Long-term climate prediction method based on dynamic graph neural network model

Abstract

The invention provides a long-term climate prediction method based on a dynamic graph neural network model, and belongs to the technical field of deep learning and time series prediction.The long-term climate prediction method comprises the steps that firstly, a climate time series is acquired, and climate data sets suitable for training are constructed and comprise a training set, a verification set and a test set; inputting the climate time sequence into a deep learning model, simulating complex spatial-temporal characteristics among climate elements, and generating future long-term climate prediction data; adjusting parameters according to the expression of the model on the verification set; testing the model on the test set, calculating an error between a prediction result and real climate data, and further optimizing the model through an error feedback mechanism; according to the method, the complex time-space relationship in the climate system can be effectively modeled, so that high-precision long-term climate prediction is realized; compared with an existing method, the method has the advantages that the method is excellent in long-time-span climate prediction task, and a more valuable climate prediction service is provided for practical application.

Description

Technical Field

[0001] The present invention belongs to the technical field of deep learning and time series prediction, and specifically, relates to a long-term climate prediction method based on a dynamic graph neural network model. Background Art

[0002] In climate prediction and environmental monitoring, long-term climate prediction is a crucial task. Especially as global climate change becomes increasingly significant, accurate predictions of climate trends in the coming months or even years have become an urgent need for all sectors of society. However, due to the complexity and variability of the climate system, achieving long-term and accurate climate predictions still faces huge challenges. At present, methods for long-term climate prediction can be mainly divided into two categories: one is climate prediction based on traditional statistical methods, and the other is climate prediction methods based on machine learning.

[0003] Climate prediction based on traditional statistical methods usually relies on the statistical characteristics of historical climate data, and predicts future climate trends through time series analysis and regression models. This type of method is simple and easy to use, and performs well when the amount of data is small, but its main disadvantage is that it is difficult to effectively capture the complex dynamic relationship between climate elements. Especially when faced with long-term prediction tasks, the prediction accuracy of traditional methods is usually not ideal, and it is difficult to cope with the diversity and uncertainty of climate change.

[0004] Climate prediction methods based on machine learning have become a hot topic in recent years. In particular, with the rapid development of deep learning technology, modeling complex spatiotemporal patterns in climate data through neural networks has become the mainstream. Compared with traditional methods, such methods can more accurately capture the interactions between climate elements, thereby generating more accurate prediction results. However, existing climate prediction methods based on deep learning also have some shortcomings. For example, most methods can only handle prediction tasks with short time spans. For climate predictions with long time spans, the model's memory capacity and ability to capture complex spatiotemporal patterns are still limited. In addition, deep learning models have high requirements for data quality. When faced with high-noise and multimodal climate data, the reliability and accuracy of prediction results are often not guaranteed.

[0005] In summary, current long-term climate prediction methods face challenges in the following aspects:

[0006] 1. Insufficient prediction accuracy: Existing climate prediction models often find it difficult to maintain high accuracy when dealing with climate change over long time spans, especially when predicting climate trends over months or even years, as the prediction results are subject to greater uncertainty.

[0007] 2. Trade-off between internal and external geographic locations: Conventional deep learning models often focus unilaterally on the time series characteristics within the channel or the time series relationship between channels in time series prediction, and cannot take both into account. Summary of the invention

[0008] In view of the above problems, the present invention proposes a long-term climate prediction method based on a dynamic graph neural network model. The method can effectively model the complex spatiotemporal relationships in the climate system through a dynamic graph neural network, thereby achieving high-precision long-term climate prediction.

[0009] The present invention is achieved through the following technical solutions:

[0010] A long-term climate prediction method based on a dynamic graph neural network model:

[0011] The method specifically comprises the following steps:

[0012] Step 1: Collect and process historical climate data, obtain climate time series, and construct a climate dataset suitable for training, including training set, validation set, and test set;

[0013] Step 2: Input the climate time series from step 1 into the deep learning model to simulate the complex spatiotemporal characteristics between climate elements;

[0014] Step 3: Generate future long-term climate prediction data through the spatiotemporal characteristics obtained in step 2;

[0015] Step 4: Adjust parameters based on the model’s performance on the validation set.

[0016] Step 5: Test the model on the test set, calculate the error between the predicted results and the real climate data, and further optimize the model through the error feedback mechanism.

[0017] Furthermore, the deep learning model consists of three core modules: a spatial interaction modeling module, a temporal feature extraction module, and a fusion prediction module;

[0018] The spatial interaction modeling module is responsible for capturing the dynamic interaction relationship between geographical locations from the input climate data;

[0019] The temporal feature extraction module is used to capture the changing relationship of climate elements over time;

[0020] The fusion prediction module fuses the features output by the spatial interaction modeling module and the temporal feature extraction module to generate future climate prediction results.

[0021] Further, in step 1, the historical climate data is two-dimensional time series data, wherein the first dimension is the time point and the second dimension is the geographical coordinate; each value represents the daily average value of the climate element of the corresponding geographical location on that day;

[0022] In order to improve the convergence speed and prediction accuracy of the model, the original climate data needs to be standardized and the daily mean climate value of each geographical location needs to be converted into a standardized numerical range to eliminate the dimensional differences between different climate elements.

[0023] Furthermore, in step 2, the deep learning model inputs a sequence length of 365, representing the data of the past year, and converts each input univariate time series x i The strategy is split into 12 blocks of sequences, representing the data for each month;

[0024] That is, the input standardized two-dimensional time series x∈R T×N is divided into x p ∈R P×N , where R is the real number domain, T is the total length of the input time series, P is the size of each time block, N is the number of time blocks, and then the channel information is fused with the batch dimension.

[0025] Furthermore, in step 2, the spatial interaction modeling module uses A∈R N×N The matrix A represents a directed graph of climate relationships among N geographic locations, and the matrix A will be optimized as the model is trained: A ij =ReLU(tanh(βET gc ))

[0026] where E∈R N×d , trainable parameters T gc ∈R d×d ; ReLU is the activation function; A ij Represents the connection relationship between point i and point j in the adjacency matrix A; tanh is the activation function and β is the hyperparameter;

[0027] In graph neural networks, given an adjacency matrix A∈R containing N nodes N×N and the initial state Node features can be calculated through K-hop neighbor aggregation:

[0028]

[0029] Where A is the original graph adjacency matrix, is the normalized graph adjacency matrix, is the output feature matrix, φ∈R D×Dis a trainable transformation matrix used to adjust the final feature representation. This approach can effectively capture information from the global neighbor structure and transform the features through the matrix Φ to obtain the final node feature representation.

[0030] Considering the case where a fixed integration time and a small step size are applied, the model can be transformed as follows, replacing the propagation step k∈{0,…,K-1} with a continuous variable

[0031] H G (t+Δt cgp )=H G (t)+Δt cgp (AI N )H G (t)#

[0032] =[(1-Δt cgp )I N +Δt cgp A]H G (t)

[0033] H G is the feature matrix representing the potential state at a specific aggregation step, Δt cgp is the single propagation step length,

[0034] The continuous dynamics of graph propagation is expressed using ODE as:

[0035]

[0036] The initialization state H G is the intermediate state of the continuous time aggregation process. To further reduce numerical errors, the model uses an attention transformation instead of the linear mapping:

[0037]

[0038] ODESolve is a black-box ordinary differential equation solver; G (t i ) represents a specific intermediate state in the process of graph information propagation; H out As a temporal feature z based on spatial interaction modeling int .

[0039] Furthermore, in step 2, the temporal feature extraction module uses TransformerEncoder to map the climate series into a latent representation:

[0040] c d =W p x p +Wpos

[0041] Where W p ∈R D×P is a trainable linear projection that maps the input time series block into a d-dimensional latent space and adds a learnable positional encoding W pos ∈R D×N Used to monitor the temporal order of sequence chunks; temporal potential representation x d As input to TransformerEncoder;

[0042] In the Transformer Encoder, each attention head h=1 in the multi-head attention, H will input x d Convert to query matrix Q h , key matrix K h and value matrix V h :

[0043]

[0044] is the query weight matrix of h attention heads, is the Key weight matrix of the h-th attention head, is the Value weight matrix of the h-th attention head,

[0045] After the normalized product operation, the attention output is obtained:

[0046]

[0047] In this multi-head attention module, a batch normalization layer and a feed-forward network are also included, which are formed by residual connections; then, the multi-head attention generates z inn , and is input into the fusion prediction module as a private temporal feature representing the internal channel.

[0048] Furthermore, in step 3, the prediction module is first fused using three different linear layers ψ q ,ψ k ,ψ v To convert z int and z inn to three embedding representations: ψ q (z int ),ψ k (z inn ),ψ v (z inn ); Then, the channel-level similarity matrix M is calculated using softmax after matrix multiplication:

[0049] M=F softmax (ψ q (z int )⊙ψ k (z inn ))

[0050] Through v (z inn ) and M T , and then apply a linear layer ω c To convert the size of the above feature map; finally, the output is obtained by matrix addition:

[0051] H mix =ω c (ψ v (z inn )⊙M T )+z int

[0052] The model contains a time series prediction module including a time decoder and a projection function, which aggregates the features H C Input to the Time Series Decoder, which maps the latent embedding to the original data space; finally, the model uses a projection function based on a feedforward network to make predictions;

[0053] The temporal decoder first aggregates the features H mix The input is normalized through the layer to obtain a normalized embedding; then, a masked multi-head self-attention layer MHSA and residual connections are used to obtain

[0054]

[0055] Where W q ,W k , W v are three different linear layers ψ q ,ψ k ,ψ v The corresponding linear projection;

[0056] Finally, the learned features is fed into a projection function for future predictions:

[0057]

[0058] in Represents the final prediction result, W p and b p are learnable parameters.

[0059] Furthermore, in step 5, the mean squared error (MSE) loss is used to measure the difference between the predicted value and the true value; the loss on each channel is collected and averaged over the time series to obtain the overall target loss L:

[0060]

[0061] E x is the expected value representing the average loss over the entire data distribution, y i is the predicted value of the i-th sample, that is, the output of the model, x i is the true value of the i-th sample, that is, the label data or target value,

[0062] The goal of the model is to minimize the MSE between the predicted value and the true value to improve the accuracy of propagation prediction.

[0063] An electronic device comprises a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the above method when executing the computer program.

[0064] A computer-readable storage medium is used to store computer instructions, and when the computer instructions are executed by a processor, the steps of the above method are implemented.

[0065] Beneficial effects of the present invention

[0066] Compared with existing methods, the present invention performs well in long-term climate forecasting tasks by introducing dynamic graph neural networks, achieves accurate prediction of future climate trends, meets practical application needs such as social decision-making, agricultural planning, and energy management, and provides more valuable climate forecasting services for the actual effectiveness of climate forecasting in responding to global climate change. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 The figure is a flow chart of the method of the present invention.

[0068] Figure 2 This is a structural diagram of the long-term climate series prediction model based on deep learning. DETAILED DESCRIPTION

[0069] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0070] The experimental methods used in the following examples are conventional methods unless otherwise specified. The materials, reagents, methods and instruments used are conventional materials, reagents, methods and instruments in the art unless otherwise specified, and can be obtained through commercial channels by those skilled in the art.

[0071] Combination Figure 1 to Figure 2 ,This invention proposes a long-term climate prediction method based on a dynamic graph neural network model,

[0072] A long-term climate prediction method based on a dynamic graph neural network model:

[0073] like Figure 1 As shown, the method specifically comprises the following steps:

[0074] Step 1: Collect and process historical climate data, obtain climate time series, and construct a climate dataset suitable for training, including training set, validation set, and test set;

[0075] Historical (raw) climate data may be stored in a variety of formats, such as CSV, NetCDF, etc. First, the historical climate data needs to be uniformly converted into a two-dimensional time series format suitable for model input. Among them, the time dimension represents continuous time points (such as daily), and the geographic dimension represents different geographic coordinates (such as longitude and latitude). The value corresponding to each location represents the daily average climate value of that location, such as temperature, precipitation, etc.

[0076] That is, the historical climate data used in the present invention is two-dimensional time series data, wherein the first dimension is the time point and the second dimension is the geographical coordinates. Each value represents the daily average value of the climate element of the corresponding geographical location.

[0077] In order to improve the convergence speed and prediction accuracy of the model, the original climate data needs to be standardized. Specifically, the daily mean climate value of each geographical location is converted into a standardized numerical interval (such as [0,1]) to eliminate the dimensional differences between different climate elements.

[0078] The standardization formula is as follows:

[0079]

[0080] Where x is the original climate data value, min(x) and max(x) represent the minimum and maximum values ​​in the data set, respectively. The standardized time series data is used in the following text.

[0081] The overall architecture of the long-term climate prediction model proposed in this invention is as follows: Figure 2As shown in Figure 1, the model mainly consists of three core modules: spatial interaction modeling module, temporal feature extraction module, and fusion prediction module. These modules work together to extract spatiotemporal features from input climate data and generate high-precision climate prediction results.

[0082] The spatial interaction modeling module is responsible for capturing the dynamic interaction relationship between geographical locations from the input climate data. This module uses dynamic graph neural networks to model climate elements at different spatiotemporal locations and generate feature maps with spatiotemporal associations.

[0083] The temporal feature extraction module aims to capture the temporal relationship of climate elements. This module uses a Transformer-based architecture to model dependencies in the temporal dimension, thereby extracting the temporal features of each geographic location.

[0084] The fusion prediction module fuses the features output by the spatial interaction modeling module and the temporal feature extraction module to generate future climate prediction results.

[0085] The goal of the model is to improve the prediction accuracy by minimizing the error between the prediction results and the real climate data. The workflow of the model is as follows:

[0086] Q = SpatialEncoder(X)

[0087] K, V = Temporal Extractor (X)

[0088] Y=Decoder(Q,K,V)

[0089] Among them, SpatialEncoder}(·) represents the spatial interaction modeling module, TemporalExtractor(·) represents the temporal feature extraction module, and Decoder(·) represents the fusion prediction module. Through this process, the present invention can generate climate prediction data that is highly similar to the actual climate change trend, providing reliable technical support for long-term climate prediction.

[0090] Step 2: Input the climate time series from step 1 into the deep learning model to simulate the complex spatiotemporal characteristics between climate elements;

[0091] The deep learning model inputs a sequence length of 365, representing the data of the past year, and converts each input univariate time series x i The strategy of splitting into 12 sequence blocks represents the data of each month. This method is beneficial to the model's feature learning and extraction of time series by providing richer contextual information.

[0092] Specifically, the input time series x∈R T×N is divided into xp ∈R P×N , where R is the real number domain, T is the total length of the input time series, P is the size of each time block, N is the number of time blocks, and then the channel information is fused with the batch dimension. The model uses a self-attention network to learn the representation of time blocks and the internal view of time blocks through this block operation.

[0093] In step 2, the spatial interaction modeling module uses A∈R N×N The matrix A represents a directed graph of climate relationships among N geographic locations, and the matrix A will be optimized as the model is trained: A ij =ReLU(tanh(βET gc ))

[0094] where E∈R N×d and trainable parameters T gc ∈R d×d . ReLU is the activation function. ij Represents the connection between point i and point j in the adjacency matrix A. tanh is the activation function and β is the hyperparameter;

[0095] In graph neural networks, given an adjacency matrix A∈R containing N nodes N×N and the initial state Node features can be calculated through K-hop neighbor aggregation:

[0096]

[0097] Where A is the normalized graph adjacency matrix, is the normalized graph adjacency matrix, is the output feature matrix, φ∈R D×D is a trainable transformation matrix used to adjust the final feature representation. This approach can effectively capture information from the global neighbor structure and transform the features through the matrix Φ to obtain the final node feature representation.

[0098] Considering the case where a fixed integration time and a small step size are applied, the model can be transformed as follows, replacing the propagation step k∈{0,…,K-1} with a continuous variable

[0099] H G (t+Δt cgp )=H G (t)+Δt cgp (AI N )H G (t)#

[0100] =[(1-Δt cgp )I N+Δt cgp A]H G (t)

[0101] H G is the feature matrix representing the potential state at a specific aggregation step, Δt cgp is the single propagation step length,

[0102] The continuous dynamics of graph propagation is expressed using ODE as:

[0103]

[0104] The initialization state H G is the intermediate state of the continuous time aggregation process. To further reduce numerical errors, the model uses an attention transformation instead of the linear mapping:

[0105]

[0106] ODESolve is a black-box ordinary differential equation solver. G (t i ) represents a specific intermediate state in the process of graph information propagation.

[0107] In order to learn from rich temporal information, the spatial ordinary differential equations are regarded as the internal process of the proposed temporal neural ordinary differential equations, which enables the model to accurately simulate the stable dynamics of multivariate time series from both spatial and temporal perspectives. The model stacks multiple residual convolution blocks to extract and aggregate temporal patterns in a non-recursive manner:

[0108]

[0109] Where TCN(·,Θ l ) is a parameterized temporal convolutional layer, Represents a truncation function.

[0110] The model is constructed by This is achieved to decouple the connection between the aggregation depth L and the integration time T. In this way, given a desired termination time T and an initial state The entire continuous temporal aggregation process can be characterized by letting Δt→0 using a set of parameters:

[0111]

[0112] The temporal aggregation process is the discretized form of the following ordinary differential equation (ODE):

[0113]

[0114] The input sequence is mapped to the latent space through independent convolutional layers. Initial state The model uses a gating mechanism to control the information flow in each integration step to implement TCN:

[0115] TCN(H T (t),θ)=f c (H T (t),Θ c )⊙f G (H T (t),Θ g )

[0116] where f c (·,Θ c ) and f G (·,Θ g ) are filtering and gated convolutions, which have similar network structures but different parameters and nonlinearities:

[0117]

[0118] where σ(·) represents the sigmoid activation function, * δ Represents a convolution operation with a scalable dilation definition.

[0119] Given two black-box ODE solvers, ODESolve 1 (·) and ODESolve 2 (·), the learned spatiotemporal representation of the input sequence can be obtained by integrating get:

[0120]

[0121] in, Based on the above two propositions, the definitions are as follows:

[0122]

[0123] In the above equation, the inner ODE solution and attention transformation, i.e., A(·,Φ), gives Let H G (0) = TCN (H (t), t, θ). By solving the 12 time series patches once using ODE, finally, H out As a temporal feature z based on spatial interaction modeling int .

[0124] The temporal feature extraction module uses Transformer Encoder to map the climate series into a latent representation:

[0125] x d =W p x p +W pos

[0126] Where W p ∈R D×P is a trainable linear projection that maps the input time series block into a d-dimensional latent space and adds a learnable positional encoding W pos ∈R D×N It is used to monitor the temporal order of sequence chunks. The temporal latent representation x d As the input of Transformer Encoder. In Transformer Encoder, each attention head h = 1 in the multi-head attention, H will input x d Convert to query matrix, key matrix and value matrix:

[0127]

[0128] is the query weight matrix of h attention heads, is the Key weight matrix of the h-th attention head, is the Value weight matrix of the h-th attention head,

[0129] After the normalized product operation, the attention output is obtained:

[0130]

[0131] In this multi-head attention module, a batch normalization layer and a feed-forward network are also included, which are formed by residual connections. After that, the multi-head attention generates z inn , and is input into the internal and external fusion prediction module as the private temporal features representing the internal part of the channel.

[0132] Step 3: Generate future long-term climate prediction data through the spatiotemporal characteristics obtained in step 2;

[0133] In step 3, the features output by the spatial interaction modeling module and the temporal feature extraction module are fused through the fusion prediction module to generate future climate prediction results;

[0134] First, the model uses three linear layers ψ q ,ψ k ,ψ v To convert z int and z inn to three embedding representations: ψ q (zint ),ψ k (z inn ),ψ v (z inn ). Then, the channel-level similarity matrix M is calculated using softmax after matrix multiplication:

[0135] M=F softmax (ψ q (z int )⊙ψ k (z inn ))

[0136] Through v (z inn ) and M T , and then apply a linear layer ω c To convert the size of the above feature map. Finally, the output is obtained by matrix addition:

[0137] H mix =ω c (ψ v (z inn )⊙M T )+z int

[0138] The model contains a time series prediction module including a time decoder and a projection function, which aggregates the features H C The input is sent to the Time Series Decoder, which maps the latent embedding to the original data space. Finally, the model uses a projection function based on a feedforward network to make predictions.

[0139] The temporal decoder first aggregates the features H mix The input is normalized through the layer to obtain a normalized embedding. Then, a masked multi-head self-attention layer MHSA and residual connections are used to obtain

[0140]

[0141] Where W q ,W k ,W v are three different linear layers ψ q ,ψ k ,ψ v The corresponding linear projection. Finally, the learned features is fed into a projection function for future predictions:

[0142]

[0143] in Represents the final prediction result, W p and b p are learnable parameters.

[0144] Step 4: Adjust parameters based on the model’s performance on the validation set.

[0145] The parameter adjustment includes but is not limited to: manual hyperparameter adjustment, or using automated hyperparameter optimization techniques (such as grid search, random search, Bayesian optimization, etc.) to automatically adjust hyperparameters such as learning rate, batch size, number of model layers, number of hidden units, etc.

[0146] In each training iteration, the model's parameters such as weights and biases are updated according to the gradient of the loss function. This is achieved through the back-propagation algorithm and optimizers (such as SGD, Adam, etc.).

[0147] Step 5: Test the model on the test set, calculate the error between the predicted results and the real climate data, and further optimize the model through the error feedback mechanism.

[0148] The mean squared error (MSE) loss is used to measure the difference between the predicted value and the true value. The loss on each channel is collected and averaged over the time series to get the overall target loss:

[0149]

[0150] E x is the expected value representing the average loss over the entire data distribution, y i is the predicted value of the i-th sample, that is, the output of the model, x i is the true value of the i-th sample, that is, the label data or target value. The goal of the model is to minimize the MSE between the predicted value and the true value to improve the accuracy of propagation prediction.

[0151] An electronic device comprises a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the above method when executing the computer program.

[0152] A computer-readable storage medium is used to store computer instructions, and when the computer instructions are executed by a processor, the steps of the above method are implemented.

[0153] The memory in the embodiments of the present application may be a volatile memory or a non-volatile memory, or may include both volatile and non-volatile memories. Among them, the non-volatile memory may be a read-only memory, ROM, a programmable read-only memory, PROM, an erasable programmable read-only memory, EPROM, an electrically erasable programmable read-only memory, EEPROM, or a flash memory. The volatile memory may be a random access memory, RAM, which is used as an external cache. By way of example but not limitation, many forms of RAM are available, such as static random access memory static RAM, SRAM, dynamic random access memory dynamic RAM, DRAM, synchronous dynamic random access memory synchronous DRAM, SDRAM, double data rate synchronous dynamic random access memory double data rate SDRAM, DDR SDRAM, enhanced synchronous dynamic random access memory enhanced SDRAM, ESDRAM, synchronous link dynamic random access memory synchlink DRAM, SLDRAM, and direct memory bus random access memory direct rambus RAM, DR RAM. It should be noted that memory of the methods described herein is intended to include, but is not limited to, these and any other suitable types of memory.

[0154] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented by software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer instructions are loaded and executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from a website site, computer, server or data center through a wired method such as coaxial cable, optical fiber, digital subscriber line digital subscriber line, DSL or wireless such as infrared, wireless, microwave, etc. to another website site, computer, server or data center. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or data center that includes one or more available media integrated. The available medium can be a magnetic medium such as a floppy disk, a hard disk, a tape, an optical medium such as a high-density digital video disc digital video disc, DVD, or a semiconductor medium such as a solid state hard disk solid state disc, SSD, etc.

[0155] In the implementation process, each step of the above method can be completed by an integrated logic circuit of hardware in a processor or an instruction in the form of software. The steps of the method disclosed in conjunction with the embodiment of the present application can be directly embodied as a hardware processor for execution, or a combination of hardware and software modules in a processor for execution. The software module can be located in a storage medium mature in the art such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory or an electrically erasable programmable memory, a register, etc. The storage medium is located in a memory, and the processor reads the information in the memory and completes the steps of the above method in conjunction with its hardware. To avoid repetition, it is not described in detail here.

[0156] It should be noted that the processor in the embodiment of the present application can be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above method embodiment can be completed by an integrated logic circuit of hardware in the processor or an instruction in the form of software. The above processor can be a general-purpose processor, a digital signal processor DSP, an application-specific integrated circuit ASIC, a field programmable gate array FPGA or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components. The methods, steps and logic block diagrams disclosed in the embodiments of the present application can be implemented or executed. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor, etc. The steps of the method disclosed in the embodiment of the present application can be directly embodied as a hardware decoding processor to perform, or the hardware and software modules in the decoding processor can be combined to perform. The software module can be located in a mature storage medium in the field such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory or an electrically erasable programmable memory, a register, etc. The storage medium is located in a memory, and the processor reads the information in the memory and completes the steps of the above method in combination with its hardware.

[0157] The above is a detailed introduction to the long-term climate prediction method based on the dynamic graph neural network model proposed in the present invention, and the principles and implementation methods of the present invention are explained. The description of the above embodiments is only used to help understand the method of the present invention and its core idea; at the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation method and application scope. In summary, the content of this specification should not be understood as a limitation on the present invention.

Claims

1. A long-term climate prediction method based on a dynamic graph neural network model, characterized in that: The method specifically comprises the following steps: Step 1: Collect and process historical climate data, obtain climate time series, and construct a climate dataset suitable for training, including training set, validation set, and test set; Step 2: Input the climate time series from step 1 into the deep learning model to simulate the complex spatiotemporal characteristics between climate elements; Step 3: Generate future long-term climate prediction data through the spatiotemporal characteristics obtained in step 2; Step 4: Adjust parameters based on the model’s performance on the validation set. Step 5: Test the model on the test set, calculate the error between the predicted results and the real climate data, and further optimize the model through the error feedback mechanism.

2. The prediction method according to claim 1, characterized in that: The deep learning model consists of three core modules: spatial interaction modeling module, temporal feature extraction module and fusion prediction module; The spatial interaction modeling module is responsible for capturing the dynamic interaction relationship between geographical locations from the input climate data; The temporal feature extraction module is used to capture the changing relationship of climate elements over time; The fusion prediction module fuses the features output by the spatial interaction modeling module and the temporal feature extraction module to generate future climate prediction results.

3. The prediction method according to claim 2, characterized in that: In step 1, the historical climate data is two-dimensional time series data, wherein the first dimension is the time point and the second dimension is the geographical coordinates; each value represents the daily average value of the climate element of the corresponding geographical location on that day; In order to improve the convergence speed and prediction accuracy of the model, the original climate data needs to be standardized and the daily average climate value of each geographical location needs to be converted into a standardized numerical range to eliminate the dimensional differences between different climate elements.

4. The prediction method according to claim 3, characterized in that: In step 2, the deep learning model inputs a sequence of length 365, representing the data for the past year, and converts each input univariate time series x i The strategy is split into 12 blocks of sequences, representing the data for each month; That is, the input standardized two-dimensional time series x∈R T×N is divided into x p ∈R P×N , where R is the real number domain, T is the total length of the input time series, P is the size of each time block, N is the number of time blocks, and then the channel information is fused with the batch dimension.

5. The prediction method according to claim 4, characterized in that: In step 2, the spatial interaction modeling module uses A∈R N×N The matrix A represents a directed graph of climate relationships among N geographic locations, and the matrix A will be optimized as the model is trained: A ij =ReLU(tanh(βET gc )) where E∈R N×d , trainable parameters T gc ∈R d×d ; ReLU is the activation function; A ij Represents the connection relationship between point i and point j in the adjacency matrix A; tanh is the activation function and β is the hyperparameter; In graph neural networks, given an adjacency matrix A∈R containing N nodes N×N and the initial state Calculate node features through K-hop neighbor aggregation: Where A is the original graph adjacency matrix, is the normalized graph adjacency matrix, is the output feature matrix, φ∈R D×D It is a trainable transformation matrix used to adjust the final feature representation; Considering the case where a fixed integration time and a smaller step size are applied, the propagation steps k∈0,…,K-1} are replaced by the continuous variable H G (t+Δt cgp )=H G (t)+Δt cgp (A-I N )H G (t)#=[(1-Δt cgp )I N +Δt cgp A]H G (t) H G is the feature matrix representing the potential state at a specific aggregation step, Δt cgp is the single propagation step length, The continuous dynamics of graph propagation is expressed using ODE as: The initialization state H G is the intermediate state of the continuous time aggregation process; to further reduce numerical errors, the model uses an attention transformation instead of the linear mapping: ODESolve is a black-box ordinary differential equation solver; G (t i ) represents a specific intermediate state in the process of graph information propagation; H out As a temporal feature z based on spatial interaction modeling int .

6. The prediction method according to claim 5, characterized in that: In step 2, the temporal feature extraction module uses Transformer Encoder to map the climate series into a latent representation: x d =W p x p +W pos Where W p ∈R D×P is a trainable linear projection that maps the input time series block into a d-dimensional latent space and adds a learnable positional encoding W pos ∈R D×N Used to monitor the temporal order of sequence blocks; Temporal latent representation x d As input to TransformerEncoder; In the Transformer Encoder, each attention head h=1 in the multi-head attention, H will input x d Convert to query matrix Q h , key matrix K h and value matrix V h : is the query weight matrix of h attention heads, is the Key weight matrix of the h-th attention head, is the Value weight matrix of the h-th attention head, After the normalized product operation, the attention output is obtained: In this multi-head attention module, a batch normalization layer and a feed-forward network are also included, which are formed by residual connections; then, the multi-head attention generates z inn , and is input into the fusion prediction module as a private temporal feature representing the internal channel.

7. The prediction method according to claim 6, characterized in that: In step 3, the prediction module is first fused using three different linear layers ψ q ,ψ k ,ψ v To convert z int and z inn to three embedding representations: ψ q (z int ),ψ k (z inn ),ψ v (z inn ); Then, the channel-level similarity matrix M is calculated using softmax after matrix multiplication: M=F softmax (ψ q (z int )⊙ψ k (z inn )) Through v (z inn ) and M T , and then apply a linear layer ω c To convert the size of the above feature map; finally, the output is obtained by matrix addition: H mix =ω c (ψ v (z inn )⊙M T )+z int The model contains a time series prediction module including a time decoder and a projection function, which aggregates the features H C Input to the Time Series Decoder, which maps the latent embedding to the original data space; finally, the model uses a projection function based on a feedforward network to make predictions; The temporal decoder first aggregates the features H mix The input is normalized through the layer to obtain a normalized embedding; then, a masked multi-head self-attention layer MHSA and residual connections are used to obtain Where W q ,W k ,W v are three different linear layers ψ q ,ψ k ,ψ v The corresponding linear projection; Finally, the learned features is fed into a projection function for future predictions: in Represents the final prediction result, W p and b p are learnable parameters.

8. The prediction method according to claim 7, characterized in that: In step 5, the mean squared error (MSE) loss is used to measure the difference between the predicted value and the true value; the loss on each channel is collected and averaged over the time series to obtain the overall target loss L: E x is the expected value representing the average loss over the entire data distribution, y i is the predicted value of the i-th sample, that is, the output of the model, x i is the true value of the i-th sample, that is, the label data or target value, The goal of the model is to minimize the MSE between the predicted value and the true value to improve the accuracy of propagation prediction.

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 8 are implemented.

10. A computer-readable storage medium for storing computer instructions, characterized in that: When the computer instructions are executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

Images