated or described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, the process, method, system, product or equipment comprising a series of steps or units need not be limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or equipment.
[0063] Figure 1 The present invention provides a flowchart of a precipitation prediction method, which includes:
[0064] S101, normalizing and down-sampling the radar echo data within T1 hours before the current moment to obtain first prediction input data.
[0065] In one example, the radar echo data is obtained from the China Meteorological Administration. Currently, the radar echo data of the China Meteorological Administration is sampled every 6 minutes, that is, the radar echo data within 1 hour includes 10 radar echo data.
[0066] It should be noted that the radar echo data is two-dimensional rectangular data, which can be regarded as a grayscale image and is treated as an image during the neural network processing.
[0067] In one example, the formula for the normalization process is as follows:
[0068]
[0069] Among them, X′ is the value of the variable after normalization, X is the initial value of the variable, and X min is the minimum value of the variable, X max is the maximum value of the variable.
[0070] Specifically, the downsampling process is implemented through maximum pooling and average pooling. The process of downsampling is as follows:
[0071] First, the first input data is downsampled to the maximum value through Max Pooling;
[0072] Secondly, use Average Pooling to further downsample the result of the maximum downsampling.
[0073] The above downsampling process can reduce the number of parameters of the feature map, improve the calculation speed, and increase the receptive field.
[0074] S102, inputting the first prediction input data into the trained radar echo prediction model to obtain radar echo prediction data within the next T2 hours.
[0075] Specifically, the radar echo prediction model includes a MIM model and a channel attention unit.
[0076] MIM model and radar echo prediction model Figure 2 As shown, A is the MIM model and B is the radar echo prediction model. The MIM model is a spatiotemporal neural network model based on deep learning, referred to as the spatiotemporal sequence model. The MIM model can capture the temporal dependency and spatial correlation between adjacent pixels by stacking multiple spatiotemporal memory modules (i.e., ST-LSTM). The main formula of MIM is as follows:
[0077]
[0078] Among them, H is the hidden state, W and b are weight parameters, σ is the sigmoid activation function, MIM-S and MIM-N represent the horizontal storage units in the stationary module and the non-stationary module respectively. is the spatiotemporal memory, l∈{1,…,L}, l and L are the layer sequence number and number of layers of the spatiotemporal memory module respectively, is a horizontal storage unit, is the differential feature learned by MIM-N and input to MIM-S, which is vertically transmitted from the previous layer to the current node at the same time step. For the memory that passes through the virtual "forget gate", i t is the input gate, f t ′ is the forget gate, o t is the output gate, g t is the input modulation gate, * is the convolution operation, and ⊙ is the Hadamard operation (corresponding to multiplication).
[0079] The structure of the channel attention unit is as follows Figure 3 As shown:
[0080] First, perform global maximum pooling and global average pooling on an input feature map F of size HxWxC in the spatial dimension to obtain two 1x1 x C feature maps;
[0081] Then, the results of global maximum pooling and global average pooling are respectively sent to a shared multi-layer perceptron (MLP) for learning, and two 1x1xC feature maps are obtained, wherein the number of neurons in the first layer of MLP is C / r (in the embodiment of the present invention, r=16), the activation function is Relu, the number of neurons in the second layer is C, and the characteristics of the channel dimension and the importance of each channel are learned based on MLP;
[0082] Finally, the results of the MLP output are added and then mapped by the Sigmoid activation function to obtain the channel attention weight matrix.
[0083] The radar echo prediction model is improved on the MIM model structure by adding a channel attention module (CAM). It can not only autonomously learn the spatial correlation and temporal dependence of local pixel values from the image, but also have a better prediction of the accumulation, deformation or dissipation of radar echoes, effectively improving the prediction accuracy.
[0084] Specifically, the channel attention unit is used to calculate the attention weight matrix according to the hidden state in the MIM model and fuse the attention weight matrix with the hidden state. The steps are as follows:
[0085] The channel attention unit first performs global maximum pooling and global average pooling according to the hidden state H in the MIM model to obtain two feature maps of different dimensions;
[0086] Then the two feature maps share a multi-layer perceptron network, first through a fully connected layer to reduce the feature dimension, and then through a fully connected layer to increase the feature dimension;
[0087] The two feature maps after feature dimensionality upgrade are stacked in the channel dimension, and then the weight of each channel in the feature map is normalized to between 0 and 1 through the Sigmoid activation function to obtain the attention weight M. C (Right now Figure 2 and Figure 3 Mc in);
[0088] Finally, the normalized attention weight is multiplied by the hidden state (i.e. Figure 2 M C *H), and then added to the hidden state (i.e. Figure 2 M C *H+H) completes fusion.
[0089] In the embodiment of the present invention, the fusion result is the radar echo prediction data within the next T2 hours.
[0090] In one example, a radar echo prediction model training method includes:
[0091] S201, selecting a first number of radar echo data within T3 consecutive hours from historical radar echo data.
[0092] Specifically, T3=T1+T2.
[0093] The historical radar echo data may be radar echo data from 2022 to 2024.
[0094] S202 , among the radar echo data within the continuous T3 hours, the radar echo data within the first T1 hour is used as the first input data, and the radar echo data within the last T2 hours is used as the first label data.
[0095] It should be noted that the radar echo data within each continuous T3 hours are divided into first input data and first label data.
[0096] S203, dividing the first number of radar echo data within T3 consecutive hours into a first training set and a first test set, normalizing all data in the first training set and the first test set, and then down-sampling all first input data in the first training set and the first test set.
[0097] The first quantity is determined according to actual conditions. When the first quantity is 9000, 8000 consecutive radar echo data within T3 hours can be selected from the data from 2022-2023 as the first training set, and 1000 consecutive radar echo data within T3 hours can be selected from the data from 2024 as the first test set.
[0098] S204, training the radar echo prediction model multiple times based on the processed first training set, and the next training is performed on the radar echo prediction model after the last training.
[0099] S205, based on the radar echo data within each continuous T3 hours in the processed first test set, input its first input data into the radar echo prediction model after each training for testing to obtain the first test data, and calculate the first TS score, the first hit rate POD and the first false alarm rate FAR according to its first label data and the first test data.
[0100] It should be noted that the first test data is actually a prediction of the first label data based on the first input data. Therefore, each radar echo data in the first test data is a prediction of the corresponding radar echo data in the first label data.
[0101] In one example, respectively calculating a first TS score, a first hit rate POD, and a first false alarm rate FAR for the first test data according to its first label data includes:
[0102] Calculating TP, FN and FP of the corresponding radar echo data in the first test data according to each radar echo data in the first label data, TP, FN and FP are respectively the number of correctly predicted echo data points, the number of incorrectly predicted echo data points and the number of incorrectly predicted non-echo data points, and each radar echo data includes multiple data points;
[0103] Calculate the TP sum, FN sum and FP sum of all radar echo data in the first test data;
[0104] The first TS score, the first hit rate POD and the first false alarm rate FAR are calculated respectively according to the TP sum, the FN sum and the FP sum.
[0105] A correctly predicted non-echo data point represents that the value of the data point in the first test data and the value of the corresponding data point in the first label data do not belong to the echo data value; an incorrectly predicted non-echo data point represents that the value of the data point in the first test data belongs to the echo data value but the value of the corresponding data point in the first label data does not belong to the echo data value; an incorrectly predicted echo data point represents that the value of the data point in the first test data does not belong to the echo data value but the value of the corresponding data point in the first label data belongs to the echo data value or the value of the data point in the first test data and the corresponding data point in the first label data both belong to the echo data value but the values of the two do not match; a correctly predicted echo data point represents that the value of the data point in the first test data and the value of the corresponding data point in the first label data both belong to the echo data value and the values of the two match.
[0106] Usually, the first verification level is used to determine whether the value of a data point belongs to an echo data value. The first verification level is a method used in meteorology to evaluate the accuracy of radar echo forecasts. If the value of a data point is within the first verification level, it belongs to an echo data value; if it is not within the first verification level, it does not belong to an echo data value.
[0107] The first verification level includes multiple first level intervals, and the value matching of the data point represents that the value of the data point is in the same first level interval.
[0108] In one example, the first level intervals are:
[0109] Greater than or equal to 10dbz and less than 20dbz;
[0110] Greater than or equal to 20dbz and less than 30dbz;
[0111] Greater than or equal to 30dbz and less than 40dbz;
[0112] Greater than or equal to 40dbz.
[0113] If the value of a data point is greater than or equal to 0 dbz and less than 10 dbz, it can be determined as not belonging to an echo data value.
[0114] The general calculation formulas for TS score, hit rate POD and miss rate FAR are as follows:
[0115]
[0116]
[0117] When calculating the first TS score, the first hit rate POD and the first null alarm rate FAR, in the general calculation formulas of the above TS score, hit rate POD and null alarm rate FAR: TP is the sum of TPs of all radar echo data in the first test data, FN is the sum of FNs of all radar echo data in the first test data, and FP is the sum of FPs of all radar echo data in the first test data.
[0118] For each radar echo prediction model after training, its first TS score, first hit rate POD and first false alarm rate FAR will be calculated.
[0119] S206: Selecting the optimal radar echo prediction model from radar echo prediction models obtained through multiple trainings based on the first TS score, the first hit rate POD, and the first false alarm rate FAR.
[0120] In one example, selecting the optimal radar echo prediction model from radar echo prediction models obtained through multiple trainings based on the first TS score, the first hit rate POD, and the first false alarm rate FAR includes:
[0121] Selecting the radar echo prediction model with the highest first TS score from the radar echo prediction models obtained through multiple trainings as the optimal radar echo prediction model;
[0122] If the first TS scores are the same, the radar echo prediction model with the highest first hit rate POD is selected from the radar echo prediction models obtained through multiple trainings as the optimal radar echo prediction model;
[0123] If the first hit rates POD are the same, the radar echo prediction model with the lowest first null alarm rate FAR is selected from the radar echo prediction models obtained through multiple trainings as the optimal radar echo prediction model.
[0124] S103, splicing the radar echo data within T1 hours before the current moment and the radar echo prediction data within T2 hours in the future in chronological order as the second prediction input data, slidingly selecting the data within T1 hours from the second prediction input data at intervals of a preset step size to input into the trained precipitation inversion model, and obtaining the precipitation prediction data within T2 hours in the future.
[0125] Specifically, the precipitation inversion model is a MIM model.
[0126] In one example, a method for training a precipitation inversion model includes:
[0127] S301, selecting a second quantity of second input data from historical radar echo data.
[0128] Specifically, the second input data is radar echo data within a continuous period of T1 hours.
[0129] S302: Select a second quantity of second label data from the historical accumulated precipitation actual data.
[0130] Specifically, the second label data is the actual accumulated precipitation data for the time period to which the second input data belongs, and the second input data corresponds one-to-one to the second label data.
[0131] It should be noted that when the second input data selected is the radar echo data from 10 o'clock to 11 o'clock, the corresponding second label data selected is the actual cumulative precipitation data at 11 o'clock, and the actual cumulative precipitation data at 11 o'clock is actually the accumulated precipitation from 10 o'clock to 11 o'clock; when the second input data selected is the radar echo data from 10:06 to 11:06, the corresponding second label data selected should be the actual cumulative precipitation data at 11:06, and the actual cumulative precipitation data at 11:06 is actually the accumulated precipitation from 10:06 to 11:06.
[0132] S303: Divide the second quantity of second input data and the corresponding second label data into a second training set and a second test set, normalize all data in the second training set and the second test set, and then downsample all second input data in the second training set and the second test set.
[0133] The second quantity is determined according to actual conditions. When the second quantity is 9000, 8000 second input data and second label data can be selected from the data from 2022-2023 as the second training set, and 1000 second input data and second label data can be selected from the data from 2024 as the second test set.
[0134] S304, training the precipitation inversion model multiple times based on the processed second training set, and the next training is performed on the precipitation inversion model after the last training.
[0135] S305, based on each second input data in the processed second test set, input it into the precipitation inversion model after each training for testing to obtain the second test data, and calculate the second TS score, the second hit rate POD and the second false alarm rate FAR according to the corresponding second label data and the second test data.
[0136] It should be noted that the second test data is actually a prediction of the second label data based on the second input data. Therefore, each accumulated precipitation actual data in the second test data is a prediction of the corresponding accumulated precipitation actual data in the second label data.
[0137] In one example, respectively calculating the second TS score, the second hit rate POD, and the second false alarm rate FAR according to the corresponding second label data and the second test data includes:
[0138] According to each actual cumulative precipitation data in the second label data, TP, FN and FP of the corresponding actual cumulative precipitation data in the second test data are calculated, where TP, FN and FP are the number of correctly predicted cumulative precipitation data points, the number of incorrectly predicted cumulative precipitation data points and the number of incorrectly predicted non-cumulative precipitation data points, respectively, and each actual cumulative precipitation data includes multiple data points;
[0139] Calculate the TP sum, FN sum and FP sum of all accumulated precipitation actual data in the second test data;
[0140] The second TS score, the second hit rate POD and the second false alarm rate FAR are calculated respectively according to the TP sum, FN sum and FP sum.
[0141] A correctly predicted non-cumulative precipitation data point means that the value of the data point in the second test data and the value of the corresponding data point in the second label data do not belong to the cumulative precipitation data value; an incorrectly predicted non-cumulative precipitation data point means that the value of the data point in the second test data belongs to the cumulative precipitation data value but the value of the corresponding data point in the second label data does not belong to the cumulative precipitation data value; an incorrectly predicted cumulative precipitation data point means that the value of the data point in the second test data does not belong to the cumulative precipitation data value but the value of the corresponding data point in the second label data belongs to the cumulative precipitation data value or the value of the data point in the second test data and the corresponding data point in the second label data both belong to the cumulative precipitation data value but the values of the two do not match; a correctly predicted cumulative precipitation data point means that the value of the data point in the second test data and the value of the corresponding data point in the second label data both belong to the cumulative precipitation data value and the values of the two match.
[0142] Usually, the second verification level is used to determine whether the value of a data point belongs to the cumulative precipitation data value. The second verification level is a method used in meteorology to verify the accuracy of precipitation. If the value of the data point is within the second verification level, it belongs to the cumulative precipitation data value; if it is not within the second verification level, it does not belong to the cumulative precipitation data value.
[0143] The second verification level includes multiple second level intervals, and the value matching of the data point represents that the value of the data point is in the same second level interval.
[0144] In one example, the second level intervals are:
[0145] Greater than or equal to 0.1mm and less than 5mm;
[0146] Greater than or equal to 5mm and less than 10mm;
[0147] Greater than or equal to 10mm and less than 20mm;
[0148] Greater than or equal to 20mm.
[0149] If the value of a data point is greater than or equal to 0 mm and less than 0.1 mm, it can be determined that it does not belong to the cumulative precipitation data value.
[0150] When calculating the second TS score, the second hit rate POD and the second false alarm rate FAR, in the general calculation formulas of the above TS score, hit rate POD and false alarm rate FAR: TP is the sum of TPs of all cumulative precipitation actual data in the second test data, FN is the sum of FNs of all cumulative precipitation actual data in the second test data, and FP is the sum of FPs of all cumulative precipitation actual data in the second test data.
[0151] S306: Selecting the optimal precipitation inversion model from precipitation inversion models obtained through multiple trainings based on the second TS score, the second hit rate POD and the second false alarm rate FAR.
[0152] In one example, selecting the optimal precipitation inversion model from precipitation inversion models obtained through multiple trainings based on the second TS score, the second hit rate POD, and the second false alarm rate FAR includes:
[0153] The precipitation inversion model with the highest second TS score is selected as the optimal precipitation inversion model from the precipitation inversion models obtained through multiple trainings;
[0154] If the second TS scores are the same, the precipitation inversion model with the highest second hit rate POD is selected from the precipitation inversion models obtained through multiple trainings as the optimal precipitation inversion model;
[0155] If the second hit rates POD are the same, the precipitation inversion model with the lowest second false alarm rate FAR is selected from the precipitation inversion models obtained through multiple trainings as the optimal precipitation inversion model.
[0156] S104, spatial interpolation and numerical amplification are performed on the precipitation forecast data within the next T2 hours to obtain the precipitation forecast result within the next T2 hours.
[0157] Numerical amplification is to amplify the data to the size of normal precipitation, which is achieved by amplifying it 50 times in the embodiment of the present invention.
[0158] In the above scheme, the MIM model is improved, and the improved radar echo prediction model with a channel attention unit is added. It can not only autonomously learn the spatial correlation and time dependence of local pixel values from the image, but also have a better prediction of the accumulation, deformation or dissipation of radar echoes, effectively improving the prediction accuracy; a precipitation inversion model is constructed based on the MIM model, and based on the radar echo data within the previous T1 hour of the current moment and the radar echo prediction data within the next T2 hours, the precipitation inversion model is used for cyclic prediction to obtain the precipitation prediction data within the next T2 hours. The fusion forecast is realized through the automatic learning fitting of machine learning, which solves the problems of difficult parameter determination and low precision in fitting using the ZR relationship, and provides important support and guarantee for the fields of meteorological forecasting, disaster prevention and mitigation.
[0159] Based on the same technical concept, Figure 4 The schematic diagram of the structure of a rainstorm monitoring and early warning device provided in this embodiment is shown, including a data processing module, a radar prediction module, a radar display module, a precipitation prediction module, a result processing module, a precipitation display module and an early warning push module, wherein:
[0160] The data processing module is used to perform normalization and down-sampling processing on the radar echo data within T1 hours before the current moment to obtain the first prediction input data. The down-sampling processing is implemented by maximum pooling and average pooling;
[0161] The radar prediction module is used to input the first prediction input data into the trained radar echo prediction model to obtain the radar echo prediction data within the next T2 hours. The radar echo prediction model includes a MIM model and a channel attention unit. The channel attention unit is used to calculate the attention weight matrix according to the hidden state in the MIM model and fuse the attention weight matrix with the hidden state.
[0162] The radar display module is used to display radar echo prediction data within the next T2 hours at the township level;
[0163] The precipitation prediction module is used to splice the radar echo data within the previous T1 hour of the current moment and the radar echo prediction data within the next T2 hours in chronological order as the second prediction input data, and slide and select the data within T1 hour from the second prediction input data at intervals of a preset step length to input the trained precipitation inversion model to obtain the precipitation prediction data within the next T2 hours, and the precipitation inversion model is a MIM model;
[0164] The result processing module is used to perform spatial interpolation and numerical amplification on the precipitation forecast data within the next T2 hours to obtain the precipitation forecast results within the next T2 hours;
[0165] The precipitation display module is used to display the precipitation forecast results within the next T2 hours at the township level;
[0166] The early warning push module is used to determine whether the maximum value in the precipitation forecast results within the next T2 hours exceeds the preset threshold. If it exceeds, an early warning is pushed at the township level.
[0167] For example, the preset threshold is set to 20mm. If the maximum value of the precipitation forecast result in a certain township within the next T2 hours is greater than 20mm, it is considered that heavy precipitation has occurred and a red alert will be issued for the township.
[0168] The above solution realizes artificial intelligence monitoring and early warning of rainstorm weather based on machine learning. The monitoring and early warning push are refined to the township level, avoiding multi-level response from the city level to the township level, improving the response speed, and ensuring the timeliness of the front-line response.
[0169] In one example, T1 hour is 1 hour, T2 hours is 2 hours, and T3 hours is 3 hours, which can ensure a warning window period of 0-3 hours.
[0170] In one example, the preset step size is set to the update interval of the radar echo data in the National Meteorological Administration, which increases the update frequency of the precipitation forecast results and achieves an update every 6 minutes.
[0171] Based on the same technical concept, an embodiment of the present invention provides a computing device, including:
[0172] A memory for storing program instructions;
[0173] The processor is used to call the program instructions stored in the memory and execute the above method according to the obtained program.
[0174] Based on the same technical concept, an embodiment of the present invention provides a computer-readable non-volatile storage medium, including computer-readable instructions. When a computer reads and executes the computer-readable instructions, the computer executes the above method.
[0175] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.
[0176] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.
[0177] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.
[0178] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.
[0179] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.
Claims
1. A precipitation prediction method, characterized in that: include: Normalizing and downsampling the radar echo data within T1 hours before the current moment to obtain first prediction input data, wherein the downsampling is achieved by maximum pooling and average pooling; Inputting the first prediction input data into a trained radar echo prediction model to obtain radar echo prediction data within the next T2 hours, wherein the radar echo prediction model includes a MIM model and a channel attention unit, and the channel attention unit is used to calculate an attention weight matrix according to a hidden state in the MIM model and fuse the attention weight matrix with the hidden state; Splicing the radar echo data within T1 hours before the current moment and the radar echo prediction data within T2 hours in the future in chronological order as the second prediction input data, slidingly selecting data within T1 hours from the second prediction input data at intervals of a preset step length to input into the trained precipitation inversion model, and obtaining precipitation prediction data within T2 hours in the future, wherein the precipitation inversion model is the MIM model; The precipitation forecast data within the next T2 hours is spatially interpolated and numerically amplified to obtain the precipitation forecast result within the next T2 hours.
2. The method according to claim 1, characterized in that: The training method of the radar echo prediction model includes: Selecting a first number of radar echo data within T3 consecutive hours from the historical radar echo data, where T3 = T1 + T2; The radar echo data within the first T1 hour of the continuous radar echo data within T3 hours is used as the first input data, and the radar echo data within the last T2 hours is used as the first label data; Dividing the first number of radar echo data within T3 consecutive hours into a first training set and a first test set, performing the normalization process on all data in the first training set and the first test set, and then performing the downsampling process on all first input data in the first training set and the first test set; The radar echo prediction model is trained multiple times based on the processed first training set, and the next training is performed on the radar echo prediction model after the last training; Based on the radar echo data in each continuous T3 hours in the processed first test set, the first input data is input into the radar echo prediction model after each training for testing to obtain the first test data, and the first TS score, the first hit rate POD and the first false alarm rate FAR are calculated according to the first label data and the first test data; An optimal radar echo prediction model is selected from radar echo prediction models obtained through multiple trainings based on the first TS score, the first hit rate POD and the first null alarm rate FAR.
3. The method according to claim 2, characterized in that: The calculating of the first TS score, the first hit rate POD and the first false alarm rate FAR for the first test data according to the first label data includes: Calculating the TP, FN and FP of the corresponding radar echo data in the first test data according to each radar echo data in the first label data, where TP, FN and FP are respectively the number of correctly predicted echo data points, the number of incorrectly predicted echo data points and the number of incorrectly predicted non-echo data points, and each radar echo data contains a plurality of said data points; Calculate the TP sum, FN sum and FP sum of all radar echo data in the first test data; The first TS score, the first hit rate POD and the first false alarm rate FAR are calculated respectively according to the TP sum, the FN sum and the FP sum.
4. The method according to claim 2, characterized in that: The selecting the optimal radar echo prediction model from radar echo prediction models obtained through multiple trainings based on the first TS score, the first hit rate POD, and the first null alarm rate FAR includes: Selecting the radar echo prediction model with the highest first TS score from the radar echo prediction models obtained through the multiple trainings as the optimal radar echo prediction model; If the first TS scores are the same, selecting the radar echo prediction model with the highest first hit rate POD from the radar echo prediction models obtained through the multiple trainings as the optimal radar echo prediction model; If the first hit rates POD are the same, the radar echo prediction model with the lowest first null alarm rate FAR is selected from the radar echo prediction models obtained through the multiple trainings as the optimal radar echo prediction model.
5. The method according to claim 1, characterized in that: The training method of the precipitation inversion model includes: Selecting a second quantity of second input data from the historical radar echo data, wherein the second input data is radar echo data within a continuous period of T1 hours; Selecting a second number of second label data from the historical accumulated precipitation real-time data, where the second label data is accumulated precipitation real-time data for the period to which the second input data belongs, and the second input data corresponds to the second label data one by one; Dividing the second number of second input data and the corresponding second label data into a second training set and a second test set, performing the normalization process on all data in the second training set and the second test set, and then performing the downsampling process on all second input data in the second training set and the second test set; The precipitation inversion model is trained multiple times based on the processed second training set, and the next training is performed on the precipitation inversion model after the last training; Based on each second input data in the processed second test set, input it into the precipitation inversion model after each training for testing to obtain second test data, and calculate the second TS score, the second hit rate POD and the second false alarm rate FAR according to the corresponding second label data and the second test data; An optimal precipitation inversion model is selected from precipitation inversion models obtained through multiple trainings based on the second TS score, the second hit rate POD and the second false alarm rate FAR.
6. The method according to claim 5, characterized in that: The calculating the second TS score, the second hit rate POD and the second false alarm rate FAR respectively according to the corresponding second label data and the second test data includes: Calculating the TP, FN and FP of the corresponding actual cumulative precipitation data in the second test data according to each actual cumulative precipitation data in the second label data, TP, FN and FP are respectively the number of correctly predicted cumulative precipitation data points, the number of incorrectly predicted cumulative precipitation data points and the number of incorrectly predicted non-cumulative precipitation data points, and each actual cumulative precipitation data contains a plurality of the data points; Calculate the TP sum, FN sum and FP sum of all accumulated precipitation actual data in the second test data; The second TS score, the second hit rate POD and the second false alarm rate FAR are calculated respectively according to the TP sum, the FN sum and the FP sum.
7. The method according to claim 5, characterized in that: The selecting the optimal precipitation inversion model from precipitation inversion models obtained through multiple trainings based on the second TS score, the second hit rate POD, and the second false alarm rate FAR includes: Selecting the precipitation inversion model with the highest second TS score from the precipitation inversion models obtained through the multiple trainings as the optimal precipitation inversion model; If the second TS scores are the same, selecting the precipitation inversion model with the highest second hit rate POD from the precipitation inversion models obtained through the multiple trainings as the optimal precipitation inversion model; If the second hit rates POD are the same, the precipitation inversion model with the lowest second false alarm rate FAR is selected from the precipitation inversion models obtained through the multiple trainings as the optimal precipitation inversion model.
8. A rainstorm monitoring and early warning device, characterized in that: It includes data processing module, radar prediction module, radar display module, precipitation prediction module, result processing module, precipitation display module and warning push module, among which: The data processing module is used to perform normalization and downsampling processing on the radar echo data within T1 hours before the current moment to obtain first prediction input data, and the downsampling processing is implemented by maximum pooling and average pooling; The radar prediction module is used to input the first prediction input data into a trained radar echo prediction model to obtain radar echo prediction data within the next T2 hours, wherein the radar echo prediction model includes a MIM model and a channel attention unit, and the channel attention unit is used to calculate an attention weight matrix according to a hidden state in the MIM model and fuse the attention weight matrix with the hidden state; The radar display module is used to display the radar echo prediction data within the next T2 hours at a township level; The precipitation prediction module is used to splice the radar echo data within the previous T1 hour of the current moment and the radar echo prediction data within the future T2 hours in chronological order as the second prediction input data, and slide and select the data within the T1 hour from the second prediction input data at intervals of a preset step length to input the trained precipitation inversion model to obtain the precipitation prediction data within the future T2 hours, and the precipitation inversion model is the MIM model; The result processing module is used to perform spatial interpolation and numerical amplification on the precipitation prediction data within the next T2 hours to obtain the precipitation prediction result within the next T2 hours; The precipitation display module is used to display the precipitation forecast results within the next T2 hours at the township level; The warning push module is used to determine whether the maximum value in the future T2 hour precipitation forecast result exceeds a preset threshold. If it exceeds, a warning push is performed at the township level granularity.
9. A computing device, characterized in that include: A memory for storing program instructions; A processor is used to call the program instructions stored in the memory and execute the method according to any one of claims 1 to 7 according to the obtained program.
10. A computer-readable non-volatile storage medium, characterized in that: The method comprises computer-readable instructions, and when a computer reads and executes the computer-readable instructions, the computer executes the method according to any one of claims 1 to 7.