A multivariate time series joint prediction method for meteorological big data

Abstract

This invention belongs to the field of time series forecasting technology, specifically relating to a multivariate time series joint forecasting method for meteorological big data. The method includes: acquiring historical multivariate features and strictly aligned solar thermal radiation and upper atmospheric dynamic parameters within the target forecast time window; encoding the preprocessed historical multivariate features using a multivariate encoder based on an inverted variable embedding strategy to obtain a multivariate global context representation; inputting the multivariate global context representation into a deep feature refinement module to obtain a refined high-dimensional dynamic representation; and processing the refined high-dimensional dynamic representation using a multi-task decoder based on sparse hybrid expert routing to obtain the forecast data. This invention demonstrates good performance in multivariate physical correlation modeling, long-term dynamic feature extraction, efficient inter-variable information interaction, and multi-task adaptive decoupling.

Description

Technical Field

[0001] This invention belongs to the field of time series forecasting technology, specifically relating to a multivariate time series joint forecasting method for meteorological big data. Background Technology

[0002] The power generation efficiency of clean and renewable energy sources such as wind and solar power directly depends on accurate weather forecasts. Wind power generation requires precise judgment of future wind speed and direction, while photovoltaic power generation relies on accurate forecasts of solar radiation intensity and cloud cover. Therefore, high-precision weather forecasting is not only a technological pursuit but also a core support for energy managers to optimize power generation plans, improve energy utilization efficiency, reduce operating costs, and ensure grid stability. Against the backdrop of the nation's vigorous promotion of energy security and carbon neutrality strategic goals, refined forecasting management based on meteorological big data is of great significance for achieving a stable supply of clean wind and solar energy.

[0003] Weather time-series forecasting is essentially a complex multivariate long-term forecasting problem, and its core challenges are reflected in the following aspects:

[0004] Multivariate physical coupling: There are deep physical connections and dynamic influences among meteorological elements such as temperature, wind speed, and air pressure. Isolated modeling of a single variable is difficult to fully utilize the synergistic information across variables.

[0005] Long-term complex time series dependencies: Meteorological data simultaneously contain short-term fluctuations (such as intraday temperature changes), medium-term trends (such as the passage of weather systems), and long-term periodic patterns (such as seasonal changes). The model needs to effectively capture multi-scale time series dynamics at the same time.

[0006] Spatiotemporal correlation: The meteorological conditions of a single automatic weather station are not only affected by its own historical data, but also closely related to the meteorological field conditions of the surrounding areas. Spatial context information is crucial for improving forecast accuracy.

[0007] Negative transfer problem in multi-task prediction: When predicting two target variables, temperature and wind speed, if the shared information is not handled properly, the mutual interference between tasks may lead to a decrease in the overall performance of the model.

[0008] Traditional statistical methods, such as autoregressive models (AR) and autoregressive moving average models (ARIMA), mainly rely on parametric modeling, making it difficult to effectively capture the highly nonlinear patterns and complex spatiotemporal correlations of meteorological systems. With the development of deep learning technology, sequence models, represented by recurrent neural networks (RNN) and long short-term memory networks (LSTM), have made some progress in meteorological element prediction, but their ability to model ultra-long sequences and their computational efficiency still have significant limitations.

[0009] In recent years, forecasting methods based on the Transformer architecture have demonstrated strong potential in time series forecasting tasks due to their ability to capture long-range dependencies through self-attention mechanisms. However, current mainstream Transformer forecasting methods generally embed multiple meteorological variables at the same time step into a single time token and apply an attention mechanism along the time dimension. This paradigm has inherent flaws: different meteorological elements represent vastly different physical measurements (e.g., temperature is measured in °C, air pressure in Pa), and forcibly combining them into the same token masks the individual characteristics between variables; simultaneously, there is an objective temporal phase shift between the observations of each variable (e.g., air pressure changes precede temperature responses), and time-aligned hybrid embedding introduces invalid noise, causing the attention map to lose its physical interpretability. Furthermore, as the number of forecast variables increases, the computational complexity of the attention mechanism increases quadratically with the number of variables, limiting the scalability of the model in large-scale meteorological data scenarios.

[0010] In multi-task learning, existing methods either employ a completely shared encoding and decoding structure (leading to mutual interference between features of different tasks) or train the model independently for each target variable (lacking the sharing of beneficial information between tasks), both of which make it difficult to achieve a good balance between prediction accuracy and computational efficiency.

[0011] In summary, existing meteorological time-series forecasting methods have significant shortcomings in multivariate physical correlation modeling, long-term dynamic feature extraction, efficient information interaction between variables, and multi-task adaptive decoupling. A new method that can systematically solve these problems is urgently needed. Summary of the Invention

[0012] To address the aforementioned technical problems, this invention provides a multivariate time-series joint forecasting method for meteorological big data, comprising:

[0013] S1: Obtain historical observation data from automatic weather stations and 3×3 grid covariates of ERA5 aligned with them in time and space to form historical multivariate features; at the same time, extract solar thermal radiation and upper atmospheric dynamic parameters that are strictly aligned with the target prediction time window from NWP data as future realistic constraint fields; perform time-series alignment and reversible instance normalization preprocessing on the above data; and input the preprocessed historical multivariate features into the prediction model.

[0014] The prediction model includes: a multivariate encoder based on an inverted variable embedding strategy, a deep feature refinement module, and a multi-task decoder based on sparse hybrid expert routing;

[0015] S2: A multivariate encoder based on an inverted variable embedding strategy is used to encode the preprocessed historical multivariate features. The complete historical sequence of each meteorological element is embedded as an independent variable token. The correlation weight between tokens is calculated in the variable dimension through a multi-head self-attention mechanism to capture the physical dependency between multiple meteorological elements and obtain a multivariate global context representation.

[0016] S3: The multivariate global context representation is input into the deep feature refinement module. The deep feature refinement module first adopts the star aggregation-redistribution module, which aggregates all variable sequence information into a global core representation through random pooling operation. Then, the global core representation is concatenated and fused with the variable sequence representations to achieve centralized information interaction between variables with linear complexity. Subsequently, combined with the temporal convolutional network, the local temporal dynamic features of meteorological elements are extracted based on the established global variable correlation through causal convolution and dilated convolution to obtain the refined high-dimensional dynamic representation.

[0017] S4: A multi-task decoder based on sparse hybrid expert routing processes the refined high-dimensional dynamic representation. The gating network of the multi-task decoder performs cross-modal fusion of the refined high-dimensional dynamic representation with the future reality constraint field, and dynamically allocates expert routing weights based on the model's self-learning features and the prior knowledge of the future physical environment. The sparse hybrid expert routing includes multiple shared expert networks. The gating network, combined with a load balancing strategy without auxiliary loss, allocates dedicated expert networks for decoupled decoding for temperature and wind speed prediction tasks, and jointly outputs the temperature and wind speed prediction results for the next 72 hours.

[0018] S6: During model training, the prediction model is optimized using a time-frequency domain difference loss function, which integrates time-domain smoothing L1 loss, frequency domain loss, and multi-step difference loss.

[0019] The beneficial effects of this invention are:

[0020] This invention employs an inverted variable embedding strategy, treating the complete time series of each meteorological element as an independent token. This allows the self-attention mechanism to directly model the physical correlations between multiple meteorological elements along the variable dimension, avoiding the variable characteristic smoothing problem caused by traditional time token hybrid embedding. In the constructed deep feature refinement module, the star-shaped aggregation-redistribution module achieves efficient centralized information interaction between variables with linear computational complexity, overcoming the quadratic complexity bottleneck of traditional attention mechanisms and effectively suppressing the negative impact of anomalous variables. Simultaneously, the temporal convolutional network further refines the local temporal dynamics through causal dilated convolution. In the multi-task decoding stage, based on S… The wiGLU-gated sparse hybrid expert routing innovatively combines a refined high-dimensional dynamic representation with future real-world constraint fields extracted from NWP forecasts of photothermal and high-altitude dynamics across modal modes. This allows the allocation of expert gating weights to both follow the inertia of historical time-series evolution and be strictly constrained by future known physical constraints, completely overcoming the technical defects of traditional deep learning models that lead to prediction distortion due to pure mathematical extrapolation. Furthermore, it achieves adaptive decoupling of temperature and wind speed prediction tasks by combining an auxiliary loss-free load balancing mechanism. The time-frequency domain differential loss function integrates multi-domain optimization objectives, significantly improving the model's ability to learn complex meteorological time-series patterns and effectively alleviating the problem of accumulated errors in long-term predictions. Attached Figure Description

[0021] Figure 1 This is a flowchart of a multivariate time-series joint forecasting method for meteorological big data provided by the present invention. Detailed Implementation

[0022] 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.

[0023] refer to Figure 1 , Figure 1 The overall flowchart of the multivariate time-series joint forecasting method for meteorological big data provided in this embodiment of the invention specifically includes:

[0024] Step 101: Data Acquisition and Preprocessing.

[0025] Historical observation data from global automatic weather stations were acquired through a meteorological data acquisition system. The dependent variables included two types of meteorological elements: temperature (°C) at 2 meters altitude and absolute wind speed (m / s) at 2 meters altitude. The historical observation window length was set to L = 168 time steps (corresponding to 7 days, with sampling intervals of 1 hour). Simultaneously, 3×3 grid covariates from ERA5 data aligned spatiotemporally were acquired, including zonal vector wind speed (m / s) at 10 meters altitude, meridional vector wind speed (m / s) at 10 meters altitude, temperature (°C) at 2 meters altitude, and uniform sea level pressure (Pa), with a sampling interval of 1 hour. The automatic station observation data and the ERA5 grid covariates were concatenated to form a historical multivariate feature matrix. ,in L represents the total number of historical channels, and L represents the length of the historical observation window.

[0026] From the NWP data, physical parameters strictly aligned with the target prediction time window are obtained. On one hand, the downward shortwave total radiation (swdown, ), direct normal shortwave radiation (swddni, ) and low cloud cover (clflo); on the other hand, extract multi-level wind speed profiles from 50 to 150 meters (wspd50-150, ) and air density (rho, Together they constitute the future reality constraint field matrix. (in (To constrain the number of parameter channels), H is the future prediction window length.

[0027] In practical applications, ensuring data quality is crucial, including handling missing values, outlier detection, and data smoothing. Common processing methods include linear interpolation and moving average filtering. Subsequently, reversible instance normalization preprocessing is performed on all variables. This involves subtracting the historical mean from each variable sequence and dividing by the standard deviation, ensuring that the mean of each variable sequence is zero and the variance is one. This effectively eliminates the interference of differences in the dimensions of different meteorological elements on model training. In the prediction output stage, inverse normalization is used to restore the prediction results to the original dimension space, ensuring the physical interpretability of the predicted values.

[0028] Step 102: Multivariate encoder based on inverted variable embedding strategy.

[0029] This step transforms the traditional modeling paradigm, which uses time steps as tokens, into a new paradigm, which uses variable sequences as tokens, through an inverted variable embedding strategy. The specific implementation is as follows:

[0030] (1) Inversion and transpose and sequence embedding: Obtaining historical multivariate feature tensors (in The length of the historical sequence, i.e., the number of time steps. (Total number of historical variables formed by automatic stations and ERA5). First, the input features... Perform an inversion and transpose operation along the time axis and variable axis to convert its tensor shape to... That is, each row of the matrix corresponds to a meteorological variable, and each column corresponds to a time step, making the last dimension of the tensor the length of the time series. Subsequently, through a parameter-sharing linear mapping layer, the length of each transposed meteorological variable is... The complete historical time sequence is treated as a whole feature vector and directly projected onto the high-dimensional hidden layer feature space. Finally, Dropout regularization is applied to the projected features to prevent overfitting, resulting in a set of independent variables called Tokens.

[0031]

[0032] in, This is the characteristic tensor after transposition. Each token independently encodes the global historical time-series evolution pattern of a single physical meteorological element, fundamentally avoiding the problems of feature physical attribute smoothing and semantic confusion caused by the mixed embedding of multiple meteorological elements in the same time step.

[0033] (2) Self-attention mechanism: in the variable token set A layer normalization is applied to eliminate the distribution differences among different meteorological elements; then, a multi-head self-attention mechanism is applied to model the correlation between meteorological elements along the token dimension of the variable:

[0034]

[0035] in, All are composed of a set of variable tokens Obtained through linear projection. Attention score matrix. The (i,j)th element reflects the strength of the physical correlation between the i-th and j-th meteorological elements, such as the thermodynamic correlation between temperature and air pressure, and the dynamic coupling between zonal and meridional wind speeds, giving the attention map clear meteorological and physical interpretability. Highly correlated variables (tokens) receive greater weight in subsequent feature aggregation, thereby enabling the targeted transfer of beneficial information between meteorological elements.

[0036] Finally, the attention weight matrix AND-value tensor Perform matrix multiplication:

[0037]

[0038] This step explicitly calculates and extracts deep physical correlations and spatial coupling information between different meteorological elements (such as temperature and longwave radiation, air pressure and wind speed) through a global receptive field.

[0039] (3) Feedforward network: Output of self-attention computation After initial residual summation, intermediate feature representations are obtained. The data is then fed into a variable-level feedforward network layer. To improve computational efficiency and the nonlinear representation of local features, the feedforward network layer uses a one-dimensional convolutional topology with a kernel size of 1 instead of a traditional fully connected layer. The complete mathematical calculation process is as follows:

[0040] First, the input features are normalized using the first layer to obtain normalized features. :

[0041]

[0042] Subsequently, Perform the transpose operation between the sequence dimension and the feature dimension (denoted as ). This satisfies the channel computation requirements of one-dimensional convolution. A first-layer one-dimensional convolution with a kernel size of 1 is used ( ) Reduce the number of feature channels from the hidden layer dimension Expansion to And sequentially pass through a nonlinear activation function ( (e.g., ReLU or GELU) and Dropout regularization yield high-dimensional extended features. :

[0043]

[0044] Next, a second one-dimensional convolution with a kernel size of 1 is applied ( )Will Channel number compressed and restored to The dimensions, after being regularized again using Dropout, are transposed back to their original shape to obtain the refined residual features. :

[0045]

[0046] Finally, normalized features Residual characteristics after refining The residuals are summed, and a second layer of normalization is applied to complete the forward propagation of a single-layer coding block, outputting the features. :

[0047]

[0048] pass The concatenated stacking of the above-mentioned coding blocks ultimately outputs a multivariate global context representation tensor that integrates global variable correlations and nonlinear local features. This serves as the input for the subsequent deep feature refinement module.

[0049] Step 103: Multi-level feature extraction by the decoder.

[0050] The multivariate global context representation output in step 102 Input deep feature refinement module. This module consists of a star-shaped aggregation-redistribution network and a temporal convolutional network connected in series. It aims to further extract cross-variable collaborative information and local temporal evolution patterns based on the established global variable correlations. The specific process is as follows:

[0051] (1) Star-shaped aggregation-redistribution module: Inspired by star-shaped topology networks in software engineering, this module introduces a global core representation to achieve centralized indirect interaction between variables, reducing the complexity from that of traditional attention. Down to The specific steps are as follows:

[0052] ① Feature projection and core aggregation: For input features The core projection representation matrix of each variable is obtained by sequentially passing it through the first linear layer, the GELU activation function, and the second linear layer for dimension mapping. Subsequently, aggregating across the variable dimensions using random pooling operations generates a global core representation. .

[0053] During training, random pooling randomly samples variables based on the Softmax probability of each variable's core projection onto the variable's channel dimension, balancing global mean and extreme value features to enhance the model's generalization ability. During inference, it is replaced by probability-weighted averaging to ensure prediction stability. This random pooling process can be formally represented as:

[0054]

[0055] in, For the first The first variable, token, is the... Core projection value, For the corresponding sampling probability, This is the index of the variable obtained by sampling according to a multinomial distribution.

[0056] ② Multilayer perceptron fusion: This involves fusing the global core state vector... Broadcast copies are performed along the variable dimension and compared with the original input features. By concatenating the feature channels, a joint feature tensor is obtained. Subsequently, dimensionality reduction and deep fusion are performed through a third linear layer, GELU activation, and a fourth linear layer to output star-shaped interaction features. :

[0057]

[0058] (2) Temporal Convolutional Network: Incorporating star-shaped interaction features Input a temporal convolutional network. This network consists of multiple stacked residual temporal modules, designed to extract local dynamic evolution patterns while strictly adhering to temporal causality. For any residual temporal module in the stacked architecture, the input features of that module are: If the porosity of the layer in which the network layer is located is... The size of the one-dimensional convolution kernel is To prevent future information leaks, a length of [value] is applied to the input features. The specific feature extraction process for zero-padding is as follows:

[0059] ① First-layer causal convolution extraction: For the input A one-dimensional dilated convolution operation with weighted normalization is performed. After convolution, a truncation operation is used to strictly remove redundant features whose ends equal to the padding length from the tensor, ensuring that the output at the current time step depends only on the current and historical inputs. Subsequently, ReLU activation and Dropout regularization are performed sequentially to obtain intermediate dynamic features. :

[0060]

[0061] ② Second-layer causal convolution extraction: ... As input, the above-described one-dimensional dilated convolution with weight normalization, truncation operation, ReLU activation, and Dropout regularization are repeated to obtain deep dynamic features. :

[0062]

[0063] ③ Residual Connection and Module Output: If the number of input and output channels of the module is inconsistent, then the initial input... A one-dimensional convolution with a kernel size of 1 is applied for downsampling dimension alignment. Finally, the aligned residual representation is then compared with... Element-wise addition is performed, and the forward propagation of the residual time module is completed by ReLU activation to obtain the module output features. :

[0064]

[0065] The final output is the refined high-dimensional dynamic representation tensor. It serves as the core input for the subsequent multi-task expert routing decoder.

[0066] Step 104: Sparse Hybrid Expert Routing Multi-task Decoding.

[0067] This step employs a multi-task decoder based on sparse hybrid expert routing to perform feature-level concatenation of the refined dynamic representation output from step 103 with the future reality constraint field. While preserving the high-order nonlinear expressive power, it independently allocates optimal expert resources for the temperature and wind speed prediction tasks through a sparse routing mechanism, ultimately outputting the prediction sequence. Its core logic is as follows:

[0068] (1) First, the refined dynamic representation output by the temporal convolutional network is processed. Flattening the vectors yields feature vectors that purely characterize the evolutionary patterns of historical timelines. (in (Based on the fundamental feature dimension). To introduce known physical boundary conditions for the future, a future reality constraint field (including prior parameters such as NWP photothermal radiation and dynamics) aligned with the target prediction time is extracted, and its dimension is nonlinearly mapped to the same dimension through a feature projection layer. To obtain the future constraint representation Subsequently, the two are physically concatenated along the feature channel dimension to construct a cross-modal joint feature vector. :

[0069]

[0070] This joint feature It incorporates both the historical inertia of the meteorological system and future known weather forcings, serving as a unified input for all subsequent gating and expert networks.

[0071] (2) Model configuration ( =8) shared expert networks. To overcome the linear bottleneck of traditional ReLU activation, each expert network adopts an advanced SwiGLU gated feedforward structure. For input features , No. A network of experts It contains three independent weight matrices: gated projection (Gate_proj), up projection (Up_proj), and down projection (Down_proj). The formula for calculating the forward pass of a single sample is as follows:

[0072]

[0073] in, This represents the element-wise product of tensors. The SwiGLU structure uses a built-in gating mechanism to selectively cross-filter cross-modal features, which can improve the model's ability to fit nonlinear abrupt changes in meteorological elements (such as sudden storm onset or temperature drop).

[0074] (3) For the two prediction tasks of temperature and wind speed, construct independent gating networks respectively. Taking the first... There are prediction tasks (of which) Taking the tasks corresponding to temperature and wind speed respectively as an example, its dedicated gating network first relies on... Calculation for all Initial probability distribution of the experts:

[0075]

[0076] To improve reasoning efficiency and prevent knowledge redundancy, a sparse Top-K selection strategy (K=2) is introduced. Only retain... The weights of the K experts with the highest scores are determined, the weights of the unselected experts are reset to zero, and the weights of the K selected experts are renormalized to obtain sparse gating weights. Subsequently, the outputs of the activated K experts are weighted and summed using this sparse weight to obtain the ... The dedicated fusion feature tower input for each task :

[0077]

[0078] This multi-gated sparsity mechanism allows wind speed and temperature tasks to freely assemble feature routes according to their respective physical needs, effectively mitigating negative transfer interference in multi-task learning.

[0079] (4) Task fusion features after decoupling (Corresponding temperature task) and (For the wind speed task), input their respective dedicated tower networks. The tower networks also employ a three-layer perceptron architecture, ultimately mapping hidden layer features to the target prediction step size:

[0080]

[0081] Finally, the dedicated tower network for each task refines the expert fusion features and jointly outputs the temperature and wind speed forecasts for the next 72 hours, denoted as... ,in , and These correspond to the predicted values ​​of the temperature series and the predicted values ​​of the wind speed series, respectively.

[0082] Step 105: Optimization of the time-frequency domain difference loss function.

[0083] During model training, a time-frequency domain difference loss function is used for multi-objective joint optimization, which integrates three complementary optimization signals. For the two prediction tasks of temperature and wind speed, the loss function is applied to their respective predicted values. Corresponding true value , The losses are calculated independently and then weighted and summed to obtain the total loss. The following will use the same method for each loss. and This represents the predicted value and the actual value for any prediction task:

[0084] (1) Time-domain smoothing L1 loss Calculate the predicted value Compared with the true value The smoothed L1 loss in the time domain is equivalent to the mean squared error loss (quadratic penalty) when the error is small, and degenerates into the mean absolute error loss (linear penalty) when the error is large. This effectively suppresses the excessive interference of outlier observations on model training while ensuring fine fitting for small errors.

[0085]

[0086] Where H is the predicted sequence length. The smooth threshold hyperparameter (usually set to 1.0) controls the transition boundary between the quadratic penalty and the linear penalty.

[0087] (2) Frequency domain loss By mapping the predicted and true values ​​to the frequency domain using Fast Fourier Transform, and calculating the mean absolute error represented by the frequency domain amplitude, the model can simultaneously learn the periodic characteristics and frequency domain patterns of the meteorological sequence, effectively suppressing overfitting due to time domain noise.

[0088]

[0089] Where L is the sequence length and K is the number of low-frequency components involved in the calculation. This represents the amplitude of the k-th frequency component in the Fourier transform result.

[0090] (3) Multi-step difference loss: For time step intervals j=1, 2, ..., N, calculate the j-step difference sequence between the predicted value and the true value, and calculate the time domain smoothing L1 loss for each difference sequence. With frequency domain absolute error loss The differential operation explicit constraint model effectively suppresses the cumulative error of long-term forecasts by enhancing its ability to predict the rate and trend of change of meteorological elements.

[0091] The final total loss function is defined as:

[0092]

[0093] Here, β is a balancing hyperparameter (usually set to 0.1~0.5), which controls the overall weight of the difference loss. The AdamW optimizer is used to minimize the total loss, the learning rate is dynamically adjusted using a cosine annealing strategy, and five-fold cross-validation is combined to improve the model's generalization ability.

[0094] 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 multivariate time-series joint forecasting method for meteorological big data, characterized in that, include: S1: Obtain historical observation data from automatic weather stations and 3×3 grid covariates of ERA5 aligned with them in time and space to form historical multivariate features; at the same time, extract solar thermal radiation and upper atmospheric dynamic parameters that are strictly aligned with the target prediction time window from NWP data as future realistic constraint fields; perform time-series alignment and reversible instance normalization preprocessing on the above data. The preprocessed historical multivariate features are then input into the prediction model. The prediction model includes: a multivariate encoder based on an inverted variable embedding strategy, a deep feature refinement module, and a multi-task decoder based on sparse hybrid expert routing; S2: A multivariate encoder based on an inverted variable embedding strategy is used to encode the preprocessed historical multivariate features. The complete historical sequence of each meteorological element is embedded as an independent variable token. The correlation weight between tokens is calculated in the variable dimension through a multi-head self-attention mechanism to capture the physical dependency between multiple meteorological elements and obtain a multivariate global context representation. S3: The multivariate global context representation is input into the deep feature refinement module. The deep feature refinement module first adopts the star aggregation-redistribution module, which aggregates all variable sequence information into a global core representation through random pooling operation. Then, the global core representation is concatenated and fused with the variable sequence representations to achieve centralized information interaction between variables with linear complexity. Subsequently, combined with the temporal convolutional network, the local temporal dynamic features of meteorological elements are extracted based on the established global variable correlation through causal convolution and dilated convolution to obtain the refined high-dimensional dynamic representation. S4: A multi-task decoder based on sparse hybrid expert routing is used to process the refined high-dimensional dynamic representation. The gating network of the multi-task decoder performs cross-modal fusion of the refined high-dimensional dynamic representation with the future reality constraint field, and dynamically allocates expert routing weights based on the model's self-learning features and the prior knowledge of the future physical environment. The sparse hybrid expert routing includes multiple shared expert networks. The gated network, combined with a load balancing strategy without auxiliary loss, allocates dedicated expert networks for temperature and wind speed prediction tasks to decouple and decode them, and jointly outputs the temperature and wind speed prediction results for the next 72 hours. S6: During model training, the prediction model is optimized using a time-frequency domain difference loss function, which integrates time-domain smoothing L1 loss, frequency domain loss, and multi-step difference loss.

2. The multivariate time-series joint forecasting method for meteorological big data according to claim 1, characterized in that, The preprocessed historical multivariate features are encoded using an inverted variable embedding strategy. The complete historical sequences of each meteorological element are embedded as independent variable tokens. A multi-head self-attention mechanism is used to calculate the correlation weights between tokens along the variable dimension, capturing the physical dependencies between multiple meteorological elements and obtaining a multivariate global context representation, including: S21: Transpose the input historical multivariate feature matrix along the time axis and the variable axis so that each variable sequence is arranged independently in the feature dimension; S22: Map each variable sequence to a unified hidden layer dimension D using a multilayer perceptron to obtain a set of variable tokens; S23: Apply layer normalization to the token dimension of the variable to eliminate the differences in the dimensions of different meteorological elements and improve training stability; S24: Global correlation modeling of the variable token set is performed on the variable dimension through a multi-head self-attention mechanism to obtain the attention score matrix, which reflects the physical coupling relationship between multiple meteorological elements; S25: The temporal representation of each variable token is nonlinearly refined through a shared feedforward network to obtain a multivariate global context representation. The feedforward network acts independently and with shared weights on each variable token to learn the intrinsic temporal rules that can be transferred across variables.

3. The multivariate time-series joint forecasting method for meteorological big data according to claim 1, characterized in that, The information interaction and temporal feature extraction process of the deep feature refinement module includes: S31: The global context representation of each variable is mapped to the dimension using a multilayer perceptron to obtain the core hidden layer representation; S32: During the training phase, the core hidden layer representations of each variable are sampled and aggregated through random pooling to generate a global core representation; during the inference phase, the global core representation is generated through a probability-weighted average method, wherein the random pooling is achieved by calculating the softmax normalized probability on the channel dimension and performing random sampling according to the probability. S33: Broadcast and copy the global core representation and then concatenate it with the global context representations of each variable along the feature dimension to form a fused representation; S34: The fusion representation is reduced in dimension and refined in feature by multilayer perceptron, and residual connections are applied to complete the indirect information interaction between variables and obtain interactive feature representation; S35: The interaction feature representation is input into a temporal convolutional network, which is composed of multiple residual temporal modules stacked together. Each residual temporal module contains two causal convolutional layers. The causal convolutional layers strictly ensure temporal causality by truncating and padding the convolutional output to prevent future information leakage. Each causal convolutional layer is followed by a ReLU activation function and a Dropout regularization layer in sequence, and a weight normalization strategy is used to constrain the parameters of the convolutional kernel. S36: The convolution kernel hole rate of each residual time module is set to increase exponentially by 2, so that the network can achieve an exponentially growing effective receptive field with fewer layers and efficiently cover long-range historical time-series dependencies. S37: A residual connection is applied between the input and output of each residual time module. When the number of input and output channels is inconsistent, a linear mapping is used for dimension alignment. The residuals are added and then processed by the ReLU activation function to finally output the refined high-dimensional dynamic representation.

4. The multivariate time-series joint forecasting method for meteorological big data according to claim 1, characterized in that, The refined high-dimensional dynamic representation is processed using a multi-task decoder based on sparse hybrid expert routing, including: S41: The refined high-dimensional dynamic representation and the future reality constraint field are cross-modal spliced ​​and fused in the feature dimension to obtain multimodal gating features; the initial routing scores of each expert are calculated by the gating linear layer in the gating network on the multimodal gating features, so that the allocation of expert routing weights is based on the historical time-series evolution law and is constrained by the prior knowledge of the future physical environment. The initial route score is activated independently using the sigmoid function, making the scores of each expert independent of each other; S42: Apply learnable expert bias parameters to the routing score. These expert bias parameters are dynamically adjusted during training based on the load of each expert, achieving expert load balancing without the need for an additional auxiliary loss function. S43: Based on the adjusted routing score, activate the corresponding K expert networks for each input using the Top-K selection strategy. Each expert network consists of a SwiGLU gated feedforward structure composed of a gated linear layer, an up-projection linear layer, and a down-projection linear layer. S44: The temperature prediction task and the wind speed prediction task are each equipped with an independent gating network. The outputs of the activated experts are weighted and summed through gating weights, and then refined by the dedicated tower network of each task before outputting the temperature and wind speed prediction results for the next 72 hours.

5. The multivariate time-series joint forecasting method for meteorological big data according to claim 1, characterized in that, The prediction model is optimized using a time-frequency domain difference loss function, including: S51: For the two prediction tasks of temperature and wind speed, calculate the smoothed L1 loss of the predicted and actual values ​​in the time domain respectively. S52: Transform the predicted and actual values ​​into the frequency domain using Fast Fourier Transform, and calculate the absolute error loss of the frequency domain amplitude for temperature and wind speed prediction tasks respectively. S53: For multiple time step intervals, calculate the first-order difference sequence between the predicted value and the true value, and calculate the time domain smoothing L1 loss and frequency domain absolute error loss for the difference sequence under each interval. S54: The time domain loss, frequency domain loss, and step size difference loss are weighted and summed according to preset weights to obtain the total loss, and the prediction model parameters are iteratively optimized based on the total loss through backpropagation.

Images