A Weather Forecasting Method Based on Physical Attention Spatiotemporal Prediction Network

Abstract

This invention belongs to the field of artificial intelligence, specifically relating to a weather forecasting method based on a physical attention-based spatiotemporal prediction network. The method includes: acquiring historical data to define a spatiotemporal prediction problem; designing a physical component extraction unit (PhyCell) and a physical decoupling network framework (PhyDNet) to decouple the spatiotemporal prediction problem into physical components and residual components, and capturing the physical dynamic features of the physical components; introducing an attention gate structure to remove irrelevant features from the residual components to obtain detailed features, and fusing these detailed features using additive attention to obtain residual dynamic features; fusing the physical dynamic features and residual dynamic features by summing them, and then reconstructing the prediction result using a decoder. This invention uses the deep architecture PastNet to focus on the dynamic information of physical constraints, thereby assisting weather forecasting spatiotemporal prediction tasks involving physical laws to produce more accurate prediction results.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence, specifically relating to a weather forecasting method based on a physical attention spatiotemporal prediction network. Background Technology

[0002] Weather forecasting has always been a focus of scientists because it contributes to improving human living standards in various aspects, including agriculture, transportation, social and economic decision-making, and disaster prevention. Since its inception at the end of the last century, numerical weather prediction (NMR) systems have dominated this task. They project the atmospheric system onto a discrete numerical grid, set initial conditions and boundary conditions based on meteorological principles, and solve nonlinear equations of fluid dynamics and thermodynamics describing weather evolution on supercomputers. This allows for the estimation of atmospheric motion and weather phenomena over a future period, thus predicting weather conditions. Despite continuous improvements in various ways, NMR systems inevitably consume significant computational resources and runtime to achieve steady-state simulations, rendering them ineffective in certain scenarios. More importantly, climate change has become an undeniable reality in recent years, leading to an increasing frequency of extreme weather events and causing loss of life and property. Therefore, there is an urgent need to improve weather forecasting performance to effectively capture potential model changes, thereby providing accurate meteorological data for crisis management and resource planning.

[0003] Meanwhile, artificial intelligence methods, represented by deep neural networks (DNNs), have demonstrated their ability to learn from nonlinear systems through a series of successful applications, including natural language understanding, visual object recognition, and drug discovery. Inspired by these successes, researchers are using DNNs to improve weather forecasting in various ways. Rasp et al. used DNNs as a post-processing method for numerical weather prediction systems to correct system biases. Inspired by image super-resolution methods, Adewoyin and Rod Rigues proposed a deep learning architecture to downscale numerical weather prediction products to obtain higher horizontal resolution forecasts. Liu et al. explored the discovery of severe weather and climate events using artificial intelligence methods. These works demonstrate that machine learning algorithms can be used in conjunction with numerical weather prediction models to improve weather forecast quality. However, their reliance on numerical weather prediction system products leads to a lack of timeliness, and their application scenarios are limited similarly to those of numerical weather prediction itself.

[0004] On the other hand, with the popularization of smart mobile devices, autonomous vehicles, and video surveillance systems, spatiotemporal data, represented by video, has experienced explosive growth. Spatiotemporal data processing has become a research hotspot in recent years, and deep learning methods have achieved remarkable success in this field. Because meteorological data describing atmospheric motion also possesses spatiotemporal correlation, spatiotemporal prediction networks derived from these works have gradually been used as a purely data-driven method for predicting weather conditions in advance. Shi, Wang, and others treated meteorological radar echoes or two-dimensional weather features as image sequences and processed them using deep learning methods, namely, using convolutional layers for spatial feature extraction and then inputting them into a recurrent neural network to learn the evolutionary trends in the temporal dimension. However, unlike video, the evolution of meteorological data is strictly constrained by physical laws, which are ignored by typical CNNs. To overcome this shortcoming, some scholars have introduced variants of recurrent neural networks to represent partial differential equations (PDEs), thereby better utilizing physical knowledge to optimize the learning process and achieving the expected excellent performance. Guen et al. used a deep encoder to map two-dimensional data sequences to a latent space and then decomposed the features into two uncorrelated components: physical features and residual features. Physical features refer to dynamic changes that conform to physical laws and can be described by partial differential equations; while residual features contain static detailed features, which can be learned through a CNN structure. This method outperformed comparative methods on datasets such as dynamic handwritten digits, human motion, and ocean surface temperature, demonstrating the effectiveness and application potential of the physical feature extraction operator. However, this "separate then combine" architecture cannot maximize the role of physical dynamic information. In fact, in meteorological spatiotemporal data, the detailed features of objects are not static but change due to their motion. Taking radar echo image sequences as an example, the shape and outline of clouds change as their position moves. In this setup, the identification of dynamic regions in the residual part will be suppressed by information from static regions that occupy a larger proportion of the target space; these regions often remain stationary between time steps. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention proposes a weather forecasting method based on a physical attention spatiotemporal prediction network, comprising the following steps:

[0006] S1: Obtain historical data from a dynamic system covering a fixed area at a fixed frequency, and define a spatiotemporal prediction problem based on the obtained historical data;

[0007] S2: Based on the spatiotemporal prediction problem, we design the physical component extraction unit PhyCell and the physical decoupling network framework PhyDNet. The spatiotemporal prediction problem is decoupled into physical components and residual components through the physical component extraction unit PhyCell.

[0008] S3: An attention gate structure is introduced to identify interesting and irrelevant feature parts within the residual components. Irrelevant feature parts are removed to obtain detailed features. Additive attention is used to fuse the detailed features to obtain the residual dynamic features.

[0009] S4: By summing and fusing the physical dynamic features and residual dynamic features, and then restoring them to the value space R of the observed data through a decoder, x is obtained. t+1 , complete the prediction.

[0010] The beneficial effects of this invention are:

[0011] This invention proposes and evaluates a deep architecture called PastNet, which focuses on the dynamic information of physical constraints to assist spatiotemporal prediction tasks involving physical laws. Experiments on the artificially synthesized Moving-MNIST dataset and real-world temperature datasets show that PastNet can produce more accurate prediction results. This indicates that PastNet outperforms baseline models in spatiotemporal data prediction tasks such as weather forecasting. In experiments conducted on both artificially synthesized and real datasets, its performance is superior to other control models, enabling it to generate weather forecast data more accurately and optimize existing weather forecast results. Attached Figure Description

[0012] Figure 1 This is an overall flowchart of the present invention;

[0013] Figure 2 This is a schematic diagram of the attention gating unit structure of the present invention. Detailed Implementation

[0014] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0015] A weather forecasting method based on a physical attention spatiotemporal prediction network, such as Figure 1 As shown, it includes:

[0016] S1: Obtain historical data from a dynamic system covering a fixed area at a fixed frequency, and define a spatiotemporal prediction problem based on the obtained historical data;

[0017] S2: Based on the spatiotemporal prediction problem, we design the physical component extraction unit PhyCell and the physical decoupling network framework PhyDNet. The spatiotemporal prediction problem is decoupled into physical components and residual components through the physical component extraction unit PhyCell.

[0018] S3: An attention gate structure is introduced to identify interesting and irrelevant feature parts within the residual components. Irrelevant feature parts are removed to obtain detailed features. Additive attention is used to fuse the detailed features to obtain the residual dynamic features.

[0019] S4: By summing and fusing the physical dynamic features and residual dynamic features, and then restoring them to the value space R of the observed data through a decoder, x is obtained. t+1 , complete the prediction.

[0020] Spatiotemporal prediction problems, including data-driven weather forecasting, utilize historical data sequences as input to generate forecasts for a future period. Historical data is obtained from observations at a fixed frequency from a dynamic system covering a fixed area. The covered area is discretized as an M×N grid, where M is the number of rows and N is the number of columns. In each observation, the data at each grid point contains C distinct physical quantities. Therefore, a single observation can be represented as a three-dimensional tensor: X∈R C ×M×N , where R represents the space of values ​​for the observations. The input data sequence driving the prediction is denoted as X. t-T+1 ,…,X t-1 ,X t The sequence to be predicted is denoted as Where T and K represent the lengths of the input and output sequences, respectively, and t represents the current time point. The spatiotemporal prediction problem can be defined in the following general form:

[0021]

[0022] Where h represents the defined spatiotemporal prediction problem, Let X represent the sequence to be predicted. t-T+1 ,…,X t-1 ,X t X represents the input data sequence driving the prediction. t+1 ,X t+1 ,…,X t+K This represents the values ​​of the input data sequence, X∈R. C×M×NThis represents the historical data observed within an M×N coverage area, R represents the value space of the observation, t represents the current time point, K represents the length of the output sequence, and T represents the length of the input sequence.

[0023] In particular, the present invention uses surface air temperature data to conduct experiments, and the observation data at each time point is a two-dimensional matrix.

[0024] Physical Units and Physics Decoupling Networks: A physical unit decoupling network model is constructed by combining the physical component extraction unit PhyCell and the physics decoupling network framework PhyDNet. This model assumes that for any dynamic system, there exists an implicit concept space. This allows the system state to be decoupled into a dynamic factor that can be described by physical laws and a residual factor that cannot be described by physical laws; formally, let the encoder decoder There exist h = E(X) and X = D(h) such that:

[0025] h = h p +h r

[0026] Among them, h p and h r Representing the physical component and residual component respectively, the physical dynamic factor h p The evolution of h is constrained by partial differential equations, while h r Otherwise; for the physical components, a network structure capable of representing partial differential equations is used for description, thereby achieving state prediction that conforms to physical laws; for the residual components, convolutional layers and recurrent neural network structures can be used for detailed feature extraction, and a Convolutional Long Short-Term Memory (ConvLSTM) structure is selected in the hidden state space. Perform detailed predictions; as shown in the following formula:

[0027]

[0028]

[0029] in, This indicates the dynamic characteristics of the captured physical components. The physical composition dynamics at the current time, x t This indicates the current system observation state, E() represents the encoder, and f p () represents a physical component learning unit, f r () represents a learning unit for residual components.

[0030] The physical component learning unit and the residual component learning unit are implemented using the PhyCell module and ConvLSTM module, respectively. First, the current system observation state x is... t Mapped to the latent space by encoder E Then the data flows into the physical branch and the residual branch respectively; in the physical branch, f p The physical state of the system in the next time step is inferred based on the mapping of the physical state and observations in the latent space at the current time step; in the residual branch, f r Then, based on the input at each step, the detailed information for the next step is inferred.

[0031] Attention gates have been shown to be able to identify interesting and irrelevant feature regions globally, helping to identify salient information carried in sparse feature maps. For example... Figure 2 As shown, physical dynamic features are used as the gating vector for each pixel, and an attention mechanism is used to determine the focal region in the corresponding residual map. The physical gating vector already contains dynamic information about the entire image, which can reveal in detail the pixel locations where evolution occurs. After passing through the attention gate, the residual parts with high physical attention coefficients are amplified, while irrelevant regions are suppressed. Additive attention is used to achieve higher prediction accuracy than multiplicative attention, and its implementation method is shown in the following equation:

[0032]

[0033]

[0034]

[0035] α=σ2(q att (r,g;Θ att ))

[0036]

[0037] in, Let α represent the residual dynamic features, σ²() represent the additive attention, σ²() represent the sigmoid activation function, r and g represent the detail features and physical dynamic features, respectively, and q represent the residual dynamic features. att () represents the expected distribution of attention, Θ att Let ψ denote the Hadema product, ψ denote the linear transformation, T denote the length of the input sequence, and W denote the length of the input sequence. r W represents the learned parameters of the detailed features. g The learning parameters, b, represent the physical dynamic characteristics. g Indicates the first bias term, b ψ Indicates the second bias term. The captured representation details features.

[0038] By summing physical dynamic features and detail features, the following is achieved:

[0039]

[0040] Among them, h t+1 Indicates the characteristics after fusion. This indicates the dynamic characteristics of the captured physical components. This represents the dynamic characteristics of the residual.

[0041] Finally, the physical features and detail features are fused through a summation operation and then restored to the value space R of the observation data through the decoder D to obtain x. t+1 Complete the prediction:

[0042]

[0043] in, h represents the predicted value. t+1 D represents the fused features, and D() represents the decoding operation.

[0044] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A weather prediction method based on a physical attention spatio-temporal prediction network, characterized in that, Comprise: S1: from the dynamic system covering the fixed area, acquire historical data at a fixed frequency, define a spatio-temporal prediction problem according to the acquired historical data; S2: design a physical component extraction unit PhyCell and a physical decoupling network framework PhyDNet according to the spatio-temporal prediction problem, decouple the spatio-temporal prediction problem into a physical component and a residual component through the physical decoupling network framework PhyDNet, and capture the physical dynamic characteristics of the physical component through the physical component extraction unit PhyCell; S3: introduce an attention gate structure to identify interesting feature parts and irrelevant feature parts in the residual component, eliminate the irrelevant feature parts, obtain detailed features, and fuse the detailed features by using additive attention to obtain residual dynamic features; Fusing the detailed features by using additive attention to obtain residual dynamic features, comprising: wherein, denotes residual dynamic feature, a denotes additive attention, a = σ2(q att (r, g; Θ att )), σ2() denotes Sigmoid activation function, r, g respectively denote detail feature, physical dynamic feature, q att () denotes attention distribution expectation, Θ att denotes Hadamard product, ψ denotes linear transformation, T denotes length of input sequence, W r denotes learning parameter of detail feature, W g denotes learning parameter of physical dynamic feature, b g denotes first bias term, b ψ denotes second bias term, captured detail feature; S4: Through the sum operation fusion of the physical dynamic characteristics and the residual dynamic characteristics, and through the decoder to restore to the value space R of the observation data, x t+1 , the prediction is completed.

2. The weather prediction method based on the physical attention spatio-temporal prediction network according to claim 1, characterized in that, Defining a spatio-temporal prediction problem, comprising: where h denotes the defined spatio-temporal prediction problem, denotes the sequence to be predicted, X t-T+1 ,…,X t-1 denotes the input data sequence driving the prediction, X t ,…,X t+1 denotes the input data sequence driving the prediction, X t+2 ,…,X t+K denotes the input data sequence taking values in R C×M×N denotes the historical data observed in a M x N coverage area, R denotes the value space of the observations, t denotes the current time node, K denotes the length of the output sequence, T denotes the length of the input sequence, and C denotes the physical quantity included in the data at each grid point.

3. The weather prediction method based on the physical attention spatio-temporal prediction network according to claim 1, characterized in that, Decoupling the spatio-temporal prediction problem into a physical component and a residual component through the physical decoupling network framework PhyDNet, comprising: h = h p + h r where h denotes the defined spatio-temporal prediction problem, h p and h r denote the physical component and the residual component, respectively.

4. The weather prediction method based on the physical attention spatio-temporal prediction network according to claim 1, characterized in that, Capturing the physical dynamic characteristics of the physical component through the physical component extraction unit PhyCell, comprising: wherein, represents a captured physical component dynamic feature, represents a current time physical component dynamic feature, x t represents a current system observation state, E() represents an encoder, f p () represents a physical component learning unit.

5. The weather prediction method based on the physical attention spatio-temporal prediction network according to claim 1, characterized in that, Introducing an attention gate structure to identify interesting feature parts and irrelevant feature parts in the residual component, irrelevant feature parts, to obtain detailed features, comprising: wherein, denotes a captured detail feature, denotes a current time detail feature, x t denotes a current system observation state, E() denotes an encoder, f r () denotes a residual component learning unit.

6. The weather prediction method based on the physical attention spatio-temporal prediction network according to claim 1, characterized in that, Fusing the physical dynamic characteristics and the detailed features by summing operation, comprising: wherein h t+1 represents the fused features, represents the captured physical component dynamic features, represents the residual dynamic features.

7. The weather prediction method based on the physical attention spatio-temporal prediction network according to claim 1, characterized in that, By decoding the decoder is restored to the observed data value space R, x t+1 , complete the prediction, including: wherein, denotes the predicted value, h t+1 denotes the fused feature, D() denotes the decoding operation.

Images