Computing system for weather forecasting

Abstract

The present invention relates generally to a computing system based on a specific computational model, and in particular to a computing system for weather forecasting. A computing system for weather forecasting includes a computing resource, a feature extraction module, a first processing module, and a prediction module. The feature extraction module is configured to extract input features based on weather data inputs associated with a plurality of sources. The first processing module is configured to determine a first output reflecting short-term motion regularity based on the input features. The second processing module is configured to determine a second output reflecting long-term motion trend based on the input features. The prediction module is configured to determine a prediction value based on the first output and the second output. A loss function associated with the computing system can be determined based on the prediction value and a benchmark value. Parameters of the modules can be updated based on the loss function to train the modules. The computing system according to embodiments predicts more accurately and is better adaptable across domains.

Description

Computational systems for weather forecasting Technical Field

[0001] This invention generally relates to computational systems based on specific computational models, and more specifically to computational systems for weather forecasting. Background Technology

[0002] Weather forecasting plays a vital role in people's daily lives and production. In recent years, with the development of artificial intelligence technology, it has played an increasingly important role in weather forecasting.

[0003] Some existing solutions rely on single-source radar emissivity data for radar echo prediction to achieve weather forecasting. This paper proposes using existing deep learning techniques to design corresponding spatiotemporal modeling / prediction models based on N frames of past single-source radar emissivity data, thereby improving the model's predictive capabilities and enabling the prediction of N future frames of single-source radar reflectivity data.

[0004] One major category of these approaches is autoregressive prediction methods. Typical methods include ConvLSTM (see https: / / proceedings.neurips.cc / paper / 5955-convolutional-lstm-network-a-machine-learningapproach-for-precipitation-nowcasting.pdf) and PredRNN (see https: / / arxiv.org / abs / 2103.09504). ConvLSTM extracts spatial structure information from radar through convolutions, and then uses LSTM to establish the temporal relationship between different frames, thereby predicting future radar reflectivity data. Another major category is non-autoregressive prediction (i.e., multiple-input multiple-output) methods. Typical methods include SimVP (see https: / / arxiv.org / abs / 2206.05099). In the SimVP method, the network simultaneously inputs multiple frames of radar images and then simultaneously predicts multiple future frames. The network structure often uses a U-Net structure, and the feature representation function is enhanced through network upsampling, downsampling, and lateral connections.

[0005] Deep learning methods for radar short-term forecasts based on a single data source lack the ability to generalize across regions and observation modalities, making it difficult to deploy and infer under different or new climate, terrain, and observation conditions.

[0006] Some existing solutions use joint training with multiple datasets / multi-source data. Multi-dataset learning aims to leverage datasets from different domains / modalities for joint training to improve the model's generalization ability and robustness.

[0007] Such approaches include cue / normalization-driven domain robustness methods (e.g., PPT-like methods (see https: / / openaccess.thecvf.com / content / CVPR2024 / papers / Wu_Towards_Large-scale_3D_Representation_Learning_with_Multi-dataset_Point_Prompt_Training_CVPR_2024_paper.pdf)), methods that align features from different domains through adaptation layers or fusion strategies (e.g., UOD-like methods (see https: / / openaccess.thecvf.com / content_CVPR_2019 / papers / Wang_Towards_Universal_Object_Detection_by_Domain_Attention_CVPR_2019_paper.pdf)), and methods that leverage informative / projective losses to facilitate cross-domain feature sharing (e.g., MultiTrain-like strategies (see https: / / arxiv.org / abs / 2209.12362)).

[0008] Multi-dataset / multi-source data joint training methods have been implemented in fields such as point clouds and natural images. However, for the field of radar extrapolation, these methods have not yet taken into account the physical commonalities and regional characteristics existing in radar motion, and have failed to fully utilize these strategies to mitigate inter-domain conflicts and negative transfer.

[0009] Some existing approaches combine motion representation learning with physical priors. Spatiotemporal prediction typically considers joint modeling of long-term evolutionary trends and short-term motion changes, and the accuracy of the motion representation has a significant impact on prediction performance. Long-term evolutionary trends are captured by superimposing temporal features through neural networks, while short-term motion changes rely on motion estimation and compensation mechanisms. These approaches include PySteps (see https: / / gmd.copernicus.org / articles / 12 / 4185 / 2019 / ), MotionRNN (see https: / / openaccess.thecvf.com / content / CVPR2021 / papers / Wu_MotionRNN_A_Flexible_Model_for_Video_Prediction_With_Spacetime-Varying_Motions_CVPR_2021_paper.pdf), and NowcastNet (see https: / / www.nature.com / articles / s41586-023-06184-4). PySteps guides short-term motion estimation through physical constraints imposed by continuity equations. Methods such as MotionRNN and NowcastNet tend to introduce implicit flows, transformation operators, or warp operations into the feature domain, making them more flexible in learning representations.

[0010] The approach that combines motion representation learning and physical priors is computationally expensive in terms of optical flow and is difficult to couple to scenarios involving training with mixed multi-source data.

[0011] There is a need for improved technologies for weather forecasting in this field. Summary of the Invention

[0012] This invention is provided to provide improved techniques for weather forecasting.

[0013] This invention provides a computational system for weather forecasting, comprising: computational resources; a feature extraction module, which invokes the computational resources to: receive meteorological data input associated with multiple sources, the meteorological data input including an image on a two-dimensional plane, each pixel of the image including a pixel value associated with meteorological data; extract input features based on the meteorological data input using a deep neural network, the input features including a time dimension associated with the moment corresponding to the meteorological data input, a channel dimension associated with the meteorological data input, and two spatial dimensions associated with the image; and a first processing module, which invokes the computational resources to: determine motion features based on the difference results between adjacent input feature components distributed along the time dimension of the input features; and determine motion features based on the motion features and a pre-constructed motion prototype. A first output; a second processing module, which invokes the computing resources to: expand the receptive field of the input features to obtain input features with expanded receptive field; and perform domain adaptive modulation on the input features with expanded receptive field to determine a second output; and a prediction module, which invokes the computing resources to: generate a predicted value based on the first output and the second output, wherein a loss function associated with the computing system is determined based on the predicted value and a baseline value, and wherein parameters associated with the feature extraction module, parameters associated with the first processing module, parameters associated with the second processing module, and parameters associated with the prediction module are iteratively updated based on the loss function to train the feature extraction module, the first processing module, the second processing module, and the prediction module.

[0014] In the computing system described above, the first processing module invokes the computing resources to: determine a difference result between an input feature component associated with a first moment in the time dimension and an input feature component associated with a second moment in the time dimension, the second moment being adjacent to the first moment; and determine a motion feature component associated with the first moment based on the difference result, wherein each motion feature component associated with each moment in the time dimension constitutes the motion feature.

[0015] In the computing system described in any of the preceding claims, the first processing module invokes the computing resources to: perform a first convolution operation on the input feature components associated with the first time step to determine a dimensionality-reduced input feature component associated with the first time step; perform the first convolution operation on the input feature components associated with the second time step to determine a dimensionality-reduced input feature component associated with the second time step; perform a second convolution operation on the dimensionality-reduced input feature components associated with the second time step to determine a convolution operation result associated with the second time step; and determine the difference result based on the dimensionality-reduced input feature components associated with the first time step and the convolution operation result associated with the second time step.

[0016] In the computing system described in any of the above embodiments, the first processing module invokes the computing resources to: perform a third convolution operation on the difference result to determine the motion feature components associated with the first time moment.

[0017] In the computing system described in any of the above embodiments, the first processing module invokes the computing resources to: determine a weight vector based on the motion features; and determine the first output based on the weight vector and the motion prototype.

[0018] In a computing system as described in any of the preceding claims, the first processing module invokes the computing resources to: apply a projection function to the motion features to calculate a similarity score; and perform a normalization operation on the similarity score to determine the weight vector.

[0019] The computational system as described in any of the above, wherein the motion prototype is pre-built through random initialization and joint optimization learning trained across datasets.

[0020] In the computing system described in any of the above, the second processing module invokes the computing resources to: perform a fourth convolution operation and a fifth convolution operation on the input features to expand the spatial receptive field of the input features, wherein the fourth convolution operation applies a depth-separable convolution kernel, and the fifth convolution operation applies a dilated depth-separable convolution kernel.

[0021] In the computing system described in any of the above, the second processing module invokes the computing resources to: perform a sixth convolution operation on the result of the fifth convolution operation to expand the temporal receptive field of the input features, wherein the sixth convolution operation applies a channel aggregation convolution kernel.

[0022] In the computing system described in any of the above, the second processing module invokes the computing resources to: determine the expanded input features of the receptive field based on the input features and the result of the sixth convolution operation.

[0023] In the computing system described in any of the preceding claims, the second processing module invokes the computing resources to: perform global average pooling on the input features of the expanded receptive field to determine a global vector; compute domain activation vectors using multiple domain adapters based on the global vector; pass the global vector through a fully connected layer and perform normalization to determine a domain weight allocation probability vector; perform a weighted summation on the activation vectors associated with each domain adapter using the domain weight allocation probability vector to determine channel activations; and use the channel activations to modulate the expanded receptive field input features via channel-wise multiplication to determine the second output.

[0024] In the computing system described in any of the above, each domain adapter samples two fully connected layers, and the second processing module invokes the computing resources to: reduce the dimensionality of the global vector first, and then increase its dimensionality.

[0025] In a computing system as described in any of the above, the feature extraction module invokes the computing resources to sample the image of the meteorological data input by a predetermined multiple in the two spatial dimensions.

[0026] The computing system as described in any of the above claims further includes: an input module that invokes the computing resources to: receive raw meteorological data input from the plurality of sources; and sample the raw meteorological data input according to a predetermined sampling rule to obtain meteorological data input associated with the plurality of sources.

[0027] The computing system as described in any of the above embodiments, wherein the predetermined sampling rule includes equally spaced sampling.

[0028] The computing system as described in any of the preceding claims, wherein the feature extraction module invokes the computing resources to: receive a second meteorological data input from a single source, the second meteorological data input comprising a second image on a two-dimensional plane, each pixel of the second image comprising a pixel value associated with the real-time meteorological data; extract a second input feature associated with the second meteorological data input using a deep neural network based on the second meteorological data input, the second input feature comprising a time dimension associated with the detection time, a channel dimension associated with the second meteorological data input, and two spatial dimensions associated with the second image; the first processing module invokes the computing resources to: determine a second motion feature based on the difference between adjacent input feature components distributed along the time dimension of the second input feature; determine a third output based on the second motion feature and a pre-constructed motion prototype; the second processing module invokes the computing resources to: expand the receptive field of the second input feature to obtain a second input feature with expanded receptive field; and perform domain adaptive modulation on the second input feature with expanded receptive field to determine a fourth output; and the prediction module invokes the computing resources to: generate a second predicted value at a subsequent time based on the third output and the fourth output.

[0029] The computing system described in any of the above embodiments, wherein the meteorological data input includes at least one of the following: radar echo images, ocean current data, and wind field data.

[0030] The computing system according to embodiments of the present invention performs weather forecasts with higher accuracy and better cross-domain adaptability. Attached Figure Description

[0031] Various embodiments of the present invention are described in conjunction with the accompanying drawings.

[0032] Figure 1 is a block diagram illustrating a computational system for weather forecasting according to some embodiments of the present invention.

[0033] Figures 2-3 are schematic diagrams illustrating the training process of a computational system for weather forecasting according to some embodiments of the present invention.

[0034] Figure 4 is a schematic diagram illustrating the short-term motion feature extraction process of a computational system for weather forecasting according to some embodiments of the present invention.

[0035] Figure 5 is a schematic diagram illustrating a cross-domain prototype feature-driven motion enhancement process for a computational system for weather forecasting according to some embodiments of the present invention.

[0036] Figure 6 is a schematic diagram illustrating the long-term trend modeling process of a computational system for weather forecasting according to some embodiments of the present invention.

[0037] Figure 7 is a schematic diagram illustrating the domain adaptive modulation process of a computational system for weather forecasting according to some embodiments of the present invention.

[0038] Figures 8 and 9 are schematic diagrams illustrating the reasoning process of a computational system for weather forecasting according to some embodiments of the present invention. Detailed Implementation

[0039] In this application, the term "meteorological data input" refers to data that can be observed by meteorological observation equipment and can be used for weather forecasting. As an example, meteorological data input may include radar echo images, ocean current data, wind field data, etc. In this application, the "source" of meteorological data input refers to equipment capable of observing meteorological data.

[0040] According to one aspect of the present invention, a calculation system for weather forecasting is provided.

[0041] Figure 1 is a block diagram illustrating a computing system 100 for weather forecasting according to some embodiments of the present invention.

[0042] In some embodiments, the computing system 100 may include computing resources 110, a feature extraction module 120, a first processing module 130, a second processing module 140, and a prediction module 150. In some embodiments, the computing system 100 may also include an optional input module 160.

[0043] In some embodiments, the computing system 100 may be a desktop computer, a laptop computer, a server, a workstation, a portable electronic device, a mobile phone, etc.

[0044] In some embodiments, computing resources 110 may include a general-purpose processing unit (CPU), a graphics processing unit (GPU), various integer processing units, floating-point processing units, tensor processing units, ray tracing cores, etc.

[0045] In some embodiments, the feature extraction module 120, the first processing module 130, the second processing module 140, the prediction module 150, and the optional input module 160 can invoke computing resources 110 to implement corresponding functions. The execution details of each module of the computing system 100 during the training process will be described below with reference to Figures 2-3, and the execution details of each module of the computing system 100 during the inference process will be described below with reference to Figures 8-9. In some embodiments, the feature extraction module 120, the first processing module 130, the second processing module 140, the prediction module 150, and the optional input module 160 can be implemented as software modules.

[0046] Figures 2-3 are schematic diagrams illustrating the training process of a computational system for weather forecasting according to some embodiments of the present invention.

[0047] For clarity, the computing resources 110 of the computing system 100 in Figure 1 are not shown in Figures 2 and 3. However, it is understood that, in the description relating to Figures 2 and 3, each module invokes computing resources (e.g., computing resources 110 in Figure 1) to perform corresponding operations. In Figures 2 and 3, the labels “A” and “B” are used to reflect the relationship between Figures 2 and 3. That is, the processes that terminate at “A” and “B” in Figure 2 continue from “A” and “B” in Figure 3, respectively.

[0048] In some embodiments, the feature extraction module 120 may be configured to receive meteorological data input 210 associated with multiple sources. The meteorological data input 210 may include an image on a two-dimensional plane, wherein each pixel in the image may include a pixel value associated with meteorological data.

[0049] As an example, meteorological data input 210 may include radar echo images. Corresponding radar echo images can be obtained by a meteorological radar at multiple moments (time frames) at predetermined time intervals. The two-dimensional plane of each radar echo image may correspond to an observation plane parallel to a reference plane (e.g., sea level), and the pixel value of each pixel on the radar echo image may be associated with reflectivity.

[0050] In Figure 2, meteorological data input 210 is shown from multiple different sources (e.g., different weather radars) using different legends. As an example, three image frames are shown for each legend, which can be associated with different time frames.

[0051] In some embodiments, the feature extraction module 120 may be configured to extract input features 220 based on meteorological data input 210 using a deep neural network. As an example, the deep learning network may include various common network frameworks, such as ResNet, Transformer, etc.

[0052] Input feature 220 may include a time dimension T associated with the time corresponding to the meteorological data input 210. As an example, the time in time dimension T may reflect the time when the corresponding meteorological data was observed. As an example, for training data, the time in time dimension T may be stored along with the meteorological data; for inference data, the time in time dimension T may be provided along with the meteorological data. Input feature 220 may include a channel dimension C associated with the meteorological data input 210. Input feature 220 may include two spatial dimensions H and W associated with the image of the meteorological data input 210. As an example, the feature space of input feature 220 may be represented as [T, C, H, W].

[0053] In some embodiments, the first processing module 130 may be configured to determine motion features 250 (see FIG3) based on the difference between adjacent input feature components in the input feature components 230 distributed along the time dimension of the input features 220. The motion features 250 may be composed of motion feature components 240 distributed along the time dimension.

[0054] In Figure 2, the input feature 220 is illustrated in three dimensions, where the time dimension T and the channel dimension C are superimposed on the third dimension, shown as "T×C". Figure 2 also shows the input feature components 230 distributed along the time dimension T, where each input feature component may include the dimension [C,H,W].

[0055] The first processing module 130 processes the input feature components 230 distributed along the time dimension T to obtain motion feature components 240 distributed along the time dimension (see Figure 3). As an example, the first processing module 130 can process the input feature components 230 distributed along the time dimension T on a per-spatial-pixel basis (pixels on the [H,W] plane). The details of this processing will be further described below with reference to Figure 4.

[0056] Referring to Figure 3, in some embodiments, the first processing module 130 may be configured to determine a first output 260 based on motion features 250 and a pre-built motion prototype. The first output 260 may be an enhanced motion feature. As an example, the first output 260 may include dimensions [T, C, H, W]. The details of this processing will be further described below with reference to Figure 5.

[0057] Returning to Figure 2, on the other hand, the input features 220 extracted by the feature extraction module 120 can also be provided to the second processing module 140.

[0058] Referring to Figure 3, in some embodiments, the second processing module 140 may be configured to expand the receptive field of the input feature 220 to obtain the input feature 270 with an expanded receptive field. The details of this processing will be further described below with reference to Figure 6.

[0059] In some embodiments, the second processing module 140 may be configured to perform domain adaptive modulation on the input features 270 with expanded receptive fields to determine a second output 280. As an example, the second output 280 may include dimensions [T,C,H,W]. The details of this processing will be further described below with reference to FIG7.

[0060] Referring to Figure 3, in some embodiments, the prediction module 150 may be configured to generate a predicted value 290 based on a first output 260 and a second output 280. As an example, the first output 260 and the second output 280 may be added together, and the predicted value 290 may be generated based on the result of the addition.

[0061] In some embodiments, for the training process, a baseline value 292 (e.g., ground truth) may be provided along with meteorological data input 210. The predicted value 290 may be compared with the baseline value 292 to determine a loss function 294 associated with the computational system (e.g., computational system 100 of FIG. 1). Based on the determined loss function 294, parameters associated with feature extraction module 120, parameters associated with first processing module 130, parameters associated with second processing module 140, and parameters associated with prediction module 150 may be iteratively updated, thereby training feature extraction module 120, first processing module 130, second processing module 140, and prediction module 150. As an example, the above process may be iteratively executed a predetermined number of times, or it may be iteratively executed until the loss function 294 reaches a preset threshold.

[0062] In Figures 2-3, the arrows pointing to the right of the feature extraction module 120, the first processing module 130, the second processing module 140, and the prediction module 150 reflect that the parameters of the corresponding modules are updated based on the loss function 294. This process is performed iteratively to train each module.

[0063] The first processing module 130 processes the input features 220 to obtain the first output 260, which focuses on the short-term motion patterns of weather processes. The second processing module 140 processes the input features 220 to obtain the second output 280, which focuses on the long-term evolution trends of weather processes and has better cross-domain adaptability.

[0064] The computational system trained using meteorological data input 210 from multiple sources, as described above and in conjunction with Figures 2-3, can predict subsequent meteorological data based on real-time observed meteorological data input during the inference process, thereby achieving weather forecasting. The inference process is described below with reference to Figures 8-9.

[0065] According to some embodiments, the computing system employs joint training with multiple datasets from multiple sources, combining short-term motion enhancement, long-term evolution trend capture, and cross-domain generalization design during training to improve the prediction accuracy and cross-domain (e.g., for different climates, terrains, and observation conditions) generalization capabilities of the computing system (e.g., its constituent modules and associated models).

[0066] According to some embodiments, the computing system combines the physical commonalities and regional characteristics of the prediction module to perform extrapolation tasks, and jointly models the short-term motion and long-term evolution trend in the extrapolation task. In this modeling process, cross-domain motion prototypes are used, and the adaptability of each module of the computing system (e.g., related models) to different data domains is indicated through a domain adaptation mechanism, thereby mitigating inter-domain conflicts and negative migration problems.

[0067] Based on the results of differential calculations between adjacent input feature components in the computational system according to some embodiments, the computational overhead is reduced compared to existing kinematic representation methods (e.g., optical flow). Furthermore, the expressive power of short-term motion representation is enhanced through a cross-domain motion prototype enhancement mechanism. Simultaneously, by expanding the receptive field and employing domain adaptive modulation mechanisms, it can better adapt to multi-source data, effectively capture the long-term evolution trend of meteorological data, and improve the prediction accuracy and robustness of the computational system (e.g., correlation models).

[0068] In some embodiments, the feature extraction module 120 may be configured to sample the image of the meteorological data input 210 by a predetermined factor in two spatial dimensions.

[0069] In some embodiments, the computing device may further include an input module 160. The input module 160 may be configured to receive raw meteorological data input from multiple sources. The input module 160 may be configured to sample the raw meteorological input according to a predetermined sampling rule to obtain meteorological data input 210 associated with multiple sources.

[0070] In some embodiments, the predetermined sampling rule may include equal-interval sampling. As an example, the raw meteorological input may include raw meteorological input from datasets SEVIR, HKO7, SRAD2018, TAASRAD19, and Pudong. As an example, equal-interval sampling with an interval of [1:1:1:1:1] may be used, that is, each dataset is sampled equally, thereby maximizing the diversity and balance of the domain input.

[0071] Figure 4 is a schematic diagram illustrating the short-term motion feature extraction process of a computational system for weather forecasting according to some embodiments of the present invention. The process shown in Figure 4 can be a specific implementation described above in conjunction with Figures 2-3, in which the first processing module 130 generates motion feature components 240 distributed along the time dimension based on the input feature components 230 distributed along the time dimension, but the scope of the present invention is not limited thereto.

[0072] In some embodiments, the first processing module (e.g., the first processing module 130 in Figures 1 and 2) may be configured to determine the input feature component H associated with a first time point in the time dimension. t and the input feature component H associated with the second time step in the time dimension t+1 The difference results between the two time points. The second time point is adjacent to the first time point. The first processing module can be configured to determine the motion feature M associated with the first time point based on the difference results. t .

[0073] In some embodiments, the first processing module may be configured to process the input feature components H associated with the first time step. t Perform the first convolution operation 410 to determine the dimensionality-reduced input feature components F associated with the first time step. t The first processing module can be configured to process the input feature components H associated with the second time step. t+1 Perform the first convolution operation 410 to determine the dimensionality-reduced input feature components F associated with the second time step. t+1 As an example, the first convolution operation 410 may include a 1×1 two-dimensional convolution operation.

[0074] The first processing module can be configured to process the dimensionality-reduced input feature components F associated with the second time step. t+1 A second convolution operation 420 is performed to determine the result of the convolution operation associated with the second time step. As an example, the second convolution operation 420 may include a channel-wise convolution operation, such as a 3×3 channel-wise convolution operation.

[0075] The first processing module can be configured to be based on the dimensionality-reduced input feature components F associated with the first time step. t The difference result is determined by combining the result of the convolution operation associated with the second time step. This difference result can be obtained through difference operation 430.

[0076] In some embodiments, the first processing module may be configured to perform a third convolution operation 440 on the difference results to determine the motion feature M associated with the first time step. t As an example, the third convolution operation 440 may include a 1×1 two-dimensional convolution operation.

[0077] Some embodiments first reduce the dimensionality of the input feature components at a certain time, then perform channel-wise convolution operations on the input feature components at adjacent time points, and then perform difference operations with the dimensionality-reduced feature classification. Finally, the motion features associated with that time point are obtained by fusing them through 1×1 convolution operations, which reflects the short-term motion change pattern.

[0078] The above-described computational process, combined with the first time step t, can be described by the following formula (1):

[0079] (1)

[0080] In this context, conv() represents the convolution operation, and its subscript indicates the size of the convolution kernel used in the convolution operation.

[0081] Motion feature components distributed along the time dimension can constitute motion features (e.g., motion feature 250 in Figure 3). As an example, the motion features distributed along the time dimension can be connected sequentially along a third dimension (e.g., the channel dimension) perpendicular to the two-dimensional spatial dimension to obtain a motion vector.

[0082] Figure 5 is a schematic diagram illustrating a cross-domain prototype feature-driven motion enhancement process of a computational system for weather forecasting according to some embodiments of the present invention. The process shown in Figure 5 can be a specific implementation described above in conjunction with Figures 2-3, in which the first processing module 130 generates a first output 260 based on motion features 250, but the scope of the present invention is not limited thereto.

[0083] In some embodiments, the first processing module may be configured to determine a weight vector 520 based on motion features 250. The first processing module may be configured to determine a first output 260 (e.g., the first output 260 in FIG3) based on the weight vector 520 and a pre-built motion prototype 530.

[0084] In some embodiments, the first processing module may be configured to use a projection function 540 on motion features 250 (e.g., channel-wise, for example, for channel-wise motion feature components 510) to compute a similarity score. As an example, the similarity score can be approximated using a multilayer perceptron (MLP). The first processing module may be configured to perform a normalization operation 550 on the similarity score to determine a weight vector 520. As an example, the normalization operation can be performed using a softmax function.

[0085] As an example, the process of determining the weight vector 520 described above can be described by the following formula (2):

[0086] (2)

[0087] in, It is a weight vector. These are motion feature components, i and j, which are associated with two spatial dimensions. () is the projection function.

[0088] As an example, the process of determining the first output described above can be described by the following formula (3):

[0089] (3)

[0090] in, These are the feature components of the first output, i and j, which are associated with two spatial dimensions, and the subscript k is a pre-set hyperparameter.

[0091] In some embodiments, the motion prototype can be pre-built by random initialization and joint optimization learning trained across a dataset. As an example, random initialization can be determined based on a Gaussian distribution.

[0092] Figure 6 is a schematic diagram illustrating the long-term trend modeling process of a computational system for weather forecasting according to some embodiments of the present invention. The process shown in Figure 6 can be a specific implementation described above in conjunction with Figures 2-3, in which the second processing module 140 generates input features 270 with expanded receptive fields based on input features 220, but the scope of the present invention is not limited thereto.

[0093] In some embodiments, the second processing module may be configured to perform a fourth convolution operation 610 and a fifth convolution operation 620 on the input features 220 to expand the spatial receptive field of the input features 220. The fourth convolution operation 610 may include applying a depthwise separable convolution kernel, and the fifth convolution operation 620 may include applying a dilated depthwise separable convolution kernel. As an example, the depthwise separable convolution kernel may include a 5×5 kernel. As an example, the dilated depthwise separable convolution kernel may include a 7×7 kernel. The convolution kernels used in the fourth convolution operation 610 and the fifth convolution operation 620 are also shown in dark blocks in Figure 6. By performing the fourth convolution operation 610 and the fifth convolution operation 620 on the input features 220, the spatial receptive field can be expanded to capture long-distance spatial context.

[0094] In some embodiments, the second processing module may be configured to perform a sixth convolution operation 630 on the result of the fifth convolution operation 620 to expand the temporal receptive field of the input features 220. The sixth convolution operation 630 may include applying a channel aggregation convolution kernel. The convolution kernel used in the sixth convolution operation 630 is also shown in darker blocks in Figure 6. By further performing the sixth convolution operation 630 on the features with expanded spatial receptive field, temporal information of the input feature components at different times can be further integrated in the temporal dimension.

[0095] In some embodiments, the second processing module may be configured to determine the receptive field-expanded input feature 270 based on the input feature 220 and the result of the sixth convolution operation 630. As an example, the input feature 220 may be multiplied by the result of the sixth convolution operation 630 to determine the receptive field-expanded input feature 270.

[0096] Figure 7 is a schematic diagram illustrating a domain adaptive modulation process of a computational system for weather forecasting according to some embodiments of the present invention. The process shown in Figure 7 can be a specific implementation described above in conjunction with Figures 2-3, in which the second processing module 140 generates a second output 280 based on the input features 270 with an expanded receptive field, but the scope of the present invention is not limited thereto.

[0097] In some embodiments, the second processing module may be configured to perform a global average pooling operation 710 on the input features 270 with expanded receptive fields to determine a global vector 720. As an example, this process can be described by the following formula (4):

[0098] (4)

[0099] Where X is the input feature after receptive field expansion, and z is the global vector. The global average pooling operation is performed in two spatial dimensions.

[0100] In some embodiments, the second processing module may be configured to compute domain excitation vectors using multiple domain adapters based on global vector 720.

[0101] As an example, the global vector 720 can be fed in parallel into domain adapter 1 732, domain adapter 2 734, ..., domain adapter N 736. Domain adapter 1 732 can generate the first channel excitation vector 742 based on the global vector 720. Domain adapter 2 734 can generate the second channel excitation vector 744 based on the global vector 720. ... Domain adapter N 736 can generate the Nth channel excitation vector 746 based on the global vector 720.

[0102] As an example, each of the domain adapters in domain adapter 1 732, domain adapter 2 734, ..., domain adapter N 736 may include an MLP. As an example, each of the domain adapters in domain adapter 1 732, domain adapter 2 734, ..., domain adapter N 736 may employ two fully connected layers to first reduce the dimensionality of the channel dimension and then increase it to obtain the excitation vector of the corresponding channel.

[0103] As an example, the process can be described by the following formula (5):

[0104] (5)

[0105] in, It is the activation vector of the i-th channel. MLP() represents multilayer perceptron computation, and N is the number of domain adapters, which is a configurable parameter.

[0106] In some embodiments, the second processing module may be configured to pass the global vector 720 through the fully connected layer 752 and through the normalization operation 754 to determine the domain weight allocation probability vector 760.

[0107] As an example, the process can be described by the following formula (6):

[0108] (6)

[0109] in, It is a probability vector for domain weight assignment. FC() indicates passing through a fully connected layer.

[0110] In some embodiments, the second processing module may be configured to use the domain weight allocation probability vector 760 to perform a weighted summation on the excitation vectors (e.g., the first channel excitation vector 742, the second channel excitation vector 744, ..., the Nth channel excitation vector 746) associated with each of the domain adapters in each of the domain adapters 1 732, 2 734, ..., Nth channel excitation vector 746 to determine the channel excitation. As an example, the first channel excitation vector 742, the second channel excitation vector 744, ..., the Nth channel excitation vector 746 may be arranged in parallel at block 780 to obtain a parallel excitation vector 790, and then the parallel excitation vector 790 may be multiplied by the domain weight allocation probability vector 760 to achieve the above-mentioned weighted summation.

[0111] As an example, the process can be described by the following formula (7):

[0112] (7)

[0113] Where s is the channel excitation.

[0114] In some embodiments, the second processing module may be configured to use channel excitation 770 to multiply the receptive field extended input features 270 channel by channel to determine the second output 280.

[0115] As an example, the process can be described by the following formula (8):

[0116] (8)

[0117] in, It is the second output. This indicates a channel-by-channel multiplication operation.

[0118] Figures 8 and 9 are schematic diagrams illustrating the reasoning process of a computational system for weather forecasting according to some embodiments of the present invention.

[0119] For clarity, the computing resources 110 of the computing system 100 in Figure 1 are not shown in Figures 8-9. However, it should be clear that, in the descriptions relating to Figures 8-9, each module invokes computing resources (e.g., computing resources 110 in Figure 1) to perform corresponding operations. The modules of the computing system in Figures 8-9 can be trained through the process described above in conjunction with Figures 2-3. In Figures 8-9, the labels “C” and “D” reflect the relationship between Figures 8 and 9. That is, the processes that terminate at “C” and “D” in Figure 8 continue from “C” and “D” respectively in Figure 9.

[0120] The inference process shown in Figures 8-9 is similar to the training process shown in Figures 2-3, except that in the inference process, the input data is real-time meteorological data from a single source, and the inference...

[0121] In some embodiments, the feature extraction module 120 may be configured to receive a second meteorological data input 810 from real-time sensing from a single source. The second meteorological data input 810 may include a second image on a two-dimensional plane. Each pixel on the second image may include a pixel value associated with the real-time sensing meteorological data.

[0122] The feature extraction module 120 can be configured to extract second input features 820 associated with the second meteorological data input 810 using a deep neural network based on the second meteorological data input 810. The second input features 820 associated with the second meteorological data input 810 may include a time dimension associated with the time of detection, a channel dimension associated with the second meteorological data input 810, and two spatial dimensions associated with the second image.

[0123] The first processing module 130 can be configured to determine the second motion feature 850 based on the difference between adjacent second input feature components in the second input feature components 830 distributed along the time dimension of the second input feature 820.

[0124] The first processing module 130 can be configured to determine a third output 860 based on the second motion feature 850 and a pre-built motion prototype.

[0125] The second processing module 140 can be configured to expand the second input feature 820 to obtain the second input feature 870 with the receptive field expanded.

[0126] The second processing module 140 can be configured to perform domain adaptive modulation on the second input feature 870 with the receptive field extended to determine the fourth output 880.

[0127] The prediction module 150 can be configured to generate a second prediction value 890 at a subsequent time based on the third output 860 and the fourth output 880.

[0128] To verify the effectiveness of the computing system according to embodiments of the present invention, various prediction techniques were evaluated using existing datasets.

[0129] For each prediction technique, the corresponding models were trained using the datasets SEVIR, HKO7, SRAD2018, TAASRAD19, and Pudong. The predictive performance of the models trained under each technique was evaluated using the datasets SEVIR, HKO7, SRAD2018, TAASRAD19, Pudong, SCWDS, and QINGPU as test datasets.

[0130] In the evaluation, existing techniques used as baselines included ConvLSTM, MotionRNN, PredRNN, RainFormer, EarthFormer, and SimVP. ConvLSTM, MotionRNN, PredRNN, and SimVP methods are described in the relevant references mentioned in the "Background Art" section. For the RainFormer method, see C. Bai, F. Sun, J. Zhang, Y. Song and S. Chen, "Rainformer: Features Extraction Balanced Network for Radar-Based Precipitation Nowcasting," in IEEE Geoscience and Remote Sensing Letters, vol. 19, pp. 1-5, 2022, Art no. 4023305, doi: 10.1109 / LGRS.2022.3162882. For the EarthFormer method, see Gao, Z., et al. (2022). Earthformer: Exploring Space-Time Transformers for Earth System Forecasting. NeurIPS 2022.

[0131] In the evaluation, the Heidke Skill Score (HSS), Critical Success Index (CSI), and Fractional Skill Score (FSS), commonly used in the field of weather forecasting, were selected as evaluation indicators. The corresponding indicators were evaluated under preset threshold conditions, and the average value of each indicator's evaluation was taken.

[0132] Tables 1-5 below show the evaluation results on the SEVIR, HKO7, Pudong, SRAD2018, and TAASRAD19 datasets, respectively. Since these test datasets are from the same datasets used to train the corresponding models, the evaluation results reflect that the embodiments according to the present invention outperform the prior art in the field where the test datasets use data from the training datasets.

[0133]

[0134]

[0135]

[0136]

[0137]

[0138] Tables 6 and 7 below show the evaluation results on the SCWDS and QINGPU datasets, respectively. Since these test datasets are not from the same datasets used to train the corresponding models, the evaluation results reflect that embodiments of the present invention also outperform the prior art in unseen domains (i.e., the test datasets do not use data from the training datasets), thus further demonstrating the cross-domain generalization capability of embodiments of the present invention.

[0139]

[0140]

[0141] Embodiments of the present invention have been described with reference to the accompanying drawings, which are illustrative and not restrictive.

Claims

1. A computational system for weather forecasting, characterized in that, include: Computing resources; The feature extraction module calls upon the computing resources to: receive meteorological data input associated with multiple sources, the meteorological data input including an image on a two-dimensional plane, each pixel in the image including a pixel value associated with meteorological data; Based on the meteorological data input, input features are extracted using a deep neural network. These input features include a time dimension associated with the moment corresponding to the meteorological data input, a channel dimension associated with the meteorological data input, and two spatial dimensions associated with the image. A first processing module calls the computing resources to: determine motion features based on the difference results between adjacent input feature components distributed along the time dimension; and determine a first output based on the motion features and a pre-built motion prototype, wherein the motion prototype is pre-built through random initialization and joint optimization learning trained across datasets. A second processing module calls the computing resources to: expand the receptive field of the input features. The system obtains input features with an expanded receptive field; performs domain adaptive modulation on the expanded receptive field input features to determine a second output; and a prediction module invokes the computing resources to: generate a predicted value based on the first output and the second output, wherein a loss function associated with the computing system is determined based on the predicted value and a baseline value, and wherein parameters associated with the feature extraction module, the first processing module, the second processing module, and the prediction module are iteratively updated based on the loss function to train the feature extraction module, the first processing module, the second processing module, and the prediction module.

2. The computing system as described in claim 1, characterized in that, The first processing module invokes the computing resources to: determine the difference between the input feature component associated with a first moment in the time dimension and the input feature component associated with a second moment in the time dimension, wherein the second moment is adjacent to the first moment; and determine the motion feature component associated with the first moment based on the difference result, wherein each motion feature component associated with each moment in the time dimension constitutes the motion feature.

3. The computing system as described in claim 2, characterized in that, The first processing module invokes the computing resources to: perform a first convolution operation on the input feature components associated with the first time step to determine the dimensionality-reduced input feature components associated with the first time step; perform the first convolution operation on the input feature components associated with the second time step to determine the dimensionality-reduced input feature components associated with the second time step; perform a second convolution operation on the dimensionality-reduced input feature components associated with the second time step to determine the convolution operation result associated with the second time step; and determine the difference result based on the dimensionality-reduced input feature components associated with the first time step and the convolution operation result associated with the second time step.

4. The computing system as described in claim 2, characterized in that, The first processing module invokes the computing resources to perform a third convolution operation on the difference result to determine the motion feature components associated with the first time step.

5. The computing system as described in claim 1, characterized in that, The first processing module invokes the computing resources to: determine a weight vector based on the motion features; and determine the first output based on the weight vector and the motion prototype.

6. The computing system as described in claim 5, characterized in that, The first processing module invokes the computing resources to: apply a projection function to the motion features to calculate a similarity score; and perform a normalization operation on the similarity score to determine the weight vector.

7. The computing system as described in claim 1, characterized in that, The second processing module invokes the computing resources to perform a fourth convolution operation and a fifth convolution operation on the input features to expand the spatial receptive field of the input features, wherein the fourth convolution operation applies a depth-separable convolution kernel and the fifth convolution operation applies a dilated depth-separable convolution kernel.

8. The computing system as described in claim 7, characterized in that, The second processing module invokes the computing resources to perform a sixth convolution operation on the result of the fifth convolution operation to expand the temporal receptive field of the input features, wherein the sixth convolution operation applies a channel aggregation convolution kernel.

9. The computing system as described in claim 8, characterized in that, The second processing module invokes the computing resources to determine the expanded input features of the receptive field based on the input features and the result of the sixth convolution operation.

10. The computing system as claimed in claim 1, characterized in that, The second processing module invokes the computing resources to: perform global average pooling on the input features of the expanded receptive field to determine a global vector; calculate domain activation vectors using multiple domain adapters based on the global vector; pass the global vector through a fully connected layer and perform normalization to determine a domain weight allocation probability vector; and perform a weighted summation on the activation vectors associated with each domain adapter using the domain weight allocation probability vector to determine channel activations. And using the channel excitation to modulate the expanded input features of the receptive field via channel-by-channel multiplication to determine the second output.

11. The computing system as described in claim 10, characterized in that, Each domain adapter samples two fully connected layers, and the second processing module calls the computing resources to: reduce the dimensionality of the global vector first, and then increase its dimensionality.

12. The computing system as described in claim 1, characterized in that, The feature extraction module calls the computing resources to sample the image of the meteorological data input at a predetermined multiple in the two spatial dimensions.

13. The computing system as described in claim 1, characterized in that, Also includes: The input module invokes the computing resources to: receive raw meteorological data input from the multiple sources; And the raw meteorological data input is sampled according to a predetermined sampling rule to obtain meteorological data input associated with the multiple sources.

14. The computing system as described in claim 13, characterized in that, The predetermined sampling rules include sampling at equal intervals.

15. The computing system as described in claim 1, characterized in that, The feature extraction module invokes the computing resources to: receive a second meteorological data input from a single source in real time, the second meteorological data input including a second image on a two-dimensional plane, each pixel in the second image including a pixel value associated with the real-time meteorological data; Based on the second meteorological data input, a deep neural network is used to extract a second input feature associated with the second meteorological data input. The second input feature includes a time dimension associated with the detection time, a channel dimension associated with the second meteorological data input, and two spatial dimensions associated with the second image. The first processing module calls the computing resources to: determine a second motion feature based on the difference results between adjacent input feature components distributed along the time dimension of the second input feature; determine a third output based on the second motion feature and a pre-built motion prototype, which is pre-built by random initialization and joint optimization learning across datasets; the second processing module calls the computing resources to: expand the receptive field of the second input feature to obtain a second input feature with an expanded receptive field; and perform domain adaptive modulation on the second input feature with an expanded receptive field to determine a fourth output; and the prediction module calls the computing resources to: generate a second predicted value at a subsequent time based on the third output and the fourth output.

16. The computing system according to any one of claims 1-15, characterized in that, The meteorological data input includes at least one of the following: radar echo images, ocean current data, and wind field data.

Images