Short-term precipitation prediction method based on multimodal data and up- and down-sampling GRU network

Abstract

The invention discloses a short-term precipitation prediction method based on multi-level GRU network of multimodal data and up-down sampling, belonging to the technical field of weather forecasting, and comprising the following steps: firstly, inputting historical radar echo map and gridded atmospheric data at time t, removing data noise, and filling missing data; then, convolving and deconvolving the spatiotemporal scales in the gridded atmospheric data that do not match the radar echo map, so as to achieve scale correspondence; then, respectively inputting the radar echo map data and the atmospheric data into the network for encoding, so as to obtain a feature map; finally, splicing the obtained two feature maps, inputting them into the network for decoding, so as to obtain the future radar echo image, and then obtaining the future regional precipitation prediction; the invention can improve the precision of precipitation prediction on multiple scales and integrate the image features of multiple meteorological data.

Description

Technical Field

[0001] The invention belongs to the technical field of weather forecasting, and in particular relates to a short-term precipitation prediction method based on multimodal data and a cyclic gated unit network with up and down sampling. Background Art

[0002] Changes in meteorological factors (such as wind speed, temperature, humidity, precipitation, etc.) have a profound impact on human life. Accurately predicting future meteorological factors can widely serve people's daily life, transportation, agriculture, forestry, animal husbandry, disaster weather and other fields. With the increase in the number of earth observation satellites and the increasing power of climate models, meteorological researchers are faced with larger-scale data.

[0003] At present, numerical weather forecasting and artificial intelligence forecasting based on numerical weather forecast data are the main methods of weather forecasting. For numerical weather forecasting methods, short-term forecasts require complex physical atmospheric model simulations. In recent years, machine learning and deep learning have begun to be applied to weather forecasting.

[0004] However, the current traditional methods mainly use 2D-CNN or 3D-CNN, and the prediction accuracy (mainly measured by the CSI index) is not high; therefore, a new solution is urgently needed to solve the above problems. Summary of the invention

[0005] The present invention aims at the defects in the prior art and provides a short-term precipitation prediction method based on multimodal data and up- and down-sampling recurrent gated unit networks. This scheme is not only conducive to model training and improving the accuracy of short-term precipitation prediction, especially the prediction accuracy of heavy rain, but also can solve the technical problems existing in the prior art, such as uneven precipitation data, low heavy rain prediction accuracy, and fewer model fusion features.

[0006] In order to achieve the above-mentioned invention object, the present invention adopts the following technical scheme: a short-term precipitation prediction method based on multimodal data and up- and down-sampling recurrent gated unit network, the method comprising the following steps:

[0007] S1: Input data and data cleaning. First, input the historical radar echo map and gridded atmospheric data at time t, supplement the input data with default value linear interpolation, and denoise it by wavelet transform to obtain the processed input data. Then, the gridded atmospheric data that does not match time and space is first convolved by 1*1 to adjust the number of feature channels, and then the scale correspondence is achieved through deconvolution. Next, the radar echo map data and gridded atmospheric data are respectively input into the GRU network containing two layers of downsampling for encoding, and the two feature maps obtained are spliced ​​to obtain the fused feature map. Finally, the fused feature map is input into the GRU network containing two layers of upsampling for decoding to obtain the predicted radar echo image. The future regional precipitation forecast is obtained through ZR transformation, and the short-term precipitation forecast result Y is output.t ,Y t+1 ,...,Y t+p , where Y t+q It represents the result of converting the predicted radar echo map at time t+q into a precipitation map, 1≤q≤p, and p represents the total number of short-term precipitation prediction moments.

[0008] The denoising process of the historical radar echo map sequence at the input time t includes the following steps:

[0009] First, use bilinear interpolation to fill the default data. The interpolation formula is as follows:

[0010]

[0011] Where Q ij is the position (x i ,y j ), i,j∈{1,2}, (x,y) is the position of the default value in the matrix.

[0012] Then, the atmospheric grid data is denoised using hard threshold wavelet transform.

[0013] S2: Data processing, the convolution and deconvolution calculations used include the following steps:

[0014] Use 1*1 convolution kernel to adjust the number of feature channels. The calculation formula is:

[0015]

[0016] Where i represents the number of target feature channels mapped by the convolution kernel, and c represents the number of feature channels of the input data.

[0017] The new feature map obtained in the previous step is resized through deconvolution operation so that the scale of the feature map is consistent with the scale of the radar echo image. The calculation formula is as follows:

[0018] O=K*I

[0019] Where * is the matrix convolution operation, I is the matrix obtained in S21, and K is the deconvolution kernel. The parameters s, k, and p of the deconvolution kernel K are set in accordance with the following formula:

[0020] i=(o-1)*s+k-2p

[0021] Where i represents the width of the meteorological data matrix, o represents the width of the radar echo map matrix, s represents the convolution kernel moving step, k represents the width of the convolution kernel, and p represents the width of the filling 0 value.

[0022] S3: Encoding feature maps and feature fusion, input radar echo image data and gridded atmospheric data into the encoder network containing two layers of down-sampling recurrent gating units for encoding, and concatenate the two obtained feature maps. The encoder network consists of a convolutional layer, a recurrent gating unit layer, a down-sampling layer, a recurrent gating unit layer, a down-sampling layer, and a recurrent gating unit layer. The loss function selects the mean square error loss function, and the optimizer selects the Adam optimizer. The radar echo image code output by the encoder network and the gridded atmospheric data code matrix are concatenated on the feature channel.

[0023] S4: Generate prediction and output precipitation forecast, input the fused feature map into the decoder network containing two layers of upsampling GRU for decoding, obtain the predicted radar echo image, obtain the future regional precipitation forecast through ZR transformation, and output the short-term precipitation forecast result Y t ,Y t+1 ,...,Y t+p , where Y t+q It represents the result after the predicted radar echo map at time t+q is converted into a precipitation map, 1≤q≤p, and p represents the total number of short-term precipitation prediction moments. The decoder network consists of a recurrent gating unit, a deconvolution upsampling, a recurrent gating unit, a deconvolution upsampling, a recurrent gating unit, and a deconvolution upsampling.

[0024] Compared with the prior art, the beneficial effects of the invention are as follows: the short-term precipitation prediction method based on multimodal data and up- and down-sampling GRU network of the present invention is a method based on the fusion of multiple meteorological features at different scales.

[0025] 1. It helps the model learn abstract geographic location information during training, and improves the accuracy of short-term precipitation prediction, especially the prediction accuracy of heavy rain;

[0026] 2. The forecast accuracy can be improved at different scales and can be adapted to regional forecasts in practical applications of weather forecasting.

[0027] 3. This method reasonably uses a variety of convolution kernel operations to reduce network parameters and speed up model operation.

[0028] 4. This method integrates multiple modal data, ultimately improving the accuracy of model prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 is a schematic diagram of step S12;

[0030] Figure 2 is a schematic diagram of step S34;

[0031] Figure 3The figure is a flowchart of the short-term precipitation prediction method based on multimodal data and up- and down-sampling GRU network. DETAILED DESCRIPTION

[0032] The present invention is further described below in conjunction with the accompanying drawings and specific implementation examples.

[0033] Example 1: See Figure 1 — Figure 3 , a short-term precipitation prediction method based on multimodal data and up- and down-sampling GRU network, comprising the following steps:

[0034] S1: Input the historical radar echo map and gridded atmospheric data at time t, perform default value linear interpolation and wavelet transform denoising on the input data, and obtain the processed input data;

[0035] S2: For the gridded atmospheric data with mismatched time and space, 1*1 convolution is first performed to adjust the number of feature channels, and then deconvolution is used to achieve scale correspondence;

[0036] S3: The radar echo map data and gridded atmospheric data are respectively input into the GRU network containing two layers of downsampling for encoding, and the two obtained feature maps are spliced ​​to obtain a fused feature map

[0037] S4: Input the fused feature map into the GRU network containing two layers of upsampling for decoding, obtain the predicted radar echo image, obtain the future regional precipitation forecast through ZR transformation, and output the short-term precipitation forecast result Y t ,Y t+1 ,...,Y t+p , where Y t+q It represents the result of converting the predicted radar echo map at time t+q into a precipitation map, 1≤q≤p, and p represents the total number of short-term precipitation prediction moments.

[0038] In step S1, the denoising process of the historical radar echogram sequence at the input time t includes the following steps:

[0039] S11: Use bilinear interpolation to fill the default data. The interpolation formula is as follows:

[0040]

[0041] Where Q ij is the position (x i ,y j ), i,j∈{1,2}, (x,y) is the position of the default value in the matrix.

[0042] S12: Denoising the atmospheric grid data using hard threshold wavelet transform;

[0043] In step S2, the convolution and deconvolution calculations used include the following steps:

[0044] S21: Use 1*1 convolution kernel to adjust the number of feature channels. The calculation formula is:

[0045]

[0046] Where i represents the number of target feature channels mapped by the convolution kernel, and c represents the number of feature channels of the input data.

[0047] S22: The new feature map obtained in the previous step is adjusted in size through deconvolution operation so that the scale of the feature map is consistent with the scale of the radar echo image. The calculation formula is as follows:

[0048] O=K*I

[0049] Where * is the matrix convolution operation, I is the matrix obtained in S21, and K is the deconvolution kernel. The parameters s, k, and p of the deconvolution kernel K are set in accordance with the following formula:

[0050] i=(o-1)*s+k-2p

[0051] Where i represents the width of the meteorological data matrix, o represents the width of the radar echo map matrix, s represents the convolution kernel moving step, k represents the width of the convolution kernel, and p represents the width of the filling 0 value.

[0052] In the step S3, the radar echo image data and the gridded atmospheric data are respectively input into an encoder network including two layers of down-sampling cyclic gating units for encoding, and the two obtained feature maps are spliced. The encoder network is composed of a convolutional layer, a cyclic gating unit layer, a down-sampling layer, a cyclic gating unit layer, a down-sampling layer, and a cyclic gating unit layer. The loss function selects a mean square error loss function, and the optimizer selects an Adam optimizer. The radar echo image code output by the encoder network and the gridded atmospheric data code matrix are spliced ​​on the feature channel.

[0053] In the step S4, the fused feature map is input into the decoder network of the GRU containing two layers of upsampling for decoding to obtain the predicted radar echo image, and the future regional precipitation prediction is obtained through ZR transformation, and the short-term precipitation prediction result Y is output. t ,Y t+1 ,...,Y t+p , where Y t+q It represents the result after the predicted radar echo map at time t+q is converted into a precipitation map, 1≤q≤p, and p represents the total number of short-term precipitation prediction moments. The decoder network consists of a recurrent gating unit, a deconvolution upsampling, a recurrent gating unit, a deconvolution upsampling, a recurrent gating unit, and a deconvolution upsampling.

[0054] The following takes a specific embodiment as an example to further illustrate this specific implementation method. Specific embodiment 1

[0056] The validation dataset for this method is the radar echo map, gridded temperature, and total precipitation provided by the Guangdong Meteorological Bureau. The time span is from March 2017 to December 2018. The resolution is 1 km, and the matrix size is 300*300*1. The data interval is 12 minutes. The ZR relationship represents the relationship between reflectivity Z and precipitation intensity R (mm / h), where dBZ = 10log 10 a+10blog 10 R, a, b are the radar parameters. In this experiment, the values ​​are a = 58.53, b = 1.56. dBZ is often used to describe precipitation. Generally, the larger the value, the greater the precipitation.

[0057] The gridded meteorological data (including humidity at 9 altitudes, wind speed at 12 altitudes and 2 rainfall intensities, a total of 13 characteristic channels) are provided by GRAPES for South China, with a matrix size of 100*100*13, a resolution of 3 km and 1 hour.

[0058] Combined with experimental experience, this experiment uses the radar echo map of the first 10 moments and the gridded temperature total precipitation to predict the radar echo map of the next 10 moments.

[0059] The precipitation prediction evaluation index in the meteorological field is the CSI score:

[0060]

[0061] Where TP is the number of correctly predicted grid points, FN is the number of missed grid points, and FP is the number of falsely reported grid points. This experiment aims to improve the CSI score of the grid sequence for precipitation prediction.

[0062] This method first inputs the historical radar echo map at time t and the gridded meteorological data (including humidity at 9 altitudes, wind speed at 12 altitudes and 2 rainfall intensities, a total of 13 feature channels), and performs missing value supplementation and denoising; then, in order to solve the problem of different temporal and spatial resolutions, the 1*1 convolution kernel is set to 1*1*1, and the parameters s, k, and p of the deconvolution kernel are set to:

[0063] s=3

[0064] k = 3

[0065] p=0

[0066] The gridded meteorological data can be mapped to the same scale as the radar echo map data; then the two feature maps are respectively input into the GRU network containing two layers of downsampling for encoding, and the two feature maps are spliced ​​to obtain a fused feature map. Finally, the fused feature map is input into the GRU network containing two layers of upsampling for decoding to obtain the predicted radar echo image, and the future regional precipitation forecast is obtained through ZR transformation, and the short-term precipitation forecast result is output.

[0067] Table 1 below is a model parameter table of the GRU network including up and down sampling. Combined with experimental experience, this experiment uses the radar echo map of the first 10 moments and gridded meteorological data to predict the radar echo map of the next 10 moments.

[0068] Table 1 contains the model parameters of the GRU network with up and down sampling

[0069]

[0070] Note: 1) The names starting with "e" are encoder structures, and the names starting with "f" are predictor structures;

[0071] 2) Depth I / O refers to the input depth to the output depth;

[0072] The r value is the rainfall intensity level obtained after Z-RC transformation, and its corresponding relationship with the precipitation report is as follows

[0073] Table 2 The corresponding relationship between r value and precipitation report is as follows

[0074] R value gear Precipitation forecast r≥0.5 Light rain r≥2 light rain r≥5 moderate rain r≥10 heavy rain r≥30 rainstorm

[0075] Traditional methods usually use 2DCNN or 3DCNN. Table 3 below shows the average scores of several models for the CSI index of the experimental data at different levels. 2D-CNN is a model that uses a two-dimensional convolutional neural network for time series images, and 3D CNN is a model that uses 3D-CNN to represent a three-dimensional convolutional neural network for time series images. The higher the CSI index, the higher the precipitation prediction accuracy. As can be seen from the table below, this method has a significant improvement in prediction scores compared to traditional methods.

[0076] Table 3 Comparison of prediction scores of the model implemented in the present invention with other methods

[0077]

[0078] It should be noted that the above embodiments are merely preferred embodiments of the present invention and are not intended to limit the protection scope of the present invention. Equivalent replacements or substitutions made on the basis of the above technical solutions all fall within the protection scope of the present invention.

Claims

1. A short-term precipitation prediction method based on multimodal data and up- and down-sampling GRU network, characterized by: The following steps are involved: S1: Input data and data cleaning: input the historical radar echo map and gridded atmospheric data at time t, perform default value linear interpolation and wavelet transform denoising on the input data, and obtain the processed input data; S2: Data processing: For the gridded atmospheric data that does not match time and space, 1*1 convolution is first performed to adjust the number of feature channels, and then deconvolution is used to achieve scale correspondence; S3: encoding feature maps and feature fusion, respectively inputting radar echo map data and gridded atmospheric data into a GRU network containing two layers of downsampling for encoding, and concatenating the two obtained feature maps to obtain a fused feature map; S4: Generate predictions and output precipitation forecasts; In step S1, the denoising process of the historical radar echogram sequence at the input time t includes the following steps: S11: Use bilinear interpolation to fill the default data. The interpolation formula is as follows: Where Q ij is the position (x i ,y j ), i,j∈{1,2}, (x,y) is the position of the default value in the matrix; S12: Denoising the atmospheric grid data using hard threshold wavelet transform; In step S2, the convolution and deconvolution calculations used include the following steps: S21: Use 1*1 convolution kernel to adjust the number of feature channels. The calculation formula is: Where i represents the number of target feature channels mapped by the convolution kernel, and c represents the number of feature channels of the input data. S22: The new feature map obtained in the previous step is adjusted in size through deconvolution operation so that the scale of the feature map is consistent with the scale of the radar echo image. The calculation formula is as follows: O=K*I Where * is the matrix convolution operation, I is the matrix obtained in S21, K is the deconvolution kernel, and the parameters s, k, and p of the deconvolution kernel K are set to conform to the following formula: i=(o-1)*s+k-2p Where i represents the width of the meteorological data matrix, o represents the width of the radar echo map matrix, s represents the convolution kernel moving step, k represents the width of the convolution kernel, and p represents the width of the filling 0 value; In the step S3, the radar echo image data and the gridded atmospheric data are respectively input into an encoder network including two layers of down-sampling cyclic gating units for encoding, and the two obtained feature maps are spliced; wherein the encoder network is composed of a convolution layer, a cyclic gating unit layer, a down-sampling layer, a cyclic gating unit layer, a down-sampling layer, and a cyclic gating unit layer, the loss function selects a mean square error loss function, and the optimizer selects an Adam optimizer; the radar echo image code output by the encoder network and the gridded atmospheric data code matrix are spliced ​​on the feature channel; In step S4, the fused feature map is input into the decoder network of the GRU containing two layers of upsampling for decoding to obtain the predicted radar echo image, and the future regional precipitation prediction is obtained through ZR transformation, and the short-term precipitation prediction result Y is output. t ,Y t+1 ,...,Y t+p , where Y t+q It represents the result after the predicted radar echo map at time t+q is converted into a precipitation map, 1≤q≤p, p represents the total number of short-term precipitation prediction moments; wherein the decoder network is composed of a recurrent gating unit, a deconvolution upsampling, a recurrent gating unit, a deconvolution upsampling, a recurrent gating unit, and a deconvolution upsampling.

Images