Deep learning-based sand and dust short-term and imminent forecasting method
By building a PredRNN network, using the spatiotemporal sequence sample set of ground stations and satellite inversion data, efficient and accurate prediction of sand and dust short-term forecasting is achieved, and the problem of insufficient precision of sand and dust short-term forecasting in the existing technology is solved.
Patent Information
- Application Number
- CN202410948641.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-16
- Publication Date
- 2024-11-08
AI Technical Summary
The existing sand and dust forecasting methods lack the forecast accuracy in a short period of time, especially in areas with complex terrain or limited observations, which makes it difficult to accurately predict the area and range of sand and dust, resulting in traffic accidents and economic losses.
Using the short-term prediction method of sand and dust based on deep learning, the PredRNN network is constructed, and the ground site observation data and satellite inversion data are used to form a spatiotemporal sequence sample set, and the sand and dust concentration prediction model is trained to achieve the sand and dust concentration in the next few hours. Accurate prediction.
It improves the accuracy and efficiency of short-term forecasts of sand and dust, can accurately predict the development of sand and dust in a short time, makes up for the shortcomings of insufficient performance of numerical models in short-term periods, and improves forecasting capabilities.
Smart Images

Figure CN118915191A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of sandstorm forecasting, and in particular to a short-term sandstorm forecasting method based on deep learning. Background Art
[0002] Dust refers to a weather phenomenon in which a large amount of dust and sand on the ground is blown into the air by strong winds under specific geographical environments and underlying surface conditions, making the air turbid and reducing horizontal visibility. It is characterized by strong suddenness, short duration, and wide impact range. It will sharply reduce visibility and seriously threaten ecological environmental safety, urban operations, and human health. Improving the accuracy of sandstorm forecasts is the focus and difficulty of current meteorological work. By using the opposite observation station network, based on weather phenomena, visibility and other observation elements, the time, location, intensity and other characteristics of sandstorms can be obtained for a long time and all-weather. At the same time, the real-time sandstorm products based on satellite inversion can monitor the development process of sandstorms continuously in time and space. These sandstorm monitoring data are the basis for conducting sandstorm research and forecasting.
[0003] Existing dust forecasts can be divided into numerical forecast methods and machine learning forecast methods according to the different implementation principles of the forecast methods. The numerical forecast method simulates dust emission and transport to dry or wet deposition of dust through a group of dynamic equations to predict the future state of dust. It is the main method currently used in dust forecast operations. Mature dust numerical model systems include global dust numerical models such as GOCART, DREAM, and ITR, as well as regional dust numerical models such as CEMSYS5, CFORS, and NARCM, as well as China's CUACE-Dust, GRAPES-CUACE / Dust, etc. Since numerical models are difficult to accurately represent small-scale initial conditions and lack sufficient resolution in both vertical and horizontal directions, they are more suitable for long-term forecasts. In addition, it is still difficult for numerical models to generate key parameters involved in the formation and evolution of dust, especially in areas with complex terrain or limited observations, which makes the forecast results have certain uncertainties.
[0004] The main idea of using machine learning related methods for sandstorm forecasting is to predict sandstorm weather by modeling the relationship between meteorological observation data and sandstorm weather related observation data. For example, Lu et al. used meteorological characteristics such as 500hPa height field, two wind fields (U, V) and temperature field as input variables to construct a sandstorm day prediction model based on the SVM method. Zhang Zhenhua et al. predicted the number of sandstorm days in Gansu Province, Xinjiang Uygur Autonomous Region, Ningxia Hui Autonomous Region and other northwestern regions based on the SMOTE sampling algorithm and decision tree algorithm. These traditional machine learning algorithms such as shallow neural networks and support vector machines are mainly used to predict the probability of sandstorms in a certain place and cannot be used in actual business applications. Relatively speaking, deep neural networks can not only provide modeling for complex nonlinear systems, but also provide a higher level of abstraction for the model, thereby improving the feature extraction ability of the model, and have been widely used in short-term precipitation forecasts.
[0005] The short-term forecast of sandstorms is an important basis for early warning and forecasting of sandstorm disasters. People pay more attention to the short-term forecast effect of sandstorms. The potential forecast method with higher forecast efficiency can help business personnel determine the area and scope of sandstorms in advance, so as to focus on early warning in key areas. The low visibility caused by sandstorms may lead to traffic accidents, flight cancellations and road closures, resulting in casualties and serious economic losses. Therefore, there is an urgent need for a more efficient and effective short-term forecast method for sandstorms to provide better meteorological service support for civil aviation centers and highway management departments. Summary of the invention
[0006] In response to the above-mentioned problems in the above-mentioned prior art, the present invention provides a short-term and impending sandstorm forecasting method based on deep learning, which converts the short-term and impending sandstorm forecasting problem into a sandstorm actual situation sequence prediction problem, and obtains the short-term and impending sandstorm forecasting map a few hours later by inputting multiple continuous historical frame sandstorm actual situation product sequence prediction maps into the network. The development of sandstorms in the near-forecast is checked by judging the area and sandstorm concentration of the prediction map, thereby realizing accurate and rapid short-term and impending sandstorm forecasting.
[0007] In order to achieve the above object, the technical solution adopted by the present invention is as follows:
[0008] A method for short-term sandstorm forecasting based on deep learning, comprising the following steps:
[0009] S10, filtering out dust process data based on the observation data of the ground station, and matching the existing dust real-time data based on satellite inversion according to the time and space of the dust process data;
[0010] S20, preprocessing the data matched in step S10 to form a dust spatiotemporal sequence sample set;
[0011] S30, constructing a deep learning model of spatiotemporal sequence based on the PredRNN network, and using the dust spatiotemporal sequence sample set of step S20 to train and obtain a dust concentration prediction model, which is expressed as:
[0012] P t+1 ~P t+n = argmaxP(X t+1 ~X t+n |X t-J+1 ~X t )
[0013] X t+1 ~X t+n
[0014] In the above formula, X t-J+1 ~X t represents the historical dust concentration data sequence from time t-J+1 to time t, X t+1 ~X t+n represents the true value data sequence of dust concentration from time t+1 to time t+n, P t+1 ~P t+n represents the dust concentration forecast result from time t+1 to time t+n, that is, the predicted dust concentration spatiotemporal data sequence with the maximum probability P, n <J;
[0015] S40. When sandstorm weather occurs, the existing actual sandstorm data based on satellite inversion is input into the sandstorm concentration prediction model obtained in step S30, and the sandstorm distribution results in the future set time period of the designated area are output to make a short-term forecast of sandstorms.
[0016] Specifically, the observation data of the ground station in step S10 includes the visibility and weather phenomenon data in the hourly full-factor data of the ground national meteorological station, and the PM10 data of the ground environmental protection station; the existing satellite-based inversion real-time sand and dust data is a sand and dust grid image containing sand and dust area and non-dust area markers in time series.
[0017] Specifically, the process of performing temporal and spatial matching according to the dust process data and the existing satellite-based inversion dust real-time data in step S10 includes:
[0018] S11. Collect observation data of ground stations during historical dust storms, and filter out time information including the start and end time of dust storms and spatial information including the latitude and longitude information of dust storm occurrence sites;
[0019] S12, according to the time information obtained in step S11, screening the data containing the same time information in the actual dust data inverted based on the satellite, and completing the time matching of the data;
[0020] S13. Under the condition of time matching, extract the inversion result data from the actual dust data based on satellite inversion according to the spatial information obtained in step S11, and judge whether the inversion result data is consistent with the observation data in the specified area. If so, include the actual dust data based on satellite inversion at this time and space into the initial sample.
[0021] Specifically, the process of preprocessing the data in step S20 to form a dust spatiotemporal sequence sample set is as follows:
[0022] S21, performing regional cropping on the data matched in step S10, and forming a 256×512 grid after cropping;
[0023] S22, using a sliding window mode to obtain a sample from the cropped data at every J+n consecutive moments, and cyclically obtain samples to form a sample set;
[0024] S23. Use the maximum and minimum normalization method to standardize each sample and assign the null value to 0 to obtain a dust spatiotemporal series sample set.
[0025] Specifically, the existing satellite-based inversion dust data in step S10 is in geotiff format, and the sample format of the dust spatiotemporal sequence sample set in step S20 is stored in npy format.
[0026] Specifically, the deep learning model of the spatiotemporal sequence constructed based on the PredRNN network in step S30 is a stacked network structure. The input dust image is converted from a single-channel image to a dust feature map with a latent space channel dimension D of W through a feature embedding layer. The dust feature map is converted into a single-channel dust prediction image output through stacked network structure learning and feature dimension reduction layer processing.
[0027] Specifically, the process of training and obtaining the dust concentration prediction model in step S30 includes:
[0028] S31, initial parameter settings for model training: the number of model recurrent unit layers is set to 4, the hidden feature dimension between basic units is set to 64; the loss function is set to the MSE loss function, and the initial learning rate is set to 0.0001; the planned sampling stop round number is set to 5000; the filter size is set to 2; the AdamW optimizer is used to minimize the MSE loss function, and the learning rate is adjusted through its adaptive gradient mechanism and momentum gradient mechanism to speed up the model convergence; the maximum iteration round is set to 200; the input spatiotemporal sequence length is set to J, and the predicted spatiotemporal sequence length is set to n;
[0029] S32, model training starts: using the sample data in the dust spatiotemporal sequence sample set as input, setting the input dust concentration image sequence X1-XJ [J, 1, 256, 512], where the input sequence length is J, the number of grayscale image channels is 1, the image grid size is 256 × 512, and the model effect is verified once in each round of training and the model parameters are saved;
[0030] S33, model learning: from the input dust concentration image sequence X1-X J [J,1,256,512] extracts a single dust concentration image X each time according to the time dimension i [1,1,256,512], i=1,2,…,J, input to the model, cycle this process J times, after feature embedding, stacked recurrent network and feature dimensionality reduction, save the feature map P output from the last cycle J+1 ;
[0031] S34, model prediction: The feature map P output in step S33 is J+1 Then input it into the model and output the predicted dust concentration image P at time J+2 J+2 , and then use this predicted dust concentration image as the model input for the next cycle, and output the predicted dust concentration image P at time J+3 J+3 Repeat this process until the predicted dust concentration image P is obtained. J+n ;
[0032] S35. Model tuning: Using the predicted dust concentration image P J+1 ~P J+n And the corresponding real dust concentration image X J+1 ~X J+n Calculate the MSE loss function and optimize the model parameters;
[0033] S36. Repeat steps S33 to S35 to iteratively optimize the model parameters. After completing the maximum number of iterations, save the model parameters to obtain a dust concentration prediction model. During this period, observe the model learning status by drawing loss curves and evaluation index curves.
[0034] Specifically, the spatiotemporal sequence deep learning model based on the PredRNN network takes the ST-LSTM unit as the network core, including multiple layers of stacked neural units, which flow up and down between layers through spatiotemporal memory units, and are transmitted between layers through long-term memory units and hidden layers. t-1 , obtain the predicted dust concentration image at output time t It is expressed as:
[0035]
[0036]
[0037] In the above formula, X t is the input dust concentration image at time t, the [] symbol represents the concatenation operation, and the * symbol represents the convolution operation. The symbol represents matrix multiplication, and They represent the long-term memory unit and spatiotemporal memory unit at the lth layer at time t, respectively. represents the value of the hidden layer at the lth layer at time t, g t 、i t and f t They represent the modulation gate, input gate and forget gate of the long-term memory unit C, respectively. t '、i t ' and f t 'represent a group of modulation gates, input gates and forget gates of the spatiotemporal memory unit M respectively. The newly added group of gates has the same formula as the traditional gate mechanism. t represents the output gate, tanh and σ are the corresponding activation functions, W xg , W xi , W xf , W hg , W hi , W hf , W′ xg , W′ xi , W′ xf , W mg , W mi , W mf , W xo , W ho , W co and W mo represents the weight of the corresponding convolution operation, b g 、b i 、b f , b' g 、b i ', b' f and b o represents the learnable offset parameter, w 1×1 Indicates for extraction The weights of the 1×1 convolution operation on the spatial features of the concatenated feature map.
[0038] Preferably, the input space-time sequence length is set to 6, and the predicted space-time sequence length is set to 3.
[0039] Compared with the prior art, the present invention has the following beneficial effects:
[0040] The present invention utilizes high-precision actual dust data generated by inversion based on geostationary satellites, and combines it with the observation data of ground stations for time-space matching to form a dust space-time series sample set to train the dust concentration prediction model, giving full play to the advantages of satellite data in monitoring the dust process, avoiding the limitation that conventional observation station site observations or radar observations in sparse areas cannot comprehensively and continuously monitor the dust process. The dust concentration prediction model obtained through training uses deep learning methods to improve the accuracy of short-term and impending dust forecasts, effectively making up for the shortcomings of insufficient performance of numerical model-based forecasting methods in short-term time periods in the business, and improving the overall short-term and impending dust forecasting capabilities. BRIEF DESCRIPTION OF THE DRAWINGS
[0041] Figure 1 It is a schematic diagram of the overall process of an embodiment of the present invention.
[0042] Figure 2 It is a schematic diagram of the processing process in the embodiment of the present invention.
[0043] Figure 3 Schematic diagram of the PredRNN network structure in an embodiment of the present invention.
[0044] Figure 4 This is a diagram of the actual dust distribution at a certain moment used in the experiment in the embodiment of the present invention.
[0045] Figure 5 This is a prediction diagram of etching dust distribution based on the experimental test results in an embodiment of the present invention. DETAILED DESCRIPTION
[0046] The present invention is further described below in conjunction with the accompanying drawings and examples. The embodiments of the present invention include but are not limited to the following examples.
[0047] like Figure 1 to Figure 2 As shown, the short-term sandstorm forecasting method based on deep learning includes the following steps:
[0048] S10. Filter out the dust process data according to the observation data of the ground station, and match the existing dust real-time data based on satellite inversion according to the time and space of the dust process data; the observation data of the ground station include the visibility and weather phenomenon data in the hourly full-factor data of the ground national meteorological station, and the PM10 data of the ground environmental protection station; the existing dust real-time data based on satellite inversion is a dust grid image containing dust area and non-dust area marks in time series. The dust real-time data based on satellite inversion can be obtained through the patented technology with application number 202111577050.2, and is usually saved in the geotiff format file.
[0049] The specific process of space-time matching is as follows:
[0050] S11. Collect observation data from ground stations during historical sandstorms, and filter out time information including the start and end time of the sandstorms, and spatial information including the longitude and latitude information of the sandstorm site.
[0051] S12, according to the time information obtained in step S11, filter the data containing the same time information in the satellite-inverted dust real-time data to complete the time matching of the data. For example, filter the files containing the same time information from the geotiff format files of the satellite-inverted dust real-time data.
[0052] S13, under the condition of time matching, extract the inversion result data from the actual dust data based on satellite inversion according to the spatial information obtained in step S11, and judge whether the inversion result data is consistent with the observation data in the specified area. If so, the actual dust data based on satellite inversion at the time and space is included in the initial sample, otherwise it is not included in the sample. According to the actual application, the specified area can be selected within the Inner Mongolia Autonomous Region.
[0053] S20, preprocessing the data matched in step S10 to form a dust spatiotemporal sequence sample set, the specific process is as follows:
[0054] S21, performing regional cropping on the data matched in step S10, and forming a 256×512 grid after cropping;
[0055] S22, using a sliding window mode to obtain a sample from the cropped data at every J+n consecutive moments, and cyclically obtain samples to form a sample set;
[0056] S23. Use the maximum and minimum normalization method to standardize each sample and assign the null value to 0 to obtain a dust spatiotemporal series sample set.
[0057] Taking the longitude and latitude range of the Inner Mongolia Autonomous Region as an example, the horizontal spatial resolution of the actual dust data based on satellite inversion is 0.05°×0.05°, and the longitude and latitude range of the cropped data is [36°N, 48.75°N, 100.95°E, 126.55°E], forming a 256×512 grid. The data used by the deep learning model contains information about the evolution of the dust process over time, and each data sample contains T n Through the final test, it is found that when T n When the value is 6, the model can achieve a good balance between calculation time and forecast performance. Therefore, the data of the past 6 hours can be used to predict the sandstorm activity in the next 3 hours. The time parameters are configured as J=6, n=3. When constructing the sandstorm spatiotemporal sequence sample set, the sliding window mode is adopted to obtain samples cyclically, and a sample is formed with 9 frames of images as a group.
[0058] The dust spatiotemporal sequence sample set is divided into three parts: training set, test set and validation set according to the set ratio, which provides guarantee for the subsequent machine learning process and is stored in npy format. In actual application, dust process data for 5 years from 2019 to 2023 are collected for testing and use.
[0059] S30, constructing a deep learning model of spatiotemporal sequence based on the PredRNN network, and using the dust spatiotemporal sequence sample set of step S20 to train and obtain a dust concentration prediction model, which is expressed as:
[0060]
[0061] In the above formula, X t-J+1 ~X t represents the historical dust concentration data sequence from time t-J+1 to time t, X t+1 ~X t+n represents the true value data sequence of dust concentration from time t+1 to time t+n, P t+1 ~P t+n represents the dust concentration forecast result from time t+1 to time t+n, that is, the predicted dust concentration spatiotemporal data sequence with the maximum probability P, n <J。
[0062] Among them, the deep learning model of spatiotemporal sequences constructed based on the PredRNN network is a stacked network structure. The input dust image is converted from a single-channel image to a dust feature map with a latent space channel dimension D of W through the feature embedding layer. The dust feature map is converted into a single-channel dust prediction image output through stacked network structure learning and feature dimension reduction layer processing.
[0063] like Figure 3 As shown in FIG. 1 , the spatiotemporal sequence deep learning model based on the PredRNN network takes the ST-LSTM unit as the network core, including multiple layers of stacked neural units W l , the maximum number of layers is 4, through the spatiotemporal memory unit Flowing up and down between layers makes it easier to capture information through long-term memory units and hidden layers The transfer between layers ensures the fluidity of memory. By inputting the input dust concentration image X at time t-1 t-1 , obtain the predicted dust concentration image at output time t This structure improves the fluidity of space and time, ensures the simultaneous update of spatial and temporal states, and makes the prediction results more effective. The ST-LSTM unit is represented as:
[0064]
[0065] In the above formula, X t is the input dust concentration image at time t, the [] symbol represents the concatenation operation, and the * symbol represents the convolution operation. The symbol represents matrix multiplication, and They represent the long-term memory unit and spatiotemporal memory unit at the lth layer at time t, respectively. represents the value of the hidden layer at the lth layer at time t, g t 、i t and f t They represent the modulation gate, input gate and forget gate of the long-term memory unit C, respectively. t '、i t ' and f t 'represent a group of modulation gates, input gates and forget gates of the spatiotemporal memory unit M respectively. The newly added group of gates has the same formula as the traditional gate mechanism. t represents the output gate, tanh and σ are the corresponding activation functions, W xg , W xi , W xf , W hg , W hi , W hf , W′ xg , W′ xi , W′ xf , W mg , W mi , W mf , W xo , W ho , W co and W mo represents the weight of the corresponding convolution operation, b g 、b i 、b f , b' g 、b i ', b' f and bo represent learnable offset parameters, w 1×1 Indicates for extraction The weights of the 1×1 convolution operation on the spatial features of the concatenated feature map.
[0066] The process of training and obtaining the dust concentration prediction model in step S30 includes:
[0067] S31. Initial parameter settings for model training: the number of model recurrent unit layers is set to 4, and the latent feature dimension between basic units is set to 64; the loss function is set to the MSE loss function, and the initial learning rate is set to 0.0001; the planned sampling stop round number is set to 5000; the filter size is set to 2; the AdamW optimizer is used to minimize the MSE loss function, and the learning rate is adjusted through its adaptive gradient mechanism and momentum gradient mechanism to speed up the model convergence; the maximum number of iterations is set to 200; the input space-time sequence length is set to J, and the predicted space-time sequence length is set to n; through the final tuning test, J=6 and n=3 are preferably selected.
[0068] S32, model training starts: using the sample data in the dust spatiotemporal sequence sample set as input, setting the input dust concentration image sequence X1-X J [J, 1, 256, 512], where the input sequence length is J, the number of grayscale image channels is 1, and the image grid size is 256 × 512. During the training process, the model effect is verified once in each round and the model parameters are saved.
[0069] S33, model learning: from the input dust concentration image sequence X1-X J [J,1,256,512] extracts a single dust concentration image X each time according to the time dimension i [1,1,256,512], i=1,2,…,J, input to the model, cycle this process J times, after feature embedding, stacked recurrent network and feature dimensionality reduction, save the feature map P output from the last cycle J+1 .
[0070] S34, model prediction: The feature map P output in step S33 is J+1 Then input it into the model and output the predicted dust concentration image P at time J+2 J+2 , and then use this predicted dust concentration image as the model input for the next cycle, and output the predicted dust concentration image P at time J+3 J+3 Repeat this process until the predicted dust concentration image P is obtained. J+n .
[0071] S35. Model tuning: Using the predicted dust concentration image P J+1 ~P J+n And the corresponding real dust concentration image X J+1 ~X J+n Calculate the MSE loss function and optimize the model parameters.
[0072] S36. Repeat steps S33 to S35 to iteratively optimize the model parameters. After completing the maximum number of iterations, save the model parameters to obtain a dust concentration prediction model. During this period, observe the model learning status by drawing loss curves and evaluation index (MAE, MSE and PSNR) curves.
[0073] The calculation formulas for the evaluation indicators MAE, MSE and PSNR are as follows:
[0074]
[0075] In the above formula, f(x i ) predict dust image, y i represents the real dust image, MAE is the mean absolute error between the predicted dust image and the real dust image, MSE is the mean square error between the predicted dust image and the real dust image, and the peak signal-to-noise ratio PSNR quantifies the difference between the original dust image and the predicted dust image by calculating the MSE between the two. MAX represents the maximum possible value of the pixel value. The higher the PSNR value, the higher the similarity between the quality of the predicted dust image and the real dust image.
[0076] According to the results of the experimental test, by dividing the dust spatiotemporal series sample set, 277 test sets were used to evaluate the dust concentration prediction model. The final result showed that the mean absolute error (MAE) was less than 60ug / m 3 , mean square error MSE is less than 800ug / m 3 , the image quality evaluation index PSNR (peak signal-to-noise ratio) is greater than 37. The evaluation index results show that the present invention has high accuracy and can be used for short-term sandstorm forecasting in business.
[0077] S40, when sandstorm occurs, input the existing satellite-based inversion sandstorm data into the sandstorm concentration prediction model obtained in step S30, output the sandstorm distribution results in the future set time period in the designated area, and make a short-term sandstorm forecast. For example, at a certain time point, the satellite-based inversion sandstorm data monitors the occurrence of sandstorm in the northern part of Xilin Gol League. Through the short-term sandstorm forecast, it is found that the Xinghuangqi Airport will be affected by the sandstorm in the next three hours, and timely provide early warning services for the Xinghuangqi Airport.
[0078] The model and process of the present invention are tested using actual historical dust process data, and the dust concentration prediction data output by the model is compared visually with the actual dust concentration data. Figure 4 and Figure 5 As shown, the dust range predicted by the model is consistent with the real dust image range, and the dust concentration value is also close to the actual situation. The qualitative evaluation results show that the present invention is effective in short-term forecasting of dust and can better predict the future development of dust.
[0079] Through the above process, the present invention achieves a relatively accurate short-term sandstorm forecast effect. Compared with the traditional sandstorm numerical model forecasting method, the present invention benefits from the full and comprehensive use of satellite data and ground station observation data, making the sandstorm process data source richer and more accurate, and providing a solid foundation for the application of deep learning models. In addition, the present invention fully considers the balance between model calculation time and prediction effect through model parameter configuration, which not only meets the timeliness requirements of short-term sandstorm forecasts, but also effectively improves the accuracy of prediction effects.
[0080] The above embodiments are only preferred embodiments of the present invention and are not intended to limit the protection scope of the present invention. Any changes made by adopting the design principles of the present invention and performing non-creative work on this basis should fall within the protection scope of the present invention.
Claims
1. A short-term sandstorm forecasting method based on deep learning, characterized in that: The following steps are involved: S10, filtering out dust process data based on the observation data of the ground station, and matching the existing dust real-time data based on satellite inversion according to the time and space of the dust process data; S20, preprocessing the data matched in step S10 to form a dust spatiotemporal sequence sample set; S30, constructing a deep learning model of spatiotemporal sequence based on the PredRNN network, and using the dust spatiotemporal sequence sample set of step S20 to train and obtain a dust concentration prediction model, which is expressed as: In the above formula, X t-J+1 ~X t represents the historical dust concentration data sequence from time t-J+1 to time t, X t+1 ~X t+n represents the true value data sequence of dust concentration from time t+1 to time t+n, P t+1 ~P t+n represents the dust concentration forecast result from time t+1 to time t+n, that is, the predicted dust concentration spatiotemporal data sequence with the maximum probability P, n <J; S40. When sandstorm weather occurs, the existing actual sandstorm data based on satellite inversion is input into the sandstorm concentration prediction model obtained in step S30, and the sandstorm distribution results in the future set time period of the designated area are output to make a short-term forecast of sandstorms.
2. The method for short-term sandstorm forecasting based on deep learning according to claim 1 is characterized in that: The observation data of the ground station in step S10 includes the visibility and weather phenomenon data in the hourly full-factor data of the ground national meteorological station, and the PM10 data of the ground environmental protection station; the existing satellite-based inversion-based sand and dust real-time data is a sand and dust grid image containing sand and dust area and non-dust area marks in time series.
3. The method for short-term sandstorm forecasting based on deep learning according to claim 2 is characterized in that: The process of performing temporal and spatial matching based on the dust process data and the existing satellite-based inversion dust actual data in step S10 includes: S11. Collect observation data of ground stations during historical dust storms, and filter out time information including the start and end time of dust storms and spatial information including the latitude and longitude information of dust storm occurrence sites; S12, according to the time information obtained in step S11, screening the data containing the same time information in the actual dust data inverted based on the satellite, and completing the time matching of the data; S13. Under the condition of time matching, extract the inversion result data from the actual dust data based on satellite inversion according to the spatial information obtained in step S11, and judge whether the inversion result data is consistent with the observation data in the specified area. If so, include the actual dust data based on satellite inversion at this time and space into the initial sample.
4. The method for short-term sandstorm forecasting based on deep learning according to claim 3 is characterized in that: The process of preprocessing the data in step S20 to form a dust spatiotemporal sequence sample set is as follows: S21, performing regional cropping on the data matched in step S10, and forming a 256×512 grid after cropping; S22, using a sliding window mode to obtain a sample from the cropped data at every J+n consecutive moments, and cyclically obtain samples to form a sample set; S23. Use the maximum and minimum value normalization method to standardize each sample, assign the null value to 0, and obtain the dust spatiotemporal series sample set.
5. The method for short-term sandstorm forecasting based on deep learning according to claim 4 is characterized in that: The existing satellite-based inversion dust data in step S10 is in geotiff format, and the sample format of the dust spatiotemporal sequence sample set in step S20 is stored in npy format.
6. The method for short-term sandstorm forecasting based on deep learning according to claim 5 is characterized in that: The deep learning model of the spatiotemporal sequence constructed based on the PredRNN network in step S30 is a stacked network structure. The input dust image is converted from a single-channel image to a dust feature map with a latent space channel dimension D of W through a feature embedding layer. The dust feature map is converted into a single-channel dust prediction image output through stacked network structure learning and feature dimension reduction layer processing.
7. The method for short-term sandstorm forecasting based on deep learning according to claim 6 is characterized in that: The process of training and obtaining the dust concentration prediction model in step S30 includes: S31, initial parameter settings for model training: the number of model recurrent unit layers is set to 4, the hidden feature dimension between basic units is set to 64; the loss function is set to the MSE loss function, and the initial learning rate is set to 0.0001; the planned sampling stop round number is set to 5000; the filter size is set to 2; the AdamW optimizer is used to minimize the MSE loss function, and the learning rate is adjusted through its adaptive gradient mechanism and momentum gradient mechanism to speed up the model convergence; the maximum iteration round is set to 200; the input spatiotemporal sequence length is set to J, and the predicted spatiotemporal sequence length is set to n; S32, model training starts: using the sample data in the dust spatiotemporal sequence sample set as input, setting the input dust concentration image sequence X1-X J [J, 1, 256, 512], where the input sequence length is J, the number of grayscale image channels is 1, the image grid size is 256 × 512, and the model effect is verified once in each round of training and the model parameters are saved; S33, model learning: from the input dust concentration image sequence X1-X J [J,1,256,512] extracts a single dust concentration image X each time according to the time dimension i [1,1,256,512], i=1,2,…,J, input to the model, cycle this process J times, after feature embedding, stacked recurrent network and feature dimensionality reduction, save the feature map P output from the last cycle J+1 ; S34, model prediction: The feature map P output in step S33 is J+1 Then input it into the model and output the predicted dust concentration image P at time J+2 J+2 , and then use this predicted dust concentration image as the model input for the next cycle, and output the predicted dust concentration image P at time J+3 J+3 Repeat this process until the predicted dust concentration image P is obtained. J+n ; S35. Model tuning: Using the predicted dust concentration image P J+1 ~P J+n And the corresponding real dust concentration image X J+1 ~X J+n Calculate the MSE loss function and optimize the model parameters; S36. Repeat steps S33 to S35 to iteratively optimize the model parameters. After completing the maximum number of iterations, save the model parameters to obtain a dust concentration prediction model. During this period, observe the model learning status by drawing loss curves and evaluation index curves.
8. The method for short-term sandstorm forecasting based on deep learning according to claim 7 is characterized in that: The deep learning model of spatiotemporal sequence constructed based on the PredRNN network takes the ST-LSTM unit as the core of the network, including multiple layers of stacked neural units, which flow up and down between layers through spatiotemporal memory units, and are transmitted between layers through long-term memory units and hidden layers. t-1 , obtain the predicted dust concentration image at output time t It is expressed as: In the above formula, X t is the input dust concentration image at time t, the [] symbol represents the concatenation operation, and the * symbol represents the convolution operation. The symbol represents matrix multiplication, and They represent the long-term memory unit and spatiotemporal memory unit at the lth layer at time t, respectively. represents the value of the hidden layer at the lth layer at time t, g t 、i t and f t They represent the modulation gate, input gate and forget gate of the long-term memory unit C, respectively. t '、i t ' and f t 'represent a group of modulation gates, input gates and forget gates of the spatiotemporal memory unit M respectively. The newly added group of gates has the same formula as the traditional gate mechanism. t represents the output gate, tanh and σ are the corresponding activation functions, W xg 、Wxi、W xf , W hg , Whi, Whf, Wx'g, Wx'i, Wx'f, Wmg, Wmi, Wmf, Wxo, Who, Wco, and Wmo represent the weights of the corresponding convolution operations, bg, bi, bf, b'g, bi', b'f, and bo represent learnable offset parameters, and w1×1 represents the weights used to extract The weights of the 1×1 convolution operation on the spatial features of the concatenated feature map.
9. The method for short-term sandstorm forecasting based on deep learning according to claim 8, characterized in that: The input spatiotemporal sequence length is set to 6, and the predicted spatiotemporal sequence length is set to 3.
Citation Information
Patent Citations
Weather prediction model generation method and device and weather prediction method and device
CN112363251A
Geostationary satellite sand storm identification method based on deep learning
CN114220024A
Multi-feature fusion sand storm prediction method based on deep neural network
CN114882373A
Machine learning-based dust jacking inversion method for stationary satellite
CN117951525A
Weather forecasting systems and methods
US10613252B1
Cited By
Sand and dust early warning method, device and system based on satellite remote sensing
CN121580330A
Satellite remote sensing-based sand-dust early warning method, device and system
CN121580330B
Near-real-time sand and dust concentration reanalysis method and device
CN122221225A
