Short-term weather prediction method

Abstract

The invention relates to a short-term weather prediction method, which comprises the following steps: low-altitude radar inversion applying MsDR filtering: in a radar inversion process, a mobile spectrum depolarization filter is adopted to remove clutter signals and noise signals, MsDR uses a one-dimensional mobile window to select a spectrum band containing rainfall signals, and the spectrum band is used for filtering the rainfall signals; smoothing noise by applying a two-dimensional moving window on the distance-Doppler spectrogram; and designing a space-time fusion short-term and temporary rainfall prediction model, combining the spatial feature extraction capability of the CNN and the time sequence modeling capability of the Transform, and analyzing the space-time features of rainfall by the space-time fusion short-term and temporary rainfall prediction model at the same time. The mobile spectrum depolarization filter provides powerful support for data processing in radar inversion, and the accuracy and reliability of rainfall inversion are greatly improved. The CNN-Transformer not only can effectively analyze the current rainfall data, but also can provide higher accuracy and stability for rainfall prediction in different environments in the future, and the precision and reliability of short-term rainfall prediction are improved.

Description

Technical Field

[0001] The invention relates to the technical field of meteorological forecasting, and in particular to a short-term meteorological forecasting method. Background Art

[0002] Short-term rainfall is usually characterized by strong suddenness and wide coverage, which can easily cause serious natural disasters such as mountain torrents and mudslides, threatening human life and property. Research on low-altitude short-term radar inversion and rainfall prediction technology can provide a more reliable scientific basis for disaster prevention and mitigation, thereby reducing the risk of sudden disasters. Radar echo signals are affected by ground clutter, buildings, vegetation and other factors, making it difficult to accurately invert rainfall. At the same time, the existing prediction models lack consideration of the temporal and spatial complexity of rainfall in the monitoring area when dealing with the dynamic changes in rainfall. Summary of the invention

[0003] The purpose of the present invention is to provide a short-term weather forecasting method to solve the technical problems mentioned in the background technology.

[0004] The technical solution adopted by the present invention to solve its technical problem is:

[0005] A short-term meteorological forecasting method comprises the following steps:

[0006] S1. Low-altitude radar inversion using MsDR filtering: In the radar inversion process, the target signal, clutter signal and noise signal are mixed in the range-Doppler spectrum to form the original radar echo signal. The moving spectrum depolarization filter is used to remove the clutter signal and the noise signal. MsDR uses a one-dimensional moving window to select the spectrum segment containing the rainfall signal, and applies a two-dimensional moving window to smooth the noise on the range-Doppler spectrum.

[0007] S2. Design a spatiotemporal fusion short-term rainfall prediction model. Combining the spatial feature extraction capability of CNN and the time series modeling capability of Transformer, the spatiotemporal fusion short-term rainfall prediction model simultaneously analyzes the spatiotemporal characteristics of rainfall.

[0008] As a further technical solution of the present invention, step S1 includes the following steps:

[0009] In the radar inversion process, the target signal, clutter signal and noise signal are mixed in the range-Doppler spectrum to form the original radar echo signal. MsDR combines the spectral differential reflectivity sZDR and the spectral co-polarization correlation coefficient sρ co To characterize the physical characteristics of the target, accurately analyze the polarization characteristics of the echo signal, thereby identifying and removing clutter and retaining important meteorological signals. The principle is shown in Formula 1:

[0010] Formula 1:

[0011] Where sZDR(r,v)=10log 10 (sZ hh (r,v) / sZ vv (r,v)), reflecting the horizontal polarization (hh) signal sZ hh (r,v) and vertically polarized (vv) signals sZ vv The reflectivity ratio (r,v) reveals the shape and arrangement of rainfall particles. The correlation between horizontally polarized and vertically polarized echo signals is measured to help determine the shape and distribution of rainfall particles;

[0012] Subsequently, MsDR uses a one-dimensional moving window to select the spectrum segment containing the rainfall signal, effectively separating the rainfall information and reducing clutter interference. Next, a two-dimensional moving window is applied to the range-Doppler spectrum to smooth the data in the local area to reduce the impact of noise and effectively identify the rainfall characteristic pattern. This process is used to recover the rainfall signal that may be lost and remove the remaining clutter. MsDR improves the extraction and reconstruction accuracy of the rainfall signal through the above steps. After the above processing, the calculation process of the reflectivity factor at the distance e is shown in Formula 2:

[0013] Formula 2:

[0014] Among them, sP hh (r,v) represents the spectral power of the horizontal polarization echo, sN is the noise power, C is the radar system constant, and the reflectivity factor Z of the horizontal polarization echo hh (r) is an important input for rainfall estimation, which directly affects the measurement of rainfall intensity and the accuracy of weather forecast. After obtaining the reflectivity factor Z, the two-parameter ZR relationship R = a·Z b ·e c·Z To estimate the rainfall at the current moment, a, b and c are parameters determined by empirical data. The two-parameter ZR relationship helps to improve the estimation accuracy of rainfall.

[0015] As a further technical solution of the present invention, step S2 includes the following steps:

[0016] Firstly, the prediction area is subdivided into intelligent grids with a resolution of p km. Each grid contains its independent geographic coordinates, meteorological characteristics and historical rainfall data. The meteorological data in each grid area includes the meteorological parameters and rainfall time series data obtained by radar inversion. The meteorological parameters include temperature, humidity and wind speed. Then, the rainfall time series in each grid is decomposed by CEEMDAN using the meteorological parameters and rainfall time series data obtained by radar inversion. CEEMDAN adds several white noises with different standard deviations to the input rainfall time series data f(t) to obtain multiple sets of data f(t). i (t), then, for each set of data f i (t) Perform empirical mode decomposition (EMD) to obtain a set of intrinsic mode function (IMF) components. i The IMF components decomposed from the rainfall time series data f(t) are averaged to obtain the first IMF component imf1(t). The first residual r1(t) is obtained by subtracting imf1(t) from the rainfall time series data f(t) to remove the high-frequency components, so that the remaining signal is more concentrated in the low-frequency part. Then white noise is added to the residual r1(t) and the decomposition process is repeated. The process of obtaining the kth residual signal is shown in Formula 3:

[0017] Formula 3:

[0018] Where w(t) is a white noise signal, ε k-1 represents the noise amplitude factor in the k-1th iteration, EMD(·) represents the empirical mode decomposition operator, n is the number of times white noise is added, r k-1 (t) is the residual signal of the k-1th iteration, that is, the input signal of the kth iteration, r k (t) is the kth residual signal, which means the remaining signal after the kth iteration, that is, the remaining part of the signal after removing the first k modes. Through multiple iterative decompositions, CEEMDAN decomposes the original signal into several intrinsic mode functions IMFs and a residual term;

[0019] In rainfall prediction, IMF can help distinguish different frequency components in rainfall signals, thereby improving the accuracy and stability of the prediction. The obtained IMF components are then input into the CNN-Transformer deep learning model. First, CNN extracts spatial features and patterns in local areas through convolution operations, which can effectively capture the spatial structure and local characteristics of rainfall data; next, the Transformer part processes these spatial feature sequences and uses the self-attention mechanism to capture long-term dependencies in time series, thereby learning the temporal patterns and trends in rainfall data. Finally, the model performs rolling predictions with a time step of m minutes. Each iteration predicts rainfall conditions in multiple time steps in the future. The model considers the previous rainfall characteristics and current meteorological conditions in each prediction, and further optimizes the prediction results through adaptive learning. The prediction results of each grid are integrated into the overall space based on its independent geographic coordinates to generate a rainfall distribution map with high spatiotemporal resolution.

[0020] The beneficial effects of the present invention are:

[0021] In the present invention, the mobile spectral depolarization filter provides strong support for data processing in radar inversion by eliminating interference, enhancing target signals and optimizing rainfall feature extraction, greatly improving the accuracy and reliability of rainfall inversion, and providing reliable technical guarantee for obtaining time series data of key meteorological parameters and rainfall.

[0022] CNN-Transformer can not only effectively analyze current rainfall data, but also provide higher accuracy and stability for rainfall forecasts under different environments in the future, providing reliable technical support for rainfall forecasts. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 A diagram of a low-altitude radar inversion process using MsDR filtering in a short-term weather forecasting method provided in an embodiment of the present invention.

[0024] Figure 2 A diagram of a short-term rainfall prediction model that integrates time and space in a short-term meteorological prediction method provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0025] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of 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.

[0026] See also Figure 1 to Figure 2 , an embodiment of the present invention provides a short-term weather forecasting method, comprising the following steps;

[0027] S1. Low-altitude radar inversion using MsDR filtering: In the radar inversion process, the target signal, clutter signal and noise signal are mixed in the range-Doppler spectrum to form the original radar echo signal. The moving spectrum depolarization filter is used to remove the clutter signal and the noise signal. MsDR uses a one-dimensional moving window to select the spectrum segment containing the rainfall signal, and applies a two-dimensional moving window to smooth the noise on the range-Doppler spectrum.

[0028] In the radar inversion process, the target signal, clutter signal and noise signal are mixed in the range-Doppler spectrum to form the original radar echo signal. In order to obtain more accurate meteorological data, it is necessary to suppress the clutter signal in the radar echo. Figure 1 The complete inversion process of low-altitude radar is shown. MsDR is obtained by combining the spectral differential reflectivity sZDR and the spectral co-polarization correlation coefficient sρ co To characterize the physical characteristics of the target, accurately analyze the polarization characteristics of the echo signal, thereby identifying and removing clutter and retaining important meteorological signals. The principle is shown in Formula 1:

[0029] Formula 1:

[0030] Where sZDR(r,v)=10log 10 (sZ hh (r,v) / sZ vv (r,v)), reflecting the horizontal polarization (hh) signal sZ hh (r,v) and vertically polarized (vv) signals sZ vv The reflectivity ratio (r,v) reveals the shape and arrangement of rainfall particles. The correlation between horizontally polarized and vertically polarized echo signals is measured to help determine the shape and distribution of rainfall particles;

[0031] Subsequently, MsDR uses a one-dimensional moving window to select the spectrum segment containing the rainfall signal, effectively separating the rainfall information and reducing clutter interference. Next, a two-dimensional moving window is applied to the range-Doppler spectrum to smooth the data in the local area to reduce the impact of noise and effectively identify the rainfall characteristic pattern. This process is used to recover the rainfall signal that may be lost and remove the remaining clutter. MsDR improves the extraction and reconstruction accuracy of the rainfall signal through the above steps. After the above processing, the calculation process of the reflectivity factor at a distance of r is shown in Formula 2:

[0032] Formula 2:

[0033] Among them, sP hh(r,v) represents the spectral power of the horizontal polarization echo, sN is the noise power, C is the radar system constant, and the reflectivity factor Z of the horizontal polarization echo hh (r) is an important input for rainfall estimation, which directly affects the measurement of rainfall intensity and the accuracy of weather forecast. After obtaining the reflectivity factor Z, the two-parameter ZR relationship R = a·Z b ·e c·Z To estimate the rainfall at the current moment, a, b and c are parameters determined by empirical data. The two-parameter ZR relationship helps to improve the estimation accuracy of rainfall.

[0034] S2. Design a spatiotemporal fusion short-term rainfall prediction model. Combining the spatial feature extraction capability of CNN and the time series modeling capability of Transformer, the spatiotemporal fusion short-term rainfall prediction model simultaneously analyzes the spatiotemporal characteristics of rainfall.

[0035] The present invention designs a spatiotemporal fusion short-term rainfall prediction model (CEEMDAN-CNN-Transformer) to solve the spatiotemporal heterogeneity problem of short-term rainfall. The model architecture is as follows Figure 2 First, the prediction area is subdivided into intelligent grids with a resolution of p km. Each grid contains its independent geographic coordinates, meteorological characteristics and historical rainfall data. The meteorological data in each grid area includes meteorological parameters obtained by radar inversion and time series data of rainfall. The meteorological parameters include temperature, humidity and wind speed. Then, the rainfall time series in each grid is decomposed by CEEMDAN using the meteorological parameters and rainfall time series data obtained by radar inversion. CEEMDAN adds several white noises with different standard deviations to the input rainfall time series data f(t) to obtain multiple sets of data f(t). i (t), then, for each set of data f i (t) Perform empirical mode decomposition (EMD) to obtain a set of intrinsic mode function (IMF) components. i The IMF components decomposed from the rainfall time series data f(t) are averaged to obtain the first IMF component imf1(t). The first residual r1(t) is obtained by subtracting imf1(t) from the rainfall time series data f(t) to remove the high-frequency components, so that the remaining signal is more concentrated in the low-frequency part. Then white noise is added to the residual r1(t) and the decomposition process is repeated. The process of obtaining the kth residual signal is shown in Formula 3:

[0036] Formula 3:

[0037] Where w(t) is a white noise signal, ε k-1 represents the noise amplitude factor in the k-1th iteration, EMD(·) represents the empirical mode decomposition operator, n is the number of times white noise is added, rk-1 (t) is the residual signal of the k-1th iteration, that is, the input signal of the kth iteration, r k (t) is the kth residual signal, which means the remaining signal after the kth iteration, that is, the remaining part of the signal after removing the first k modes. Through multiple iterative decompositions, CEEMDAN decomposes the original signal into several intrinsic mode functions IMFs and a residual term;

[0038] In rainfall prediction, IMF can help distinguish different frequency components in rainfall signals, thereby improving the accuracy and stability of prediction. The obtained IMF components are then input into the CNN-Transformer deep learning model. First, CNN extracts spatial features and patterns in local areas through convolution operations, which can effectively capture the spatial structure and local features of rainfall data. Next, the Transformer part processes these spatial feature sequences and uses the self-attention mechanism to capture long-term dependencies in time series, thereby learning the temporal patterns and trends in rainfall data. Finally, the model performs rolling predictions with a time step of m minutes. Each iteration predicts the rainfall conditions in multiple time steps in the future. The model considers the previous rainfall characteristics and current meteorological conditions in each prediction, and further optimizes the prediction results through adaptive learning. The prediction results of each grid are integrated into the overall space based on its independent geographic coordinates to generate a rainfall distribution map with high spatiotemporal resolution. This m-minute rolling prediction method ensures accurate predictions on the time axis, thereby greatly improving the real-time and accuracy of the prediction. Through gridded rainfall prediction, the model can provide detailed rainfall assessments for different regions, making the rainfall intensity changes in local areas more controllable and visualized.

[0039] In the radar inversion process, the target signal, clutter signal and noise signal are mixed in the range-Doppler spectrum to form the original radar echo signal. This signal mixing often leads to a lot of interference, affecting the accuracy of rainfall inversion. To solve this problem, the mobile spectrum depolarization filter plays a vital role. First, it can effectively remove clutter and noise signals and retain the target signal related to rainfall, thereby improving the accuracy of radar inversion. By processing the spectral information in the range-Doppler spectrum, the filter can highlight the low-frequency characteristics of the rainfall signal and reduce the interference of high-frequency noise, helping to better extract key information such as the spatial distribution, intensity and dynamic changes of rainfall, and provide a more accurate data basis for subsequent rainfall inversion. In addition, radar inversion can not only obtain time series data of rainfall, but also obtain key meteorological parameters such as temperature, humidity, and wind speed, which are crucial for comprehensive monitoring and prediction of weather conditions. Since the polarization characteristics reflect the shape, size and distribution of raindrops, the mobile spectrum depolarization filter can more accurately analyze the polarization characteristics of rainfall by reducing polarization noise, further improving the accuracy and stability of radar inversion. Its real-time dynamic processing capability enables it to quickly capture changes in the range-Doppler spectrum, ensuring that accurate rainfall forecast data is quickly generated under complex weather conditions.

[0040] CNN-Transformer plays a key role in analyzing rainfall time series. It combines the advantages of convolutional neural network (CNN) and Transformer and can effectively capture the spatiotemporal characteristics of rainfall data. First, CNN is good at extracting local spatial features in data. In rainfall time series analysis, the distribution of rainfall has spatial dependence. CNN can extract these local spatial information through convolution operations, thereby helping the model to better understand the distribution of rainfall in different regions. However, CNN has limitations in capturing long-range dependencies, while Transformer can effectively model long-range dependencies in the time dimension through the self-attention mechanism. In rainfall time series, rainfall changes not only depend on current weather conditions, but are also closely related to meteorological data over a period of time. Transformer's self-attention mechanism enables the model to capture this long-range dependency in time, thereby more accurately predicting future rainfall trends. By combining CNN's spatial feature extraction capabilities and Transformer's time series modeling capabilities, the CNN-Transformer model can simultaneously analyze the spatiotemporal characteristics of rainfall and improve the accuracy of rainfall prediction. In addition, the complexity of rainfall time series is reflected in the fact that it is affected by factors such as season and geographical location, and rainfall intensity, frequency, and duration will fluctuate significantly. The multi-layer structure of CNN-Transformer can extract features at different levels layer by layer and adapt to data changes in different environments, thereby improving the generalization ability of the model.

[0041] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or device including the element.

[0042] The above are only preferred embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A short-term weather forecasting method, characterized in that: The steps include: S1. Low-altitude radar inversion using MsDR filtering: In the radar inversion process, the target signal, clutter signal and noise signal are mixed in the range-Doppler spectrum to form the original radar echo signal. The moving spectrum depolarization filter is used to remove the clutter signal and the noise signal. MsDR uses a one-dimensional moving window to select the spectrum segment containing the rainfall signal, and applies a two-dimensional moving window to smooth the noise on the range-Doppler spectrum. S2. Design a spatiotemporal fusion short-term rainfall prediction model. Combining the spatial feature extraction capability of CNN and the time series modeling capability of Transformer, the spatiotemporal fusion short-term rainfall prediction model simultaneously analyzes the spatiotemporal characteristics of rainfall.

2. A short-term weather forecasting method according to claim 1, characterized in that: Step S1 includes the following steps: In the radar inversion process, the target signal, clutter signal and noise signal are mixed in the range-Doppler spectrum to form the original radar echo signal. MsDR combines the spectral differential reflectivity sZDR and the spectral co-polarization correlation coefficient sρ co To characterize the physical characteristics of the target, accurately analyze the polarization characteristics of the echo signal, thereby identifying and removing clutter and retaining the meteorological signal. The principle is shown in Formula 1: Formula 1: Where sZDR(r,v)=10log 10 (sZ hh (r,v) / sZ vv (r,v)), reflects the horizontal polarization signal sZ hh (r,v) and vertical polarization signal sZ vv The reflectivity ratio (r,v) reveals the shape and arrangement of rainfall particles. The correlation between horizontally polarized and vertically polarized echo signals is measured to help determine the shape and distribution of rainfall particles; Subsequently, MsDR uses a one-dimensional moving window to select the spectrum segment containing the rainfall signal, effectively separating the rainfall information and reducing clutter interference. Next, a two-dimensional moving window is applied to the range-Doppler spectrum to smooth the data in the local area to reduce the influence of noise and effectively identify the rainfall characteristic pattern. The calculation process of the reflectivity factor at a distance of r is shown in Formula 2: Formula 2: Among them, sP hh (r,v) represents the spectral power of the horizontal polarization echo, sN is the noise power, C is the radar system constant, and the reflectivity factor Z of the horizontal polarization echo hh (r) is an important input for rainfall estimation, which directly affects the measurement of rainfall intensity and the accuracy of weather forecast. After obtaining the reflectivity factor Z, the two-parameter ZR relationship R = a·Z b ·e c·Z To estimate the rainfall at the current moment, a, b and c are parameters determined by empirical data. The two-parameter ZR relationship helps to improve the estimation accuracy of rainfall.

3. A short-term weather forecasting method according to claim 1, characterized in that: Step S2 The following steps are involved: Firstly, the prediction area is subdivided into intelligent grids with a resolution of pkm. Each grid contains its independent geographic coordinates, meteorological characteristics and historical rainfall data. The meteorological data in each grid area includes the meteorological parameters obtained by radar inversion and the time series data of rainfall. The meteorological parameters include temperature, humidity and wind speed. Then, the rainfall time series in each grid is decomposed by CEEMDAN using the meteorological parameters and rainfall time series data obtained by radar inversion. CEEMDAN adds several white noises with different standard deviations to the input rainfall time series data f(t) to obtain multiple sets of data f(t). i (t), then, for each set of data f i (t) Perform empirical mode decomposition (EMD) to obtain a set of intrinsic mode function (IMF) components. i The IMF components decomposed from the rainfall time series data f(t) are averaged to obtain the first IMF component imf1(t). The first residual r1(t) is obtained by subtracting imf1(t) from the rainfall time series data f(t) to remove the high-frequency components, so that the remaining signal is more concentrated in the low-frequency part. Then white noise is added to the residual r1(t) and the decomposition process is repeated. The process of obtaining the kth residual signal is shown in Formula 3: Formula 3: Where w(t) is a white noise signal, ε k-1 represents the noise amplitude factor in the k-1th iteration, EMD(·) represents the empirical mode decomposition operator, n is the number of times white noise is added, r k-1 (t) is the residual signal of the k-1th iteration, that is, the input signal of the kth iteration, r k (t) is the kth residual signal, which means the remaining signal after the kth iteration, that is, the remaining part of the signal after removing the first k modes. Through multiple iterative decompositions, CEEMDAN decomposes the original signal into several intrinsic mode functions IMFs and a residual term; In rainfall prediction, the obtained IMF components are then input into the CNN-Transformer deep learning model. First, CNN extracts spatial features and patterns in local areas through convolution operations, effectively capturing the spatial structure and local features of rainfall data. Next, the Transformer part processes these spatial feature sequences and uses the self-attention mechanism to capture long-term dependencies in time series, thereby learning the temporal patterns and trends in rainfall data. Finally, the model performs rolling predictions with a time step of m minutes. Each iteration predicts rainfall conditions in multiple time steps in the future. The model considers previous rainfall characteristics and current meteorological conditions in each prediction, and further optimizes the prediction results through adaptive learning. The prediction results of each grid are integrated into the overall space based on its independent geographic coordinates to generate a rainfall distribution map with high spatiotemporal resolution.

Images