Rainfall prediction system based on laser radar
Through the lidar-based rainfall prediction system, which integrates the spatiotemporal prediction module and the full-dimensional replacement hybrid module, the problems of large computational complexity and insufficient accuracy in the existing technology are solved, and efficient and accurate rainfall prediction is achieved.
Patent Information
- Application Number
- CN202510781552.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-12
- Publication Date
- 2025-09-23
AI Technical Summary
Existing rainfall prediction methods have deficiencies in computational complexity and accuracy. Traditional numerical weather forecasts are sensitive to initial conditions and require large amounts of computation, while deep learning methods are highly complex and computationally expensive. In addition, lidar is insufficiently used in the field of rainfall prediction.
A rainfall prediction system based on lidar is adopted. By integrating the spatiotemporal prediction module and the full-dimensional substitution mixing module and combining the dynamic weighting mechanism, efficient fusion and decoding of lidar multi-channel physical data are achieved. Dual memory stream spatiotemporal modeling and substitution mixing mechanism are used to capture the physical transmission chain from changes in high-altitude extinction coefficient to ground rainfall, and predictions are made in combination with the laws of atmospheric dynamics.
It achieves high-precision rainfall forecasts, reduces computing costs, improves the model's adaptability and forecast accuracy, and complies with the laws of atmospheric dynamics.
Smart Images
Figure CN120686380A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of meteorological forecasting, and in particular to a rainfall forecasting system based on laser radar. Background Art
[0002] Nowcasting plays a vital role in meteorological forecasting. It is a forecast of rainfall within the next few hours. In recent years, although the nowcasting technology based on Doppler radar has been widely used, especially the deep learning method using convolutional recurrent neural network architecture, the application research of lidar in the field of rainfall forecasting is still insufficient. In recent years, climate anomalies and extreme weather have increased, including heavy rainfall in the short term. It poses a major threat to many human activities such as transportation, agricultural production and tourism. Therefore, precipitation nowcasting technology is particularly important. Precipitation nowcasting refers to the task of predicting future rainfall information in a specific area in a relatively short period of time based on radar observations;
[0003] Existing rainfall prediction methods are mainly divided into traditional numerical weather prediction (NWP) and radar echo pattern extrapolation. NWP is based on atmospheric physics equations and can simulate atmospheric dynamics and thermodynamic processes under given initial conditions and boundaries. 2 However, it is extremely sensitive to initial conditions, and small errors may lead to significant prediction deviations. At the same time, the computational complexity is huge, making it unsuitable for frequently updated near-forecasting tasks. The radar echo map extrapolation method uses Doppler weather radar detection data and uses optical flow method or deep learning technology to predict future echo map changes. The optical flow method assumes that the image pixel values remain unchanged and does not consider changes in rainfall intensity, which is not in line with the actual situation. The most popular deep learning methods currently include recursive neural networks, such as ConvLSTM. 5 and pure convolutional neural networks. This approach enables short-term nowcasting with significantly less computational effort than NWP. However, it also has limitations. For example, to achieve high-quality forecasts, the model becomes increasingly complex, leading to increased computational costs. Therefore, this paper proposes a lidar-based rainfall prediction system to address the challenges of existing technologies. Summary of the Invention
[0004] In response to the above problems, the purpose of the present invention is to propose a rainfall prediction system based on lidar. The rainfall prediction system based on lidar realizes efficient fusion and decoding of lidar multi-channel physical data by integrating the spatiotemporal sequence processing capability of the spatiotemporal prediction module with the improved full-dimensional permutation hybrid module combined with a dynamic weighting mechanism. The dominant role of the extinction coefficient is enhanced by adaptively allocating channel weights through the dynamic weighting module. The physical conduction chain of "sudden increase in high-altitude extinction coefficient → low-altitude humidity saturation → ground rainfall" is explicitly captured through dual-memory flow-based spatiotemporal modeling and permutation hybrid mechanism. The model decision is consistent with the laws of atmospheric dynamics.
[0005] To achieve the purpose of the present invention, the present invention is implemented through the following technical solutions: a rainfall prediction system based on lidar, including a data preprocessing module, a spatiotemporal prediction module, a full-dimensional permutation hybrid module and a training module, the data preprocessing module is used to perform physical calibration and spatial reshaping of the aerosol extinction coefficient and humidity data collected by the lidar, the data preprocessing module includes a physical calibration module and a spatial reshaping module, the physical calibration module is used to return the negative value of the aerosol extinction coefficient to zero, and perform forward filling on the missing data at a single moment, the spatial reshaping module is used to use normalization processing to unify the dimensions of different features to between 0 and 1, the spatiotemporal prediction module is used to use a spatiotemporal memory flow with memory decoupling to extract the spatiotemporal dependency features of the data to form a two-dimensional spatial feature map, the full-dimensional permutation hybrid module is used to fuse multi-channel physical data through a dynamic weighting mechanism and output the rainfall probability, and the training module is used to convert the original meteorological data into high-precision rainfall prediction capabilities through a multi-stage joint optimization strategy.
[0006] A further improvement is that the output dimension of the spatial reshaping module is T×8×22×22 (T=12), and the convolution dimension increase adopts a 3×3 convolution kernel.
[0007] Further improvements are as follows: the spatiotemporal prediction module includes a time memory module and a space memory module. The time memory module is used to use a spatiotemporal memory flow with memory decoupling to extract the spatiotemporal dependency features of the data to form a two-dimensional spatial feature map. The space memory module is used to convert the profile data of the lidar into a learnable spatial evolution law through a dynamic memory flow in the vertical height dimension.
[0008] Further improvements are: the full-dimensional permutation mixing module includes a feature slice embedding layer module, a permutation mixing multi-layer perception module and a dynamic weighting module, the feature slice embedding layer module is used to convert the two-dimensional spatial feature map output by the spatiotemporal prediction module into a structured fragment sequence, providing an input basis for subsequent channel mixing and fragment mixing, the permutation mixing multi-layer perception module is used to realize cross-dimensional information interaction by alternately executing feature mixing of channel mixing paths and fragment mixing paths, and the dynamic weighting module is used to dynamically allocate feature importance known by physical laws.
[0009] A further improvement is that the output layer of the dynamic weighting module maps the fusion features to interval probability values of [0,1] through linear projection.
[0010] A further improvement is that the weighting formula used by the dynamic weighting module is:
[0011]
[0012]
[0013] Among them, σ represents the GELU activation function, W i , i∈{1,2,3,4} is the weight of the fully connected layer, C and S represent the dimensions of the image after patch embedding, X is the input data, is the final output of the model.
[0014] A further improvement is that the training module includes a function encoding module and a multi-stage joint optimization module, the function encoding module is used to encode atmospheric dynamics knowledge into the neural network, and the multi-stage joint optimization module is used to embed atmospheric dynamics equations into the neural network in a differentiable form.
[0015] A further improvement is that the loss function in the function encoding module uses Binary CrossEntropyLoss, and the formula is:
[0016]
[0017] Where N is the number of samples, y i is the i-th sample label, p i is the probability that the model predicts that the i-th sample is a positive class.
[0018] The beneficial effects of the present invention are as follows: the present invention realizes efficient fusion and decoding of multi-channel physical data of lidar by integrating the spatiotemporal sequence processing capability of the spatiotemporal prediction module with the improved full-dimensional permutation hybrid module in combination with the dynamic weighting mechanism, and adaptively allocates channel weights through the dynamic weighting module, thereby strengthening the dominant role of the extinction coefficient. By based on dual-memory flow spatiotemporal modeling and permutation hybrid mechanism, the physical conduction chain of "sudden increase in high-altitude extinction coefficient → low-altitude humidity saturation → ground rainfall" is explicitly captured, and the model decision is consistent with the laws of atmospheric dynamics. BRIEF DESCRIPTION OF THE DRAWINGS
[0019] Figure 1 This is a system architecture diagram of the present invention;
[0020] Figure 2 This is a visualization diagram of the model result prediction of the present invention. DETAILED DESCRIPTION
[0021] In order to deepen the understanding of the present invention, the present invention will be further described in detail below with reference to the examples. The examples are only used to explain the present invention and do not constitute a limitation on the scope of protection of the present invention.
[0022] according to Figure 1 、 Figure 2 As shown, this embodiment provides a rainfall prediction system based on lidar, including a data preprocessing module, a spatiotemporal prediction module, a full-dimensional permutation hybrid module and a training module. The data preprocessing module is used to perform physical calibration and spatial reshaping of the aerosol extinction coefficient and humidity data collected by the lidar. The data preprocessing module includes a physical calibration module and a spatial reshaping module. The physical calibration module is used to reset the negative value of the aerosol extinction coefficient to zero and perform forward filling for missing data at a single moment, thereby ensuring temporal continuity and reflecting the trend of smooth changes in humidity and aerosol concentration in the atmosphere, thereby maintaining the physical integrity of the data. For missing data segments at multiple consecutive moments, they will be discarded because long-term absences may cause the spatiotemporal dependency to fail, so as to ensure that the data has consistent physical meaning to the model. The extinction coefficient of aerosol refers to the absorption and scattering of light by aerosol particles within each kilometer of air. Aerosols act as cloud condensation nuclei and affect the formation, size and distribution of cloud droplets, thereby having a direct or indirect impact on precipitation22,23, with units of km -1 Its extinction coefficient exhibits significant spatiotemporal dependence: high values at high altitudes indicate cloud formation at that altitude, a precursor to rain. As clouds descend to lower altitudes, the high extinction coefficient gradually shifts to the lower layers, signaling the onset of rain. This spatiotemporal dependence reveals the dynamic evolution of aerosol distribution and changes at different altitudes and time points, providing important insights into rainfall prediction.
[0023] The spatial reshaping module uses normalization to normalize the dimensions of different features to between 0 and 1. The output dimensions of the spatial reshaping module are T×8×22×22 (T=12), and the convolutional kernel is used for dimensionality enhancement. The original data is of the shape T×C×L, where T=12 represents 12 timestamps, C=2 represents the two channels of extinction and humidity, and L=496 represents the data length. Because the data contains two different physical meanings and their values and dimensions differ significantly, normalization is required to normalize the dimensions of the different features to between 0 and 1. The spatiotemporal prediction module uses a spatiotemporal memory flow with memory decoupling to extract the spatiotemporal dependence of the data and form a two-dimensional spatial feature map. The full-dimensional permutation hybrid module fuses multi-channel physical data through a dynamic weighting mechanism and outputs a rainfall probability. The training module uses a multi-stage joint optimization strategy to transform the raw meteorological data into a high-precision rainfall prediction capability.
[0024] The spatiotemporal prediction module includes a time memory module and a space memory module. The time memory module is used to use a spatiotemporal memory flow with memory decoupling to extract the spatiotemporal dependency features of the data to form a two-dimensional spatial feature map. The spatial memory module is used to convert the lidar profile data into a learnable spatial evolution law through a dynamic memory flow in the vertical height dimension.
[0025] The full-dimensional permutation mixing module includes a feature slice embedding layer module, a permutation mixing multi-layer perception module and a dynamic weighting module. The feature slice embedding layer module is used to convert the two-dimensional spatial feature map output by the spatiotemporal prediction module into a structured fragment sequence, providing an input basis for subsequent channel mixing and fragment mixing. The permutation mixing multi-layer perception module is used to realize cross-dimensional information interaction by alternately executing feature mixing of channel mixing paths and fragment mixing paths. The dynamic weighting module is used to dynamically allocate the importance of features known by physical laws. The output layer of the dynamic weighting module maps the fused features to interval probability values of [0,1] through linear projection.
[0026] The weighting formula used by the dynamic weighting module is:
[0027]
[0028]
[0029] Among them, σ represents the GELU activation function, W i , i∈{1,2,3,4} is the weight of the fully connected layer, C and S represent the dimensions of the image after patch embedding, X is the input data, is the final output of the model.
[0030] The training module includes a function encoding module and a multi-stage joint optimization module. The function encoding module is used to encode atmospheric dynamics knowledge into the neural network. The loss function in the function encoding module uses Binary CrossEntropy Loss, and the formula is:
[0031]
[0032] Where N is the number of samples, y i is the i-th sample label, p i is the probability that the model predicts that the i-th sample is a positive class.
[0033] A multi-stage joint optimization module is used to embed the atmospheric dynamics equations into the neural network in a differentiable form.
[0034] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the foregoing embodiments. The foregoing embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.
Claims
1. A rainfall prediction system based on laser radar, characterized by: It includes a data preprocessing module, a spatiotemporal prediction module, a full-dimensional permutation hybrid module and a training module. The data preprocessing module is used to physically calibrate and spatially reshape the aerosol extinction coefficient and humidity data collected by the lidar. The data preprocessing module includes a physical calibration module and a spatial reshaping module. The physical calibration module is used to return the negative value of the aerosol extinction coefficient to zero and perform forward filling on the missing data at a single moment. The spatial reshaping module is used to use normalization processing to unify the dimensions of different features to between 0 and 1. The spatiotemporal prediction module is used to use a spatiotemporal memory flow with memory decoupling to extract the spatiotemporal dependency features of the data to form a two-dimensional spatial feature map. The full-dimensional permutation hybrid module is used to fuse multi-channel physical data through a dynamic weighting mechanism and output the rainfall probability. The training module is used to convert the original meteorological data into high-precision rainfall prediction capabilities through a multi-stage joint optimization strategy.
2. The laser radar-based rainfall prediction system according to claim 1, characterized in that: The output dimension of the spatial reshaping module is T×8×22×22 (T=12), and the convolution dimension is increased using a 3×3 convolution kernel.
3. The laser radar-based rainfall prediction system according to claim 1, characterized in that: The spatiotemporal prediction module includes a time memory module and a space memory module. The time memory module is used to use a spatiotemporal memory flow with memory decoupling to extract the spatiotemporal dependency features of the data to form a two-dimensional spatial feature map. The space memory module is used to convert the profile data of the lidar into a learnable spatial evolution law through a dynamic memory flow in the vertical height dimension.
4. The laser radar-based rainfall prediction system according to claim 1, characterized in that: The full-dimensional permutation mixing module includes a feature slice embedding layer module, a permutation mixing multi-layer perception module and a dynamic weighting module. The feature slice embedding layer module is used to convert the two-dimensional spatial feature map output by the spatiotemporal prediction module into a structured fragment sequence, providing an input basis for subsequent channel mixing and fragment mixing. The permutation mixing multi-layer perception module is used to realize cross-dimensional information interaction by alternately executing feature mixing of channel mixing paths and fragment mixing paths. The dynamic weighting module is used to dynamically allocate feature importance known by physical laws.
5. The laser radar-based rainfall prediction system according to claim 4, characterized in that: The output layer of the dynamic weighting module maps the fusion features to interval probability values of [0, 1] through linear projection.
6. The laser radar-based rainfall prediction system according to claim 4, characterized in that: The weighting formula used by the dynamic weighting module is: Among them, σ represents the GELU activation function, W i , i∈{1,2,3,4} is the weight of the fully connected layer, C and S represent the dimensions of the image after patch embedding, X is the input data, is the final output of the model.
7. The laser radar-based rainfall prediction system according to claim 1, characterized in that: The training module includes a function encoding module and a multi-stage joint optimization module. The function encoding module is used to encode atmospheric dynamics knowledge into the neural network, and the multi-stage joint optimization module is used to embed atmospheric dynamics equations in a differentiable form into the neural network.
8. The laser radar-based rainfall prediction system according to claim 7, characterized in that: The loss function in the function encoding module uses Binary CrossEntropy Loss, and the formula is: Where N is the number of samples, y i is the i-th sample label, p i is the probability that the model predicts that the i-th sample is a positive class.