Method and device for predicting vertical accumulation liquid water content based on cross-scale operator learning
Through the cross-scale operator learning method, a multimodal input layer and spatial characterization aggregator are constructed, which solves the problems of multimodal information redundancy and model deviation in the prediction of vertical cumulative liquid moisture content, and achieves more accurate prediction results.
Patent Information
- Application Number
- CN202411032783.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-30
- Publication Date
- 2024-11-15
AI Technical Summary
The prior art has problems of multimodal information redundancy and model prediction results deviate from the real physical process in the prediction of vertical cumulative liquid moisture, resulting in complex models, inefficient and lack of effective information interaction.
Using a method based on cross-scale operator learning, the multi-modal input layer, cross-scale operator learner, spatial characterization aggregator, time module and decoder are used to extract and integrate physical background information to generate accurate conforming to atmospheric dynamics. Predict results.
The accuracy of vertical cumulative liquid moisture content prediction is improved, the model prediction is aligned with the actual physical process, effectively reducing redundant information, and improving information interaction and model efficiency.
Smart Images

Figure CN118962857A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of intelligent meteorological technology, and in particular relates to a method and device for predicting vertical cumulative liquid water content based on cross-scale operator learning. Background Art
[0002] Vertically accumulated liquid water content is one of the effective indicators for assessing severe convective weather, and its accurate forecast is crucial. If the vertically accumulated liquid water content of a region can be accurately predicted, meteorological changes can be predicted, people's production and life can be guided, meteorological resources can be used efficiently, and meteorological disasters can be reduced. It is of great significance in agricultural planning, transportation, tourism, and environmental protection. However, in a highly complex and interconnected atmospheric system, the evolution of vertically accumulated liquid water content is usually potentially related to other physical factors such as cloud cover, water vapor, and temperature. Methods that only rely on the statistical laws of vertically accumulated liquid water content itself and ignore these physical factors may lead to problems in which the prediction results deviate significantly from the actual physical laws.
[0003] To this end, it is necessary to make full use of the multimodal physical spatiotemporal data acquired by satellites to achieve reliable predictions aligned with atmospheric dynamics, but how to efficiently extract and integrate physical background information remains to be studied. In the prior art, some people use the method of pre-fusion mechanism to align multimodal data such as radar, spectral images, temperature and precipitation into the same format using interpolation and downsampling, and integrate them into a single tensor as the input of the channel dimension. There are also other methods that use mid-level fusion mechanisms to add multimodal meteorological data such as solar irradiance and wind intensity to the input, fuse multimodal information in the encoded feature space, and assist in time series prediction. There are also some more complex feature extraction and association strategies that introduce attention mechanisms to adjust the degree of influence of other modalities on the target modality, so as to achieve better information interaction and fusion.
[0004] However, this type of technology has limited improvements in model performance, and there are still problems such as model complexity, low efficiency, and lack of effective information interaction. On the one hand, while multimodal data brings effective information gain, it also brings useless redundant information, and more complex feature extraction and association strategies need to be designed to achieve effective multimodal information extraction and fusion. On the other hand, the existing vertical cumulative liquid water content prediction model lacks guidance based on mathematical stability mechanisms, such as nonlinear operators that can describe dynamic systems, and cannot obtain prediction results that are aligned with the actual complex atmospheric dynamic system. Summary of the invention
[0005] In view of the above-mentioned deficiencies in the prior art, the present invention provides a method and device for predicting vertical cumulative liquid water content based on cross-scale operator learning, which solves the problems of multimodal information redundancy and deviation of model prediction results from the real physical process, and improves the accuracy of vertical cumulative liquid water content prediction.
[0006] In order to achieve the above objectives, the technical solution adopted by the present invention is as follows: On the one hand, the present solution provides a vertical cumulative liquid water content prediction method based on cross-scale operator learning, comprising the following steps:
[0007] S1, randomly sampling at least one multimodal spatiotemporal series from the spatiotemporally aligned vertical cumulative liquid water content training dataset;
[0008] S2. According to the multimodal spatiotemporal sequence, a water content prediction model based on cross-scale operator learning is constructed, and the water content prediction model is trained;
[0009] S3. Use the trained water content prediction model to predict the future prediction results of the vertical cumulative liquid water content spatiotemporal series.
[0010] The beneficial effects of the present invention are as follows: the vertical cumulative liquid water content prediction method based on cross-scale operator learning provided by the present invention models the multimodal vertical cumulative liquid water content prediction task from the perspective of operator learning for the first time, extracts the spatial representation of atmospheric evolution, and effectively combines it with the spatial details of the vertical cumulative liquid water content. It effectively extracts and integrates physical background information, aligns the model prediction with the actual physical process, and effectively improves the accuracy of vertical cumulative liquid water content prediction compared with the baseline model during the reasoning process.
[0011] Furthermore, the step S2 comprises the following steps:
[0012] S201, construct a multimodal input layer, convert the multimodal spatiotemporal sequence into feature maps of different scales, and obtain discrete multi-scale features;
[0013] S202, construct a cross-scale operator learner and generate a spatial representation of atmospheric evolution based on discrete multi-scale features;
[0014] S203, construct a spatial representation aggregator to fuse the spatial representation of atmospheric evolution with the spatial representation vertical accumulated liquid water content VIL in multi-scale features;
[0015] S204, constructing a time module and a decoder, and obtaining the future prediction result of the currently input vertical cumulative liquid water content time-space sequence by using the time module and the decoder according to the fusion result, and completing the construction of the water content prediction model;
[0016] S205. Use the loss function to update the parameters of the moisture content prediction model until the moisture content prediction model converges, thereby completing the training of the moisture content prediction model. In the training process, AdamW is used as the optimizer, and preheating and cosine annealing are used as the learning rate scheduling strategies.
[0017] The beneficial effect of the above further scheme is that the water content prediction model can convert multimodal spatiotemporal data into discrete multi-scale features, and learn a nonlinear operator to map the discrete multi-scale features into a dynamic atmospheric spatial representation that can capture the evolution process, and then generate an enhanced spatiotemporal representation with spatial details of water content and atmospheric physical background. Through end-to-end constraints, the water content prediction model can make full use of multimodal information related to water content, extract a spatiotemporal representation that is more consistent with the real atmospheric physical background, and generate reliable and accurate prediction results that conform to the overall atmospheric dynamics.
[0018] Furthermore, the step S201 includes the following steps:
[0019] A1. Take the multimodal spatiotemporal sequence as input and construct a multimodal input layer;
[0020] A2. Convert the multimodal spatiotemporal sequence into feature maps of different scales using a parameter-sharing encoder to obtain discrete multi-scale features.
[0021] The beneficial effects of the above further scheme are: converting multimodal spatiotemporal sequences into feature maps of different scales, being able to map high-dimensional and complex multimodal data into discrete input space, generating discrete multi-scale features for further feature extraction and fusion. By sharing parameters, the number of parameters of the model can be significantly reduced, thereby reducing the complexity and computational cost of the model, while increasing the model's generalization ability for unknown data.
[0022] Furthermore, the step S202 includes the following steps:
[0023] B1. Construct a cross-scale operator learner;
[0024] B2. Based on the discrete multi-scale features, the cross-scale operator learner is used to learn the fitting of nonlinear operators to generate a spatial representation of the atmospheric evolution.
[0025] The beneficial effect of the above further scheme is that the cross-scale operator learner learns a nonlinear operator from discrete multi-scale features, which means that a universal representation of different discrete functions is learned, mapping the discrete input space to the hidden physical space, thereby obtaining an atmospheric space representation of the atmospheric evolution process.
[0026] Furthermore, the cross-scale operator learner includes:
[0027] A branch network with parallel multi-scale convolution kernels is used to use the kernel function to act on the discrete multi-scale features corresponding to each modality, and use spatial interpolation to generate spatially aligned multi-modal stacked features;
[0028] A backbone network based on the association generator includes a pooling layer, a first fully connected layer, a ReLU activation layer, a second fully connected layer and a Softmax layer connected in sequence. The backbone network is used to generate a modal association vector that adaptively changes with the input, and to achieve the fitting of a nonlinear operator by calculating the dot product of multi-modal stacking features and the adaptively changing modal association vector, thereby generating a spatial representation of the atmospheric evolution.
[0029] The beneficial effects of the above further scheme are: the branch network adopts parallel multi-scale convolution kernels, which can adapt to the heterogeneity of spatial scales and facilitate scale-invariant feature extraction. The backbone network adopts a correlation generator, which can adaptively change with the input to form an adaptive vector. By calculating the dot product of the multi-modal stacking feature and the adaptive modal correlation vector, the fitting of the nonlinear operator is realized to generate a spatial representation of the atmospheric evolution.
[0030] Furthermore, the spatial representation aggregator includes a convolutional layer, a layer normalization unit, and a Leaky ReLU activation function layer connected in sequence;
[0031] The convolution layer is used to use the local perception convolution kernel to act on the spatial representation of atmospheric evolution and the spatial representation vertical cumulative liquid water content VIL in the multi-scale feature to generate an aggregate feature;
[0032] The layer normalization unit is used to normalize each feature channel of the aggregated features;
[0033] The Leaky ReLU activation function layer is used to perceive the normalized aggregate features and generate a fusion result.
[0034] The beneficial effect of the above further scheme is that through these components, the spatial representation aggregator effectively combines detailed water content spatial data with broader atmospheric spatial information, and enhances the spatiotemporal representation from the perspective of atmospheric physical background. The enhanced spatiotemporal representation has more comprehensive spatial details, which is conducive to improving the accuracy of the spatiotemporal dynamics captured by the model and improving the final prediction accuracy of the model.
[0035] Furthermore, the time module is used to generate spatiotemporal sequence coding based on the fusion result using a Transformer converter model with residual connection;
[0036] The decoder is used to obtain the future prediction result of the current input vertical cumulative liquid water content spatiotemporal sequence according to the spatiotemporal sequence encoding by using multi-layer convolution.
[0037] The beneficial effect of the above further scheme is that the time module uses a Transformer converter with residual connections, which can capture trends and periodicity in the spatiotemporal representation, thereby effectively learning and predicting future changes. Finally, the decoder uses this rich spatial and temporal representation as input to generate the final time series prediction results, achieving reliable and accurate water content prediction.
[0038] Furthermore, the loss function is expressed as follows:
[0039]
[0040] N=K×H×W
[0041] x'[T+1:T+K]=F(x[1:T],y[1:T])
[0042] Among them, l mse represents the loss function, N represents the number of pixels in all future frames, K represents the number of future frames to be predicted, H represents the height of the image frame, W represents the width of the image frame, i represents the i-th pixel, T represents the number of historical frames, x'[T+1:T+K] represents the predicted frame, x[T+1:T+K] represents the real future frame, F() represents the water content prediction model, x[1:T] represents the historical sequence of vertically accumulated liquid water content, y[1:T] represents the historical sequence combination of 0.6μm visible light satellite channel, 6.9μm and 10.7μm infrared channels and overall lightning number data, x represents the real data of vertically accumulated liquid water content mode, T represents the number of historical frames, K represents the number of future frames to be predicted, and x' represents the prediction result of vertically accumulated liquid water content mode.
[0043] On the other hand, the present invention provides a vertical cumulative liquid water content prediction device based on cross-scale operator learning, including a processor and a memory, wherein the processor reads a computer program in the memory to execute the vertical cumulative liquid water content prediction method. BRIEF DESCRIPTION OF THE DRAWINGS
[0044] Figure 1 The figure is a flow chart of the method of the present invention.
[0045] Figure 2 It is a schematic diagram of the device structure of the present invention.
[0046] Among them, 21 is a processor, and 22 is a memory. DETAILED DESCRIPTION
[0047] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.
[0048] Example 1
[0049] The present invention first generates a deep model for vertical cumulative liquid water content prediction (referred to as a water content prediction model). The water content prediction model uses physical spatiotemporal data obtained from satellites to achieve effective information gain, and introduces a cross-scale operator learner and a spatial representation aggregator to effectively extract and integrate physical background information. In one embodiment of the present invention, the water content prediction model includes four parts: a multimodal input layer, a cross-scale operator learner, a spatial representation aggregator, a time module, and a decoder. Figure 1 As shown, the present invention provides a vertical cumulative liquid water content prediction method based on cross-scale operator learning, which is characterized by comprising the following steps:
[0050] S1, randomly sampling at least one multimodal spatiotemporal series from the spatiotemporally aligned vertical cumulative liquid water content training dataset;
[0051] S2. According to the multimodal spatiotemporal sequence, a water content prediction model based on cross-scale operator learning is constructed, and the water content prediction model is trained. The implementation method is as follows:
[0052] S201, construct a multimodal input layer, convert the multimodal spatiotemporal sequence into feature maps of different scales, and obtain discrete multi-scale features. The implementation method is as follows:
[0053] A1. Take the multimodal spatiotemporal sequence as input and construct a multimodal input layer;
[0054] A2. Convert the multimodal spatiotemporal sequence into feature maps of different scales using a parameter-sharing encoder to obtain discrete multi-scale features.
[0055] S202, construct a cross-scale operator learner, and generate a spatial representation of the atmospheric evolution based on the discrete multi-scale features, the implementation method of which is as follows:
[0056] B1. Construct a cross-scale operator learner;
[0057] B2. Based on the discrete multi-scale features, the cross-scale operator learner is used to learn the fitting of nonlinear operators to generate the spatial representation of atmospheric evolution;
[0058] S203, construct a spatial representation aggregator to fuse the spatial representation of atmospheric evolution with the spatial representation vertical accumulated liquid water content VIL in multi-scale features;
[0059] S204, constructing a time module and a decoder, and obtaining the future prediction result of the currently input vertical cumulative liquid water content time-space sequence by using the time module and the decoder according to the fusion result, and completing the construction of the water content prediction model;
[0060] S205, using the loss function to update the parameters of the water content prediction model until the water content prediction model converges, completing the training of the water content prediction model, wherein AdamW is used as the optimizer during the training process, and preheating and cosine annealing are used as the learning rate scheduling strategy;
[0061] In this embodiment, the cross-scale operator learner includes:
[0062] A branch network with parallel multi-scale convolution kernels is used to use the kernel function to act on the discrete multi-scale features corresponding to each modality, and use spatial interpolation to generate spatially aligned multi-modal stacked features;
[0063] A backbone network based on the association generator includes a pooling layer, a first fully connected layer, a ReLU activation layer, a second fully connected layer, and a Softmax layer connected in sequence, wherein the backbone network is used to generate a modal association vector that changes adaptively with the input, and to achieve fitting of a nonlinear operator by calculating a dot product of a multi-modal stacking feature and the adaptively changing modal association vector, thereby generating a spatial representation of the atmospheric evolution;
[0064] The spatial representation aggregator includes a convolutional layer, a layer normalization unit, and a Leaky ReLU activation function layer connected in sequence;
[0065] The convolution layer is used to use the local perception convolution kernel to act on the spatial representation of atmospheric evolution and the spatial representation vertical cumulative liquid water content VIL in the multi-scale feature to generate an aggregate feature;
[0066] The layer normalization unit is used to normalize each feature channel of the aggregated features;
[0067] The Leaky ReLU activation function layer is used to perceive the normalized aggregate features and generate a fusion result.
[0068] In this embodiment, the convolution layer is used to use the local perception convolution kernel to act on the atmospheric space representation and VIL space representation to achieve cross-channel information interaction and aggregation and generate fusion features. The layer normalization unit is used to normalize each feature channel of the aggregated feature so that the variance is close to 1 and the mean is close to 0, which helps to further generate a fusion feature with stable distribution and achieve a certain regularization effect. The Leaky ReLU activation function layer is used to perceive small changes in the fusion feature, increase the nonlinear ability of the network, ensure that the subtle atmospheric dynamics can be captured, and produce the final fusion result.
[0069] S3. Use the trained water content prediction model to predict the future prediction results of the vertical cumulative liquid water content spatiotemporal series.
[0070] In this embodiment, the present invention randomly samples at least one multimodal spatiotemporal sequence from a spatiotemporally aligned vertical cumulative liquid water content training data set, including vertical cumulative liquid water content (VIL), 0.6 μm visible light satellite channel (VIS), 6.9 μm and 10.7 μm infrared channels (IR069, IR107) and overall lightning quantity data (LGHT), establishes a multimodal input layer, and converts the multimodal spatiotemporal sequence into feature maps of different scales and modes using an encoder; establishes a cross-scale operator learner, inputs discrete multi-scale features, learns the fitting of nonlinear operators, realizes the mapping of feature space to hidden physical space, and generates a spatial representation of atmospheric evolution; establishes a spatial representation aggregator, fuses the extracted atmospheric spatial representation and the generated vertical cumulative liquid water content VIL modal features, and realizes the fusion of the vertical cumulative liquid water content VIL spatial details with the atmospheric physical background; establishes a time module and a decoder, inputs the fusion result, and obtains the future prediction result of the currently input vertical cumulative liquid water content spatiotemporal sequence.
[0071] In this embodiment, a multimodal spatiotemporal sequence {x[1:T], y[1:T]} is randomly sampled from the spatiotemporally aligned vertical cumulative liquid water content training data set, where x[1:T] represents the historical sequence of vertical cumulative liquid water content (VIL), and y[1:T] is a combination of the historical sequences of 0.6 μm visible light satellite channel (vis), 6.9 μm and 10.7 μm infrared channels (ir069, ir107) and overall lightning quantity data (lght). The above multimodal spatiotemporal sequence is used as input to establish a multimodal input layer, and the multimodal data such as vertical cumulative liquid water content VIL are converted into features of different scales through a parameter-sharing encoder to achieve mapping from input to feature space. In the multimodal input layer, the parameter-sharing encoders respectively contain 3 two-dimensional convolution blocks, and the connection relationship between them can refer to the existing Earth system spatiotemporal converter model Earthformer model, which will not be repeated here.
[0072] In this embodiment, a cross-scale operator learner is established, the obtained discrete multi-scale features are input, the fitting of nonlinear operators is learned, and the mapping from multimodal feature space to hidden physical space is realized.
[0073] In this embodiment, the cross-scale operator learner includes two parts: the branch network and the backbone network: Assume that the operator representation of the multimodal feature space u is G(u), and v is used as the label of different modes. Then the output of the operator can be regarded as G(u)(v). This means that the cross-scale operator learner can learn different representations of [u(VIL),u(vis),u(ir069),u(ir107),u(lght)]. The cross-scale operator is as follows:
[0074]
[0075] This formula means that for any ε>0, there exists a branch network function g and a backbone network function f that make the formula valid. In the cross-scale operator learner, the branch network uses parallel multi-scale convolution kernels, uses the kernel function to act on the features of each modality, and uses spatial interpolation to generate spatially aligned multi-modal stacked features. This structure adapts to the heterogeneity of spatial scales and facilitates scale-invariant feature extraction; the backbone network is based on the association generator, which sequentially connects the pooling layer, the first fully connected layer, the ReLU activation layer, the second fully connected layer, and the Softmax layer, and finally generates a modal association vector that can change adaptively with the input. Finally, the spatial representation of atmospheric evolution is generated by calculating the dot product of the multi-modal stacked features and the adaptive modal association vector.
[0076] In this embodiment, a spatial representation aggregator is established to fuse the extracted atmospheric spatial representation and the generated vertical cumulative liquid water content (VIL) modal features into a new spatiotemporal representation, thereby realizing the fusion of the vertical cumulative liquid water content (VIL) spatial details with the atmospheric physical background to enhance the overall model prediction capability.
[0077] In this embodiment, the atmospheric spatial representation and the vertical cumulative liquid water content (VIL) spatial representation enter the convolution layer at the same time and are connected to the layer normalization unit. The Leaky ReLU activation function is then used to sense small changes in the input to ensure that subtle atmospheric dynamics can be captured.
[0078] In this embodiment, a time module and a decoder are established, and the obtained fusion result is input to obtain the future prediction result of the vertical cumulative liquid water content spatiotemporal sequence currently input.
[0079] In one embodiment of the present invention, the constructed loss function l mse The calculation formula is as follows:
[0080]
[0081] x'[T+1:T+K]=F(x[1:T],y[1:T])
[0082] Among them, l mse represents the loss function, N represents the number of pixels in all future frames, K represents the number of future frames to be predicted, H represents the height of the image frame, W represents the width of the image frame, i represents the i-th pixel, T represents the number of historical frames, x'[T+1:T+K] represents the predicted frame, x[T+1:T+K] represents the real future frame, F() represents the water content prediction model, x[1:T] represents the historical sequence of vertically accumulated liquid water content, y[1:T] represents the historical sequence combination of 0.6μm visible light satellite channel, 6.9μm and 10.7μm infrared channels and overall lightning number data, x represents the real data of vertically accumulated liquid water content mode, T represents the number of historical frames, K represents the number of future frames to be predicted, and x' represents the prediction result of vertically accumulated liquid water content mode.
[0083] In this embodiment, the water content prediction model predicts the future K frames based on the historical T frames, that is, the prediction result is x'[T+1:T+K], and the real future frame x[T+1:T+K] is used as supervision, that is, x[T+1:T+K]∈R K×H×W×1 and x'[T+1:T+K]∈R K ×H×W×1 represent the real and predicted future frames respectively, and N=K×H×W represents the number of pixels in all future frames.
[0084] In this embodiment, the loss function l mse The role of is to ensure the authenticity and accuracy of the vertical cumulative liquid water content output by the vertical cumulative liquid water content prediction model, and to constrain the vertical cumulative liquid water content predicted by each vertical cumulative liquid water content prediction model to make it close to the true value.
[0085] In this embodiment, the obtained water content prediction model is used in the training process through the loss function l mse The network parameters are updated until the vertical cumulative liquid water content prediction model is finally fully converged. In one embodiment of the present invention, AdamW is used as an optimizer during the training process, and preheating and cosine annealing are used as learning rate scheduling strategies.
[0086] In this embodiment, in order to verify the actual effect of the vertical cumulative liquid water content prediction method provided by the embodiment of the present invention, it is put into use in a real vertical cumulative liquid water content prediction task scenario. Specifically, the inventor selected meteorological events sampled in the continental United States from January 2017 to December 2019. The data contains five modal frames of vertical cumulative liquid water content (VIL), 0.6μm visible light satellite channel (vis), 6.9μm and 10.7μm infrared channels (ir069, ir107) and total lightning number data (lght) at each moment. The spatial scales are 768×768, 192×192, 192×192, 384×384, and 48×48, respectively, and the spatial resolutions are 0.5km, 2km, 2km, 1km, and 8km, respectively, representing the same area. The time interval between different frames is 5 minutes, which records the movement of the atmospheric system. In order to make a fair comparison with other methods, the present invention uses the past 13 frames as input to predict the next 12 frames, and uses the commonly used critical success index (CSI) as the evaluation index. The CSI between the predicted results and the true value is shown in Table 1:
[0087] Table 1
[0088]
[0089]
[0090] It can be seen from Table 1 that the present invention can achieve relatively excellent performance in the real vertical cumulative liquid water content prediction task scenario and has good practical value.
[0091] Example 2
[0092] like Figure 2 As shown, the present invention provides a vertical cumulative liquid water content prediction device based on cross-scale operator learning, including a processor and a memory, wherein the processor reads a computer program in the memory to execute the vertical cumulative liquid water content prediction method described in any one of Example 1.
[0093] like Figure 2 As shown, the vertical cumulative liquid water content prediction device includes one or more processors 21 and a memory 22. The memory 22 is coupled to the processor 21 and is used to store one or more programs. When the one or more programs are executed by the one or more processors 21, the one or more processors 21 implement the vertical cumulative liquid water content prediction method based on cross-scale operator learning in the above-mentioned embodiment 1.
[0094] Among them, the processor 21 is used to control the overall operation of the vertical cumulative liquid water content prediction device based on cross-scale operator learning to complete all or part of the steps of the vertical cumulative liquid water content prediction method based on cross-scale operator learning. In an embodiment of the present invention, the processor 21 is preferably a GPU (graphics processing unit), but it can also be an FPGA (field programmable gate array), ASIC (application-specific integrated circuit), DSP (digital signal processor), etc. The memory 22 is used to store various types of data to support the operation of the vertical cumulative liquid water content prediction method based on cross-scale operator learning, and these data may include, for example, instructions for any application or method operating on the vertical cumulative liquid water content prediction device based on cross-scale operator learning, as well as application-related data.
[0095] The memory 22 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, etc.
[0096] In this embodiment, the vertical cumulative liquid water content prediction device based on cross-scale operator learning can be implemented by a computer chip or entity, or by a product with a certain function, for executing the above-mentioned vertical cumulative liquid water content prediction method based on cross-scale operator learning, and achieving the same technical effect as the above-mentioned method. A typical embodiment is a computer. Specifically, the computer can be, for example, a personal computer, a laptop computer, a vehicle-mounted human-computer interaction device, a public security checkpoint inspection device, a cellular phone, a camera phone, a smart phone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or a combination of any of these devices.
[0097] In this embodiment, the present invention further provides a computer-readable storage medium including program instructions, and when the program instructions are executed by a processor, the steps of the vertical cumulative liquid water content prediction method based on cross-scale operator learning in any of the above embodiments are implemented. For example, the computer-readable storage medium can be a memory including program instructions, and the above program instructions can be executed by a processor of a vertical cumulative liquid water content prediction device based on cross-scale operator learning to complete the above vertical cumulative liquid water content prediction method based on cross-scale operator learning, and achieve the same technical effect as the above method.
[0098] Compared with the prior art, the vertical cumulative liquid water content prediction method and device based on cross-scale operator learning provided by the present invention models the multimodal vertical cumulative liquid water content prediction task from the perspective of operator learning for the first time, extracts the spatial representation of atmospheric evolution, and effectively combines it with the spatial details of vertical cumulative liquid water content. It effectively extracts and integrates physical background information, aligns the water content prediction model prediction with the actual physical process, and effectively improves the accuracy of vertical cumulative liquid water content prediction compared with the baseline model during the reasoning process.
[0099] The vertical cumulative liquid water content prediction method and device based on cross-scale operator learning provided by the present invention are described in detail above. For those skilled in the art, any obvious changes made to it without departing from the essence of the present invention will constitute an infringement of the patent right of the present invention and will bear corresponding legal responsibilities.
Claims
1. A vertical cumulative liquid water content prediction method based on cross-scale operator learning, characterized in that: The following steps are involved: S1, randomly sampling at least one multimodal spatiotemporal series from the spatiotemporally aligned vertical cumulative liquid water content training dataset; S2. According to the multimodal spatiotemporal sequence, a water content prediction model based on cross-scale operator learning is constructed, and the water content prediction model is trained; S3. Use the trained water content prediction model to predict the future prediction results of the vertical cumulative liquid water content spatiotemporal series.
2. The vertical cumulative liquid water content prediction method based on cross-scale operator learning according to claim 1 is characterized in that: The step S2 comprises the following steps: S201, construct a multimodal input layer, convert the multimodal spatiotemporal sequence into feature maps of different scales, and obtain discrete multi-scale features; S202, construct a cross-scale operator learner and generate a spatial representation of atmospheric evolution based on discrete multi-scale features; S203, construct a spatial representation aggregator to fuse the spatial representation of atmospheric evolution with the spatial representation vertical accumulated liquid water content VIL in multi-scale features; S204, constructing a time module and a decoder, and obtaining the future prediction result of the currently input vertical cumulative liquid water content spatiotemporal sequence by using the time module and the decoder according to the fusion result, and completing the construction of the water content prediction model; S205. Use the loss function to update the parameters of the moisture content prediction model until the moisture content prediction model converges, and complete the training of the moisture content prediction model. During the training process, AdamW is used as the optimizer, and preheating and cosine annealing are used as the learning rate scheduling strategies.
3. The vertical cumulative liquid water content prediction method based on cross-scale operator learning according to claim 2 is characterized in that: The step S201 comprises the following steps: A1. Take the multimodal spatiotemporal sequence as input and construct a multimodal input layer; A2. Convert the multimodal spatiotemporal sequence into feature maps of different scales using a parameter-sharing encoder to obtain discrete multi-scale features.
4. The vertical cumulative liquid water content prediction method based on cross-scale operator learning according to claim 2 is characterized in that: The step S202 comprises the following steps: B1. Construct a cross-scale operator learner; B2. Based on the discrete multi-scale features, a cross-scale operator learner is used to learn the fitting of nonlinear operators to generate a spatial representation of atmospheric evolution.
5. The vertical cumulative liquid water content prediction method based on cross-scale operator learning according to claim 4 is characterized in that: The cross-scale operator learner comprises: A branch network with parallel multi-scale convolution kernels is used to use the kernel function to act on the discrete multi-scale features corresponding to each modality, and use spatial interpolation to generate spatially aligned multi-modal stacked features; A backbone network based on the association generator includes a pooling layer, a first fully connected layer, a ReLU activation layer, a second fully connected layer and a Softmax layer connected in sequence. The backbone network is used to generate a modal association vector that adaptively changes with the input, and to achieve the fitting of a nonlinear operator by calculating the dot product of multi-modal stacking features and the adaptively changing modal association vector, thereby generating a spatial representation of the atmospheric evolution.
6. The vertical cumulative liquid water content prediction method based on cross-scale operator learning according to claim 2 is characterized in that: The spatial representation aggregator includes a convolutional layer, a layer normalization unit, and a Leaky ReLU activation function layer connected in sequence; The convolution layer is used to use the local perception convolution kernel to act on the spatial representation of atmospheric evolution and the spatial representation vertical cumulative liquid water content VIL in the multi-scale feature to generate an aggregate feature; The layer normalization unit is used to normalize each feature channel of the aggregated features; The Leaky ReLU activation function layer is used to perceive the normalized aggregate features and generate a fusion result.
7. The vertical cumulative liquid water content prediction method based on cross-scale operator learning according to claim 2 is characterized in that: The time module is used to generate spatiotemporal sequence coding based on the fusion result by using a Transformer converter model with residual connection; The decoder is used to obtain the future prediction result of the current input vertical cumulative liquid water content spatiotemporal sequence according to the spatiotemporal sequence encoding by using multi-layer convolution.
8. The vertical cumulative liquid water content prediction method based on cross-scale operator learning according to claim 2 is characterized in that: The expression of the loss function is as follows: N=K×H×W x'[T+1:T+K]=F(x[1:T],y[1:T]) Among them, l mse represents the loss function, N represents the number of pixels in all future frames, K represents the number of future frames to be predicted, H represents the height of the image frame, W represents the width of the image frame, i represents the i-th pixel, T represents the number of historical frames, x'[T+1:T+K] represents the predicted frame, x[T+1:T+K] represents the real future frame, F() represents the water content prediction model, x[1:T] represents the historical sequence of vertically accumulated liquid water content, y[1:T] represents the historical sequence combination of 0.6μm visible light satellite channel, 6.9μm and 10.7μm infrared channels and overall lightning number data, x represents the real data of vertically accumulated liquid water content mode, T represents the number of historical frames, K represents the number of future frames to be predicted, and x' represents the prediction result of vertically accumulated liquid water content mode.
9. A vertical cumulative liquid water content prediction device based on cross-scale operator learning, comprising a processor and a memory, characterized in that: The processor reads the computer program in the memory to execute the vertical cumulative liquid water content prediction method described in any one of claims 1-8.
Citation Information
Cited By
Short-term surface solar irradiance prediction method and system
CN121459206A
A short-term surface solar irradiance forecasting method and system
CN121459206B
