Search

CN-122020500-A - Discrete site precipitation short-term forecasting method based on convolution and dimension reduction deconvolution neural network

CN122020500ACN 122020500 ACN122020500 ACN 122020500ACN-122020500-A

Abstract

The invention discloses a discrete site precipitation short-term forecasting method based on a convolution and dimension reduction deconvolution neural network, and belongs to the application field of artificial intelligence in weather and related industries for short-term precipitation forecasting. According to the method, a dataset is constructed through normalization of multivariable two-dimensional predictors and precipitation data of discrete sites, and accurate prediction of future multi-day precipitation of the multiple discrete sites is achieved through multi-layer multi-core two-dimensional convolution, attention mechanism fusion, flattening processing, one-dimensional deconvolution up-sampling and loss function optimization. The method can automatically learn complex nonlinear relations, effectively extract multi-scale weather features, solve the problem of 'field-point' forecasting, meet the service timeliness requirement, and has good portability and wide popularization and application prospects.

Inventors

  • ZHAO YUCHUN
  • WANG YEHONG
  • HUANG YIPENG
  • XUN AIPING
  • ZHENG HUI

Assignees

  • 厦门市气象台(厦门市海洋气象台、海峡气象开放实验室)

Dates

Publication Date
20260512
Application Date
20260127

Claims (10)

  1. 1. The discrete site precipitation short-term forecasting method based on the convolution and dimension reduction deconvolution neural network is characterized by comprising the following steps of: S1, carrying out normalization processing on a multivariate two-dimensional predictor X and precipitation data Y of N discrete stations in the future D days, and establishing a predictor data set and a predictor data set according to the corresponding relation between the predictor and a predictor; S2, carrying out multi-layer and multi-core two-dimensional convolution, nonlinear activation, standardization and pooling operation on the normalized multivariable two-dimensional predictor X, respectively adding a spatial attention mechanism and a channel attention mechanism in a second layer and a penultimate layer, and carrying out dropout operation in a last layer; S3, carrying out one-dimensional flattened fusion on the feature map output in the step S2; S4, carrying out one-dimensional deconvolution up-sampling calculation on the flattened fusion result, adjusting deconvolution parameters to enable output dimensions to meet requirements, and carrying out alignment treatment; s5, constructing a mean square error loss function, minimizing the loss function through an optimization algorithm, and updating model parameters; S6, inputting new multivariable two-dimensional predictor data, and forecasting by using the trained model to obtain rainfall predicted values of N discrete stations in the future D days.
  2. 2. The method according to claim 1, wherein the normalization of the multivariate two-dimensional predictor X in step 1 uses a min-max normalization with the formula: ; wherein the dimension of X is% ,C, , ), For the number of samples, C is the number of channels or the number of variables, In order to input the data width of the data, In order for the height of the data to be entered, And The multi-variable two-dimensional forecasting factors comprise 850hPa, 500hPa and 200hPa on the equal pressure surfaces of temperature, humidity, uwind, V wind, potential height, equivalent temperature, wet static energy, convection effective potential energy, wind vertical shear and wet vortex, sea level air pressure, ground 2m temperature, ground 2m humidity, ground 10m U wind and ground 10m V wind.
  3. 3. The method of claim 1, wherein the normalization formula for the N discrete site precipitation data Y in step S1 is: ; wherein the dimension of Y is ,N, For the number of samples, N is the number of the water-reducing stations, Is the maximum value of the precipitation data of N sites in the sample dimension.
  4. 4. The method of claim 1, wherein the specific process of creating the dataset in step S1 is to set time t=0, extract the normalized predictor data at time t=0, and extract the precipitation data of N discrete sites on D days after time t=0 according to the number of days D in advance of the prediction and arrange in time sequence, slide time t, repeat the above extraction operations to form dimensions B, , , And B, N x D of predictor datasets and forecast object datasets, wherein B is the number of samples that entered training.
  5. 5. The method of claim 1, wherein the convolution kernel size k e {3,5,7,9} used in step S2 is a separate branch for each kernel, performing a J-layer convolution operation, J being greater than or equal to 4, J = 1,2, the J-th layer J = 1,2 Convolution step length =1, Fill The method is characterized in that the method comprises the steps of setting according to the size of a core to keep the space size, adopting batch normalization for standardization, adopting a ReLU function for nonlinear activation, adopting 2X 2 maximum pooling for pooling, and adopting the step length of 2.
  6. 6. The method of claim 1, wherein the implementation process of the spatial attention mechanism in the step S2 is that after the second layer of pooling, a1×1 convolution kernel is used for convolution to output a single-channel feature map, spatial weights are obtained through a Sigmoid function, element multiplication operation is carried out on the pooling layer result and the spatial weights, and the implementation process of the channel attention mechanism is that after the second layer of pooling, global average pooling is carried out first, channel weights are obtained through two layers of full-connection layers and Sigmoid functions, and channel multiplication operation is carried out on the pooling layer result and the channel weights.
  7. 7. The method of claim 1, wherein the one-dimensional flattening fusion in step S3 includes flattening the J-layer operation output feature map corresponding to each core into vectors, and splicing and fusing the flattened vectors corresponding to all cores into a single one-dimensional vector through a merging operation.
  8. 8. The method of claim 1, wherein the one-dimensional deconvolution in step S4 uses convolution kernels in order of magnitude from odd to small, including odd kernels and even kernels, with a channel number gradually decreasing to 1, and repeating deconvolution once when the channel number is 1, deconvolution times M, step size And filling By iterative adjustment, the deconvolution output length Satisfy the following requirements And finally, performing quasi-symmetrical shearing on the deconvolution result to align with the NxD, and removing the channel dimension to obtain a result with the dimension (B, nxD).
  9. 9. The method of claim 1, wherein the mean square error loss function formula in step 5 is: ; Wherein, the In order to output the result of the deconvolution, The optimization algorithm adopts Adam algorithm to calculate the variation of loss function The optimization is stopped when the time is over, To set a threshold.
  10. 10. The method according to claim 1, wherein the forecasting process in step 6 is that the new input data is processed according to the normalization method in step 1 and then is input into a trained model to obtain a normalized predicted value, and the normalized predicted value is subjected to inverse normalization operation Obtaining a final precipitation prediction value, wherein Predicted values were normalized for the D-th day of model output, d=1, 2.

Description

Discrete site precipitation short-term forecasting method based on convolution and dimension reduction deconvolution neural network Technical Field The invention belongs to the technical field of applying artificial intelligence technology to weather and related industries for carrying out short-term rainfall forecast, and particularly relates to a discrete site rainfall short-term forecast method based on a convolution and dimension reduction deconvolution neural network. Background The short-term quantitative rainfall forecast is one of core tasks in meteorological service, and has important significance in the fields of disaster prevention and reduction, water resource management, agricultural production and the like. At present, short-term quantitative rainfall forecasting technology and method are mainly divided into three main categories, namely a numerical weather forecasting and set forecasting method, a numerical forecasting post-processing method and a statistical and artificial intelligence method. The numerical weather forecast and aggregate forecast method predicts by solving the atmospheric transport physical equation set, is a physical basis of modern weather forecast, is influenced by uncertainty such as a model power frame, an initial value, physical process parameterization and the like, has limited rainfall forecast capability with obvious nonlinear characteristics, and has high calculation cost. The numerical prediction post-processing method (such as MOS method) corrects systematic deviation by establishing a statistical relation between numerical mode output and observed value, but severely depends on the length and quality of historical data, and retrains after upgrading the numerical mode, and has insufficient suitability for dual characteristics (discrete type and continuous type) of precipitation. Statistical and artificial intelligence methods have become research hotspots by virtue of big data processing and complex pattern learning capabilities, but the existing methods still have a plurality of defects: 1. The linear regression model assumes that the forecasting factors and the rainfall are in a linear relation, cannot process complex nonlinear problems, depends on a large amount of historical data, and fails after the model is upgraded; 2. the nonlinear regression model needs to manually preset a nonlinear function form, the prior knowledge is excessively relied on, and the effect of the model is greatly reduced due to the fact that the function form is selected incorrectly; 3. the U-Net convolutional neural network method has high reasoning speed, but the physical consistency is lost, the prediction is good at short-term prediction, the prediction performance decay is fast in 1-4 days, and the problem of 'field-point' prediction cannot be solved; 4. ConvLSTM space-time sequence prediction method is limited by convolution kernel vision, is difficult to capture large-range environmental field information, has a fuzzy effect easily caused by a prediction result, accumulates errors along with prediction duration, and cannot cope with discontinuous daily rainfall prediction; 5. Although the improved space-time sequence prediction method is optimized in a specific scene, the problems of insufficient long-term memory, high calculation overhead, error accumulation and the like still exist, and the precipitation prediction requirement of 1-4 days of discrete sites is difficult to meet. The ideal discrete site short-term quantitative precipitation prediction method has the capabilities of modeling complex nonlinearity and space-time dependence, optimizing 'field-point' prediction, sensing a large-range environment field, solving discontinuous daily precipitation prediction, meeting service timeliness requirements and the like. Therefore, there is an urgent need for a new forecasting method that overcomes the drawbacks of the prior art. Disclosure of Invention The invention aims to provide a discrete site precipitation short-term forecasting method based on a convolution and dimension reduction deconvolution neural network. In order to achieve the above purpose, the present invention provides the following technical solutions: Data set construction The multi-variable two-dimensional forecasting factor X is subjected to minimum-maximum normalization to ensure that the value range is 0,1, and the normalization formula is thatWherein the dimension of X is%,C,,),For the total number of samples, C is the number of channels or the number of variables,In order to input the data width of the data,In order for the height of the data to be entered,AndThe minimum and maximum values in the other dimensions after the channel dimension are reserved for the input data X, respectively. The multivariate two-dimensional predictors include temperature, humidity, uwind, V wind, potential height, equivalent potential temperature, wet static energy, convection effective potential energy, wind vertical shear