Discrete station rainfall short-term forecasting method based on convolution and dimension reduction deconvolution neural network

Abstract

The invention discloses a discrete station rainfall short-term forecasting method based on a convolution and dimensionality reduction deconvolution neural network, and belongs to the application field of artificial intelligence in short-term rainfall forecasting in meteorology and related industries. According to the method, a data set is constructed through a normalized multivariable two-dimensional forecasting factor and discrete site rainfall data, and through multilayer multi-kernel two-dimensional convolution, attention mechanism fusion, flattening processing, one-dimensional deconvolution up-sampling and loss function optimization, future multi-day rainfall accurate forecasting of multiple discrete sites is realized. The method can automatically learn a complex nonlinear relationship, effectively extract multi-scale meteorological features, solve the problem of'field-point 'forecast, meet the requirement of service timeliness, and is good in portability and wide in popularization and application prospect.

Description

Technical Field

[0001] This invention belongs to the technical field of applying artificial intelligence technology to meteorology and related industries for short-term precipitation forecasting, specifically involving a method for short-term precipitation forecasting at discrete stations based on convolutional and dimensionality-reduced deconvolutional neural networks. Background Technology

[0002] Short-term quantitative precipitation forecasting is one of the core tasks in meteorological operations, and it is of great significance for disaster prevention and mitigation, water resource management, and agricultural production. Currently, short-term quantitative precipitation forecasting techniques and methods are mainly divided into three categories: numerical weather prediction and ensemble forecasting methods, numerical forecast post-processing methods, and statistical and artificial intelligence methods.

[0003] Numerical weather prediction and ensemble forecasting methods rely on solving atmospheric motion equations for prediction, forming the physical foundation of modern weather forecasting. However, due to uncertainties in the model's dynamic framework, initial values, and physical process parameterization, their ability to forecast precipitation with significant nonlinear characteristics is limited, and their computational costs are high. Numerical forecast post-processing methods (such as the MOS method) correct systematic biases by establishing a statistical relationship between numerical model outputs and observations, but they heavily depend on the length and quality of historical data. Furthermore, they require retraining after numerical model upgrades and lack adaptability to the dual characteristics of precipitation (discrete and continuous).

[0004] Statistical and artificial intelligence methods have become a research hotspot due to their ability to process big data and learn complex patterns, but existing methods still have many shortcomings:

[0005] 1. The linear regression model assumes a linear relationship between forecast factors and precipitation, which cannot handle complex nonlinear problems and relies on a large amount of historical data. The model fails after the model is upgraded.

[0006] 2. Nonlinear regression models require manual pre-setting of the nonlinear function form, which relies excessively on prior knowledge. Incorrect selection of the function form can lead to a significant decrease in model performance.

[0007] 3. The U-Net convolutional neural network method has a fast inference speed, but lacks physical consistency. It is good at short-term forecasts, but its performance degrades rapidly in 1-4 day forecasts, and it cannot solve the "field-point" forecast problem.

[0008] 4. The ConvLSTM spatiotemporal sequence prediction method is limited by the field of view of the convolution kernel, making it difficult to capture large-scale environmental field information. The forecast results are prone to ambiguity, and the error accumulates with the forecast duration, making it unable to cope with non-continuous daily precipitation forecasts.

[0009] 5. Although the improved spatiotemporal series prediction method has been optimized in specific scenarios, it still has problems such as insufficient long-term memory, large computational overhead, and error accumulation, making it difficult to meet the needs of 1-4 day discrete station precipitation forecast.

[0010] An ideal method for short-term quantitative precipitation forecasting at discrete stations should possess the capabilities to model complex nonlinearities and spatiotemporal dependencies, optimize field-point forecasts, perceive large-scale environmental fields, address discontinuous daily precipitation forecasting, and meet operational timeliness requirements. Therefore, a novel forecasting method is urgently needed to overcome the shortcomings of existing technologies. Summary of the Invention

[0011] The purpose of this invention is to provide a method for short-term precipitation forecasting at discrete sites based on convolutional and dimensionality-reduced deconvolutional neural networks.

[0012] To achieve the above objectives, the present invention provides the following technical solution:

[0013] Dataset Construction

[0014] The multivariate two-dimensional forecast factor X is subjected to min-max normalization to ensure its value ranges from [0,1]. The normalization formula is as follows: The dimension of X is ( C , ), Where C is the total number of samples, and C is the number of channels or variables. For the width of the input data, For the height of the input data, and These represent the minimum and maximum values ​​of the input data X after retaining the channel dimension in other dimensions. The multivariate two-dimensional forecast factors include temperature, humidity, U-wind, V-wind, geopotential height, equivalent potential temperature, wet hydrostatic energy, convective effective potential energy, vertical wind shear, and moist potential vortex on the 850hPa, 500hPa, and 200hPa isobaric surfaces, as well as sea level pressure, surface 2m temperature, surface 2m humidity, and surface 10m U-wind and 10m V-wind.

[0015] Minimum-maximum normalization is performed on precipitation data Y from N discrete stations. The normalization formula is as follows: The dimension of Y is ( (N) Let be the maximum value of precipitation data from N stations in the sample dimension.

[0016] At time t=0, normalized forecast factor data at time t=0 and precipitation data from N discrete stations for D days after time t=0 are extracted and arranged in chronological order; the above extraction operation is repeated over a sliding time t, resulting in data with dimensions (B, C, ...). , The forecast factor dataset and forecast object dataset are (B, N×D), where B is the number of samples entering the training and D is the number of days in advance of the forecast.

[0017] Fusion of multi-layer, multi-kernel convolution with attention mechanism

[0018] The normalized multivariate two-dimensional forecast factors were evaluated using kernel sizes k∈{3,5,7,9}. Perform J-layer (J≥4) neural network operations. After each kernel undergoes J layers of convolution, the number of feature maps is less than 0.25×N×D. Each kernel corresponds to an independent branch. Each layer operation includes:

[0019] 1. Two-dimensional convolution: convolution stride =1, fill The number of output channels for the j-th layer is set according to the kernel size to maintain space dimensions. The convolution formula is

[0020] ;

[0021] in , For convolution kernel weights, For bias.

[0022] 2. Batch Normalization: Calculate the mean for each channel. and variance The convolution results are normalized and then processed using learnable parameters. and Adjustment, the formula is as follows

[0023] ;

[0024] ε is a small constant.

[0025] 3. ReLU nonlinear activation: .

[0026] 4. Max Pooling: Uses a 2×2 pooling window with a stride of 2, and the formula is as follows: The output width and height are halved.

[0027] 5. Attention Mechanism: A spatial attention mechanism is added after the second pooling layer. Spatial weights are obtained through 1×1 convolution and the Sigmoid function, and element-wise multiplication is performed with the pooling result. A channel attention mechanism is added after the penultimate pooling layer. Channel weights are obtained through global average pooling, fully connected layers, and the Sigmoid function, and channel multiplication is performed with the pooling result.

[0028] 6. Dropout operation: Performed after the last pooling layer, randomly discarding some elements with a probability of (0-0.3), as shown in the formula. , It is a mask matrix.

[0029] One-dimensional flattened fusion

[0030] The output feature map corresponding to each kernel k Flatten the vector using the Flatten operation. The dimension is (B, , Then, the Concat operation is used to concatenate and merge the vectors corresponding to all kernels, resulting in a dimension of (B, The fusion vector V of F), where , .

[0031] One-dimensional deconvolution upsampling

[0032] The dimensions of the fusion vector V are adjusted to (B, F, ... A series of one-dimensional deconvolution operations are performed, and the F-dimensional dimension is upsampled. The deconvolution kernels are selected in descending order, including odd-numbered kernels and even-numbered kernels. The number of channels... The number of channels is gradually reduced to 1, and deconvolution is repeated once when the number of channels is 1. The formula for the deconvolution of the l-th layer is: ,in Step size, These are the weights of the deconvolution kernel. The deconvolution iteration M and stride are adjusted accordingly. and filling Make the output length satisfy Finally, the results are quasi-symmetrically sheared and aligned to N×D, and the channel dimension is removed to obtain an output with dimension (B, N×D). .

[0033] Model training optimization

[0034] Constructing the mean squared error loss function The Adam algorithm is used to minimize the loss function, and the parameter update formula is as follows: Where θ represents the model parameters and η represents the learning rate. The gradient. When the change in the loss function... Training stops when τ is the set threshold.

[0035] Precipitation Forecast

[0036] For new multivariate two-dimensional forecast factor data After processing using the normalization method in step 1, the input is fed into the trained model to obtain normalized predicted values ​​with dimension (1, N×D). Convert it to a (D, N) matrix and perform inverse normalization. The final precipitation forecast values ​​for N discrete stations over the next d days (d=1,2,…,D) are obtained.

[0037] The beneficial effects of this invention are as follows:

[0038] This invention eliminates the need to assume a combined relationship between forecast factors and precipitation. It automatically learns complex nonlinear correspondences from data through multi-layer, multi-kernel convolutional and deconvolutional networks, overcoming the limitations of linear regression and traditional nonlinear regression models. Employing a "convolution-dimensionality reduction deconvolution" architecture, it directly uses a multi-dimensional meteorological grid field as input and discrete station precipitation as output, dynamically extracting the most relevant features for each station. This effectively solves the "field-point" forecasting problem and avoids errors caused by interpolating station data to the grid.

[0039] By expanding the perceptual scope of the environmental field through multi-kernel convolutional branches, and combining spatial attention and channel attention mechanisms, meteorological features at different scales can be fully extracted, enhancing the perception capability of large-scale environmental fields and improving forecast accuracy. Constructing models using multivariate forecasting factors with clear physical meaning can, to some extent, solve the problem of non-continuous daily precipitation forecasting and extend the effective forecast duration.

[0040] The method employs dimensionality reduction deconvolution technology to directly model discrete sites, eliminating the need for interpolation and time-dimensional convolution. This significantly reduces the number of model parameters and computational requirements, ensuring that training and inference speeds meet the timeliness requirements of business forecasts, and facilitating portability and promotion.

[0041] The above description is only an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, the preferred embodiments of the present invention and the accompanying drawings are described in detail below. Attached Figure Description

[0042] Figure 1 This is a schematic diagram of a discrete-site precipitation short-term forecasting method based on convolutional and dimensionality-reduced deconvolutional neural networks, as shown in Embodiment 1 of the present invention. Detailed Implementation

[0043] The technical solution of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. In the description of the present invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the present invention. In addition, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0044] Example 1

[0045] See Figure 1 In this embodiment, the forecast advance days are D=4 days, the number of discrete stations is N=70, and the multivariate two-dimensional forecast factors include temperature, humidity, U wind, V wind, geopotential height, equivalent potential temperature, wet hydrostatic energy, convective effective potential energy, wind vertical shear and wet potential vortex on the 850hPa, 500hPa and 200hPa isobaric surfaces, as well as sea level pressure, surface 2m temperature, surface 2m humidity, surface 10m U wind, and surface 10m V wind, for a total of C=3×10 + 5=35 channels.

[0046] Dataset Construction

[0047] Multivariate two-dimensional forecast factor data from the past 60 years and daily precipitation observation data from 70 discrete stations were collected, with a total sample size of 21,900. 70% was used as the training set (15,330 samples), 20% as the validation set (4,380 samples), and 10% as the test set (2,190 samples). Min-max normalization was applied to the forecast factor data. and The minimum and maximum values ​​of each variable across all samples and spatial dimensions are given. The normalized data dimension is (21900, 35, 128, 128).

[0048] Normalize the precipitation data from the stations. =350mm, the maximum precipitation observed at all stations, and the normalized data dimension is (21900, 70).

[0049] Samples are extracted from the training set using a sliding time window, with t ranging from 0 to 15330 - 1. Forecast factor data at time t and station precipitation data for four days after time t are extracted and arranged chronologically to form the forecast object data. The final training sample size is B = 15330. The dimension of the training set forecast factor dataset is (15330, 35, 128, 128), and the dimension of the forecast object dataset is (15330, 70 × 4 = 280). Similarly, the dimension of the validation set forecast factor dataset is (4380, 35, 128, 128), and the dimension of the forecast object dataset is (4380, 70 × 4 = 280); the dimension of the test set forecast factor dataset is (2190, 35, 128, 128), and the dimension of the forecast object dataset is (2190, 70 × 4 = 280).

[0050] Fusion of multi-layer, multi-kernel convolution with attention mechanism

[0051] Four convolutional branches with kernel sizes k=3, 5, 7, 9 are used, and each branch performs J=4 layers of convolutional operations. The number of output channels in layers 1-4 are 16×2¹=32, 16×2²=64, 16×2³=128, 16×2, and 16×2, respectively. 4 =256.

[0052] The stride of each convolutional layer is 1, and the padding is set according to the kernel size: P=1 for k=3, P=2 for k=5, P=3 for k=7, and P=4 for k=9, ensuring that the spatial dimensions remain unchanged after convolution; batch normalization uses ε=10⁻. 5 The ReLU activation function is calculated according to the formula; max pooling uses a 2×2 window with a stride of 2. After 4 pooling layers, the spatial size of the output feature map of each branch is 128÷2. 4 =8, that is, (15330, 256, 8, 8).

[0053] Spatial attention mechanism: After the second pooling layer (spatial size is 128÷2²=32), the number of channels is reduced from 64 to 1 through 1×1 convolution, and the spatial weights are obtained by the Sigmoid function. The weights are then multiplied element-wise with the pooling result (15330, 64, 32, 32).

[0054] Channel attention mechanism: After the penultimate pooling layer (spatial size is 128÷2³=16), the feature map (15330, 128, 16, 16) is subjected to global average pooling to obtain (15330, 128). The channel weights are obtained by passing through two fully connected layers (128→64→128), ReLU activation, and Sigmoid function. The channel weights are then multiplied with the pooling result.

[0055] Dropout operation: Performed after the fourth pooling layer, with a dropout probability set to 0.1.

[0056] One-dimensional flattened fusion

[0057] The output feature maps (15330, 256, 8, 8) of the four branches are flattened, and the flattened vector of each branch has a length of 8×8=64. After concatenation by the Concat operation, the fused vector has a dimension of (15330, 256, 64×4=256), and the feature dimension and channel dimension are swapped.

[0058] One-dimensional deconvolution upsampling

[0059] The deconvolution kernels are selected in the order of 9, 7, 6, 5, 4, 3, 1, 1, with the initial input dimension being (15330, 256, 256). The number of channels is gradually halved starting from 256: 256→128→64→32→16→8→4→1→1.

[0060] Adjust the deconvolution stride and padding to make the final output length ≤ 280. ≤420, and cut both ends as needed, so that =280 (N×D=70×4=280), deconvolution times M=9, output dimension is (15330, 280, 1), after removing the channel dimension we get (15330, 280).

[0061] Model training optimization

[0062] Set the learning rate η = 0.001 and the loss function threshold τ = 10⁻ 5 The Adam algorithm was used to train the model, with 100 iterations. Training stopped when the change in the loss function was less than τ for 5 consecutive iterations.

[0063] Model testing

[0064] Input multivariate two-dimensional forecast factor data (2190, 35, 128, 128) from the test set, normalize it, and input it into the trained model to obtain normalized predicted values ​​(2190, 280). This is then converted into a matrix of (2190, 4, 70) and subjected to inverse normalization (multiplied by ). =350mm) to obtain the daily precipitation forecast values ​​for the next 4 days for 2190 samples and 70 discrete stations. The forecast object dataset (2190, 280) on the test machine is converted into a matrix of (2190, 4, 70) for comparison and evaluation, so as to understand the forecast performance of the model.

[0065] Precipitation Forecast

[0066] Input new multivariate two-dimensional forecast factor data (1, 35, 128, 128), normalize it, and input it into the trained model to obtain normalized predicted values ​​(1, 280). This is then converted into a (4, 70) matrix and subjected to inverse normalization (multiplied by 1 / 2). =350mm) to obtain the daily precipitation forecast values ​​for the next 4 days for 70 discrete stations.

[0067] Example 2

[0068] The difference between this embodiment and Embodiment 1 is that dilated convolution is used instead of multi-kernel convolution, with dilation rates of 1, 2, 3, and 4, respectively. The remaining steps are the same as in Embodiment 1. By expanding the receptive field without increasing the number of parameters, dilated convolution can also fully extract meteorological features at different scales, achieving accurate "field-point" precipitation forecasting.

[0069] Example 3

[0070] In this embodiment, the forecast advance period D = 3 days, the number of discrete stations N = 200, the deconvolution output length is set to N × D = 600, and the deconvolution result is adjusted to 600 ≤ ≤900 (1.5 times N×D), cut and aligned to 600, the remaining parameters and steps are the same as in Example 1. This example verifies the adaptability of the present invention to different forecast durations and numbers of stations.

[0071] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0072] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A method for short-term precipitation forecasting at discrete stations based on convolutional and dimensionality-reduced deconvolutional neural networks, characterized in that, Includes the following steps: S1. Normalize the multivariate two-dimensional forecast factor X and the precipitation data Y of N discrete stations for the next D days. Based on the correspondence between the forecast factor and the forecast object, establish the forecast factor dataset and the forecast object dataset. S2. Perform multi-layer, multi-kernel two-dimensional convolution, nonlinear activation, standardization and pooling operations on the normalized multivariate two-dimensional forecast factor X. Add spatial attention mechanism and channel attention mechanism to the second layer and penultimate layer respectively. Perform dropout operation in the last layer. S3. Perform one-dimensional flattening and fusion on the feature map output in step S2; S4. Perform one-dimensional deconvolution upsampling calculation on the flattened fusion result, adjust the deconvolution parameters to make the output dimension meet the requirements and perform alignment processing. S5. Construct the mean squared error loss function, minimize the loss function using an optimization algorithm, and update the model parameters; S6. Input new multivariate two-dimensional forecast factor data, use the trained model to make forecasts, and obtain the precipitation prediction values ​​for N discrete stations over the next D days.

2. The method according to claim 1, characterized in that, In step 1, the normalization of the multivariate two-dimensional forecast factor X adopts minimum-maximum normalization, and the normalization formula is as follows: ； Where the dimension of X is ( C , ), Where C is the number of samples, and C is the number of channels or variables. For the width of the input data, For the height of the input data, and These are the minimum and maximum values ​​of the input data X after retaining the channel dimension in other dimensions; the multivariate two-dimensional forecast factors include temperature, humidity, U-wind, V-wind, geopotential height, equivalent potential temperature, wet hydrostatic energy, convective effective potential energy, wind vertical shear and wet potential vortex on the 850hPa, 500hPa and 200hPa isobaric surfaces, as well as sea level pressure, surface 2m temperature, surface 2m humidity and surface 10m U-wind and 10m V-wind.

3. The method according to claim 1, characterized in that, The normalization formula for precipitation data Y from N discrete stations in step S1 is: ； Where the dimension of Y is N, Where N is the sample size, and N is the number of precipitation stations. Let be the maximum value of precipitation data from N stations in the sample dimension.

4. The method according to claim 1, characterized in that, The specific process of establishing the dataset in step S1 is as follows: Set time t=0, extract the normalized forecast factor data at time t=0, and simultaneously, based on the forecast lead time D, extract precipitation data from N discrete stations for D days after time t=0 and arrange them in chronological order; repeat the above extraction operation over a sliding time t to form datasets with dimensions B, ... , , , and B, N×D of the forecast factor dataset and forecast object dataset, where B is the number of samples entering the training.

5. The method according to claim 1, characterized in that, In step S2, the kernel size k∈{3,5,7,9} is used, each kernel corresponds to an independent branch, and J layers of convolution operations are performed, where J≥4, and the number of output channels in the j-th layer is j=1,2,…,J. Convolution stride =1, fill The kernel size is set to maintain the space size; batch normalization is used for standardization, ReLU function is used for nonlinear activation, and 2×2 max pooling is used with a step size of 2.

6. The method according to claim 1, characterized in that, The implementation process of the spatial attention mechanism in step S2 is as follows: After the second pooling layer, a 1×1 convolution kernel is used to output a single-channel feature map. The spatial weights are obtained through the Sigmoid function, and the pooling layer result is multiplied with the spatial weights element by element. The implementation process of the channel attention mechanism is as follows: After the penultimate pooling layer, global average pooling is performed first, and then the channel weights are obtained through two fully connected layers and the Sigmoid function. The pooling layer result is multiplied with the channel weights channel by channel.

7. The method according to claim 1, characterized in that, Step S3, one-dimensional flattening fusion, includes: flattening the feature map output by the J layer operation corresponding to each kernel into a vector, and then merging the flattened vectors corresponding to all kernels into a single one-dimensional vector through a merging operation.

8. The method according to claim 1, characterized in that, In step S4, the one-dimensional deconvolution uses convolution kernels in descending order of size, including odd-numbered and even-numbered kernels, with the number of channels gradually decreasing to 1. The deconvolution is repeated once when the number of channels reaches 1. The number of deconvolution iterations M and the stride are also specified. and filling The deconvolution output length is adjusted iteratively. satisfy Finally, the deconvolution result is quasi-symmetrically clipped and aligned with N×D. After removing the channel dimension, a result with dimension (B, N×D) is obtained.

9. The method according to claim 1, characterized in that, The formula for the mean squared error loss function in step 5 is: ； in, For the deconvolution output, The data is normalized station precipitation data; the Adam algorithm is used for optimization, when the change in the loss function... Stop optimization at this time. To set a threshold.

10. The method according to claim 1, characterized in that, The prediction process in step 6 is as follows: New input data is processed using the normalization method from step 1 and then input into the trained model to obtain normalized predicted values, which are then subjected to inverse normalization. The final precipitation forecast value was obtained, among which Let d be the normalized predicted value of the model output on day d, where d = 1, 2, ..., D.

Images