CN-121980534-A - Visibility time sequence prediction method based on DeepAR model and weather prediction field

CN121980534ACN 121980534 ACN121980534 ACN 121980534ACN-121980534-A

Abstract

The application discloses a visibility time sequence prediction method based on DeepAR models and weather prediction fields, which mainly relates to the technical field of visibility time sequence prediction and is used for solving the problems that the existing scheme is insufficient in fusion of weather prediction field characteristics, only outputs a single numerical value and the existing DeepAR models cannot splice weather prediction field characteristics. The method comprises the steps of generating comprehensive feature vectors according to feature results, historical visibility time sequence observation data and preset additional covariates, inputting the comprehensive feature vectors into a hidden layer of an initial DeepAR model, completing construction of a future meteorological field extraction module in the DeepAR model, training and optimizing a DeepAR model constructed with the future meteorological field extraction module based on the historical visibility time sequence observation data to obtain a trained DeepAR model, and predicting visibility time sequence by utilizing the trained DeepAR model.

Inventors

ZHANG ZHANSHUO
LIU XIN
HAN QIQI
ZHANG JIE
GUAN SHUHUI
GUO YING

Assignees

山东省计算中心（国家超级计算济南中心）

Dates

Publication Date: 20260505
Application Date: 20251126

Claims (10)

1. A method for predicting a visibility time sequence based on DeepAR model and weather prediction field, the method comprising: The method comprises the steps of acquiring historical visibility time sequence observation data and historical weather forecast field data, carrying out time dimension alignment treatment on the historical weather forecast field data so that the historical visibility time sequence observation data is identical to the forecast time of the historical weather forecast field data; constructing an initial DeepAR model based on the LSTM model; carrying out multichannel self-attention mechanism calculation on the historical visibility time sequence observation data and the spatial correlation characteristics to obtain the spatial characteristics of the channel number of each meteorological element related to the historical weather forecast field data; The method comprises the steps of calculating the respective attention of each meteorological element channel by using spatial features, weighting and unifying the spatial association features of each meteorological element channel according to the channel attention to obtain a feature result, generating a comprehensive feature vector according to the feature result, historical visibility time sequence observation data and a preset additional covariate, inputting the comprehensive feature vector into a hidden layer of an initial DeepAR model, and completing the construction of a future meteorological field extraction module in a DeepAR model; training and optimizing DeepAR models with future meteorological field extraction modules based on historical visibility time sequence observation data to obtain a trained DeepAR model, and predicting visibility time sequence by using the trained DeepAR model.
2. The method for predicting a visibility time sequence based on DeepAR model and weather prediction field according to claim 1, wherein obtaining historical visibility time sequence observation data and historical weather prediction field data specifically includes: inputting and acquiring historical visibility time sequence observation data and historical weather forecast field data; Wherein, historical weather forecast field data The system consists of a meteorological element channel number C, a reporting time T, a forecasting time T and a spatial grid resolution H, W.
3. The method for predicting a time sequence of visibility based on DeepAR model and weather forecast field according to claim 1, wherein the alignment of time dimension is performed on the historical weather forecast field data so that the time sequence of the historical visibility is the same as the forecast time of the historical weather forecast field data, specifically comprising: acquiring forecast aging of historical weather forecast field data and historical visibility time sequence observation data; and carrying out time dimension interpolation processing on the historical weather forecast field data so as to lead the forecast time of the historical weather forecast field data and the historical visibility time sequence observation data to be the same.
4. The method for predicting the visibility time sequence based on DeepAR models and weather prediction fields according to claim 1, wherein the collection range of the historical weather prediction field data is: weather field data of a preset distance around a site is taken as a center.
5. The method for predicting the visibility time sequence based on DeepAR model and weather prediction field according to claim 1, wherein the extracting of spatial correlation features is performed on the preprocessed historical weather prediction field data, specifically comprising: historical weather forecast field data for four-dimensional tensors Performing time slicing processing to obtain three-dimensional tensors corresponding to each time t Wherein, C represents the channel number of meteorological elements, T represents forecast aging, H, W represents the resolution of a space grid; and then using a 2D convolution to check three-dimensional tensors of different meteorological element channels for spatial feature extraction: obtaining the spatial correlation characteristic of the c weather element channel ; Wherein, the To be the convolution kernel for the c-th meteorological element channel, The function is activated for the purpose of GLEU, Representing the three-dimensional tensor of the c-th meteorological element channel, and b representing a preset bias term.
6. The method for predicting the visibility time sequence based on DeepAR model and weather prediction field according to claim 1, wherein the method is characterized by performing multichannel self-attention mechanism calculation on historical visibility time sequence observation data and spatial correlation characteristics to obtain the spatial characteristics of the channel number of each weather element related to the historical weather prediction field data, and specifically comprises the following steps: Carrying out space flattening on the space correlation characteristics of the channel number of each meteorological element to obtain characteristic results; By passing through Calculating a Q value by using the historical visibility time sequence observation data Z; By: 、 Using characteristic results Calculating a K value and a V value; Wherein, the 、、 Representing a preset weight matrix; by the formula: Calculating an attention weight A; by the formula: And calculating the spatial characteristics of the channel number of each meteorological element.
7. The method for predicting the visibility time series based on DeepAR model and weather prediction field according to claim 1, wherein the calculating the respective attention of each weather element channel by using the spatial features specifically includes: by the formula: Calculating the attention of the current meteorological element channel ; Wherein, the 、 Representing a preset parameter matrix, O represents the spatial characteristics of the current meteorological element channel, Representing a modified linear element activation function, Representing the remodeling operational function.
8. The method for predicting a time sequence of visibility based on DeepAR model and weather prediction field according to claim 1, wherein the method for weighting and unifying the spatial correlation features of each weather element channel according to the channel attention to obtain feature results specifically comprises: by the formula: obtaining characteristic results of the c-th meteorological element channel ; Wherein, the Representing the attention of the current weather element channel, And representing the spatial correlation characteristics of the current meteorological element channel.
9. The method for predicting a time sequence of visibility based on DeepAR model and weather prediction field according to claim 1, wherein generating the comprehensive feature vector according to the feature result, the historical visibility time sequence observation data and the preset additional covariates specifically includes: by the formula: Generating a comprehensive feature vector ; Wherein, the Representing the characteristic results, Z representing historical visibility time sequence observation data, X representing a preset additional covariate, Representing the stitching function.
10. The method for predicting a visibility time sequence based on DeepAR models and weather prediction fields according to claim 1, wherein the training and optimizing the DeepAR model constructed with the future weather field extraction module based on the historical visibility time sequence observation data, the obtaining a trained DeepAR model specifically includes: dividing historical visibility time sequence observation data into a training set and a verification set; And learning a visibility change rule of the historical visibility time sequence observation data in the training set through a loss function, and adjusting super parameters by using a verification set.

Description

Visibility time sequence prediction method based on DeepAR model and weather prediction field Technical Field The application relates to the technical field of visibility time sequence prediction, in particular to a method for predicting a visibility time sequence based on DeepAR models and weather prediction fields. Background The accurate prediction of visibility is an important technical support for aviation safe operation, low-altitude economic development and ocean activities. In the aviation field, low visibility weather (such as fog, haze and the like) frequently occurs, so that problems of flight preparation, delay, even cancellation and the like are increasingly prominent, and particularly in the take-off and landing stage of an airport terminal area, sudden performance visibility dip directly threatens flight safety. According to the civil aviation operation specification, when the visibility is reduced to 1000 meters or the cloud bottom is higher than 90 meters and the visibility is in a descending trend, low-visibility early warning needs to be issued, and when the visibility is further reduced to 800 meters, a low-visibility operation program is started. In the ocean field, sea fog is a weather phenomenon with atmospheric visibility lower than 1000 meters, is commonly found on the sea, islands and target sea areas, and seriously affects navigation safety and production operation. Therefore, the accurate forecasting visibility has important significance for the safety guarantee of civil aviation, ocean navigation and related production activities. The following core problems exist in the existing visibility prediction technology: 1. The traditional deep learning method mainly relies on time sequence modeling of historical observation data, and cannot effectively integrate grid space information of a future high-resolution weather forecast field, so that two-dimensional and three-dimensional distribution characteristics of key weather elements such as a temperature field, a humidity field and a wind field are not fully utilized, and prediction accuracy is limited. 2. Only a single numerical value is output, namely, the prior art adopts LSTM, GRU, transformer and other deterministic models, and only single numerical value prediction is output. However, the strong randomness and burstiness characteristics of low visibility events place higher demands on probability prediction. 3. DeepAR the model application limitation is that although the DeepAR model has probability generation capability, effective information interaction with weather forecast field characteristics is not realized in the field of visibility prediction, and the potential of the model in complex weather scenes cannot be fully exerted. Disclosure of Invention The application provides a visibility time sequence prediction method based on DeepAR models and a weather prediction field, which aims to solve the problems that the existing scheme has insufficient fusion of weather prediction field characteristics, only outputs a single numerical value and the existing DeepAR model cannot predict weather prediction field characteristics. In a first aspect, the present application provides a method for predicting a visibility time sequence based on DeepAR model and weather prediction field, the method comprising: Acquiring historical visibility time sequence observation data and historical weather forecast field data; carrying out time dimension alignment treatment on the historical weather forecast field data so as to enable the historical visibility time sequence observation data to be identical with the forecast time efficiency of the historical weather forecast field data; preprocessing historical visibility time sequence observation data and historical weather forecast field data; constructing an initial DeepAR model based on the LSTM model; carrying out multichannel self-attention mechanism calculation on the historical visibility time sequence observation data and the spatial correlation characteristics to obtain the spatial characteristics of the channel number of each meteorological element related to the historical weather forecast field data; the spatial correlation characteristics of each meteorological element channel are weighted and unified according to the channel attention, and a characteristic result is obtained; Generating a comprehensive feature vector according to the feature result, the historical visibility time sequence observation data and the preset additional covariates, inputting the comprehensive feature vector into a hidden layer of the initial DeepAR model, and completing the construction of a future meteorological field extraction module in the DeepAR model; training and optimizing DeepAR models with future meteorological field extraction modules based on historical visibility time sequence observation data to obtain a trained DeepAR model, and predicting visibility time sequence by using the trained DeepAR mod