Multi-element meteorological element data prediction method based on deep neural network fine tuning training

Abstract

A multivariate meteorological element data prediction method based on deep neural network fine tuning training comprises the following steps: preprocessing a multivariate meteorological data set to obtain a preliminary prediction result, and segmenting the preliminary prediction result according to meteorological elements or time steps; respectively matching each time sequence with a universal lightweight fine tuning module for fine tuning, and fully extracting features through a multi-layer perceptron composed of a plurality of linear layers; multiplying the output of the linear layer by a weight V obtained by layer normalization to obtain a prediction result corresponding to the lightweight fine tuning module; and finally, combining the results of the lightweight fine tuning modules. By introducing an advanced time sequence analysis technology and a dynamic adjustment mechanism, the performance of the model can be monitored in real time, and the model can be automatically adjusted when necessary, so that accumulation of prediction errors is reduced, long-term stability and reliability of the model are kept, and the method is suitable for any meteorological element prediction expressed by a time sequence and has a wide application prospect. And the user can select a proper trained deep learning model by himself / herself.

Description

Technical Field

[0001] The present invention relates to a multivariate meteorological element data prediction method based on deep neural network fine-tuning training, and belongs to the technical field of meteorological forecasting and spatiotemporal big data mining. Background Art

[0002] In the field of meteorology, time series forecasting technology plays a vital role. Faced with the uncertainty brought about by global climate change, the need to accurately predict meteorological changes has become particularly urgent, which is crucial in many fields such as agricultural production, disaster prevention, and resource planning. Through in-depth analysis and research of historical meteorological data, time series models that predict future weather changes can be constructed. These models are key tools for understanding and responding to climate change.

[0003] At present, deep learning models have made significant progress in the field of meteorological spatiotemporal forecasting. These models can predict future trends and patterns by analyzing historical data. They use complex algorithms and a large amount of computing resources to conduct in-depth analysis of time series data and spatial distribution data, showing good forecasting results. However, although these models have strong predictive capabilities in theory, their performance is largely limited by the quality and quantity of available data. The limited nature of data is the main bottleneck encountered by deep learning models when processing spatiotemporal data. The incompleteness of data may cause the model to fail to capture all important features and patterns, thereby affecting the accuracy of the forecast. The sparsity or uneven distribution of data may make the model insufficient for predicting certain meteorological elements or time periods. In addition, spatiotemporal data are often accompanied by high dimensionality and nonlinearity, which further increases the difficulty of model training.

[0004] As model performance approaches saturation, researchers and engineers begin to seek new breakthroughs and try to expand the training set through data enhancement technology, or use methods such as transfer learning to improve the generalization ability of meteorological models. At the same time, some researchers are committed to developing more efficient algorithms and model architectures to achieve better prediction results under limited data conditions. Due to the limited data, it is often impossible to deploy large-scale high-performance models, which limits the learning and generalization capabilities of the models. Summary of the invention

[0005] The purpose of the present invention is to provide a meteorological element prediction method based on deep neural network fine-tuning training to enhance the prediction performance of existing deep learning models.

[0006] The method for predicting multivariate meteorological element data based on fine-tuning training of a deep neural network is characterized by comprising the following steps: 1) Preprocessing of multivariate meteorological dataset In the multivariate meteorological dataset X∈ R T×K×N middle, T is the time step, K is the spatial range, N is the number of meteorological elements, data X ∈ R T×K×N is the spatial range K The time step in is T And contains N The data of meteorological elements are based on their predicted spatial range. K medium future S time steps and contains N Forecast data of meteorological elements Ῡ ∈ R S×K×N ; Preprocessing includes removing outliers and missing values ​​in the sequence data and normalizing the data; 2) Obtain preliminary prediction results For historical data X= { x 1 ,...,x T}∈ R T×K×N , x i ∈ R K×N , i =1,…, T , select any trained deep learning model for preliminary training to obtain preliminary prediction results Ŷ ={ y 1 ,...,y S}∈ R S×K×N , y j ∈ R K×N , j =1,…, S ; 3) Segmentation preliminary prediction results The preliminary forecast results are divided into N time series, or split into S Time series; 4) Fine-tune each time series Match each time series separately N or S A universal lightweight fine-tuning module for fine-tuning; each lightweight fine-tuning module is equipped with Adam (Adaptive Moment Estimation) optimizer; The training process of each lightweight fine-tuning module is as follows: S1: First, normalize the mean and variance of each time series to obtain the normalized series Ŷ var_norm ; S2: Use LayerNorm to perform layer normalization on each time series to obtain the weight V; S3: The normalized sequence obtained from S1 Ŷ var_norm The data is fed into a multi-layer perceptron consisting of 4 linear layers to change the feature dimension of the data, so that the number of channels of the time series changes from the initial 1 channel to h1 channels, then to h2 channels, and then back to h1 channels to fully extract the features. Finally, the sequence is compressed back to a 1-dimensional channel, completing the complete cycle of channel change and obtaining the linear layer result. Ŷ var_hidden ; After each linear layer, the GELU (Gaussian Error Linear Unit) activation function is introduced; S4: Finally, the output of the linear layer Ŷ var_hidden Multiply it by the weight V obtained by layer normalization to get the prediction result corresponding to the lightweight fine-tuning module; 5) Generate prediction results The results of each lightweight fine-tuning module are combined to obtain the prediction results of multivariate meteorological element data Ῡ ∈ R S×K×N .

[0007] The meteorological element prediction method based on deep neural network fine-tuning training is characterized in that the deep learning model in step 2) includes but is not limited to Transformer Model and decomposition for long-term time series forecasting Autoformer Model, with interpretability and efficient prediction DLinear The MICN model uses a multi-layer convolutional network to capture data features at different time scales, and the encoder-decoder structure Unet Model.

[0008] The meteorological element prediction method based on deep neural network fine-tuning training is characterized in that in the step 2), the model configuration of its deep learning model adopts its original settings, including the setting of model initialization parameters, the definition of loss function and the optimization of model reasoning process.

[0009] The meteorological element prediction method based on deep neural network fine-tuning training is characterized by the mean variance normalization described in step 4) S1, which is specifically as follows: calculate the mean of all decomposed data, calculate the variance of the decomposed data, and for each data point in each time series, subtract the mean and divide it by the standard deviation. After this processing, the data of each feature will be converted to a distribution with a mean of 0 and a standard deviation of 1, and a normalized sequence will be obtained. Ŷ vari_norm Through mean-variance normalization, not only can the influence of different dimensions and magnitudes be eliminated, but also the convergence speed of the algorithm and the generalization ability of the model can be improved.

[0010] The meteorological element prediction method based on deep neural network fine-tuning training is characterized in that S1-S4 in the lightweight fine-tuning module training process all include the following steps: Forward propagation: calculate the model's predicted value; Error calculation: The prediction performance is evaluated by MSE; Back propagation: calculate gradient information based on MSE; Parameter Update: Update the weights and biases of the multilayer perceptron to reduce the prediction error and improve the predictive power of the model.

[0011] In the current field of deep learning, the performance of multivariate spatiotemporal meteorological prediction models mostly depends on large models. However, in practical applications, due to the limitations of hardware equipment, users often cannot choose the latest large models that require higher computing resources. Even so, users' demand for improving prediction performance has never decreased. In order to maximize the use of existing resources and models, it is particularly important to develop a meteorological element prediction method based on deep neural network fine-tuning training. In response to this challenge, the present invention proposes an innovative solution, which aims to meet users' continuous pursuit of performance improvement by providing a meteorological element prediction method based on deep neural network fine-tuning training. Such a method can select a suitable trained deep learning model according to the performance of the user's device and obtain better prediction performance. And it can maximize the use of existing models and resources to achieve more accurate predictions. Obviously, the present invention can predict any meteorological element that can be expressed by a time series.

[0012] The core advantage of the present invention lies in its flexibility and adaptability. First, users can independently select a suitable basic model according to their needs and resource conditions. This means that whether it is a resource-constrained environment or a scenario that pursues high performance, users can find a suitable starting point. Secondly, the present invention allows users to freely design fine-tuning modules, which provides users with great creative space. Users can customize the model parameters according to their professional knowledge and understanding of specific tasks to obtain the best prediction effect.

[0013] In addition, another significant advantage of the present invention is that it can significantly improve the performance of existing deep learning models. Through a carefully designed fine-tuning strategy, the present invention can tap into the potential capabilities of the basic model, so that the performance of the model on specific tasks is further optimized.

[0014] More importantly, the present invention can effectively improve the effect of prediction error increasing over time. In weather change prediction, prediction errors tend to accumulate over time, resulting in a decrease in the long-term prediction ability of the model. By introducing advanced time series analysis technology and dynamic adjustment mechanisms, the present invention can monitor model performance in real time and automatically adjust when necessary, thereby reducing the accumulation of prediction errors and maintaining the long-term stability and reliability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 Overall flow chart of the present invention.

[0016] Figure 2 Schematic diagram of the fine-tuning framework for spatiotemporal meteorological forecasting.

[0017] Figure 3 The average MSE results of the five meteorological elements of the embodiment.

[0018] Figure 4 The average MSE results of the five meteorological elements at 20 time steps in the embodiment.

[0019] Figure 5 The spatiotemporal comparison diagram of the predicted two-meter ground temperature T2M before and after fine-tuning in the first time step of the embodiment, where (a) is the result of the original deep learning model, (b) is the result of fine-tuning after freezing the original deep learning model, and (c) is the true value.

[0020] Figure 6 The spatiotemporal comparison diagram of the predicted two-meter ground temperature T2M before and after fine-tuning at the twentieth time step of the embodiment, where (a) is the result of the original deep learning model, (b) is the result of fine-tuning after freezing the original deep learning model, and (c) is the true value.

[0021] Figure 7 The spatiotemporal comparison diagram of the predicted ten-meter latitudinal wind speed V10 in the embodiment before and after fine-tuning at the first time step, where (a) is the result of the original deep learning model, (b) is the result of fine-tuning after freezing the original deep learning model, and (c) is the true value.

[0022] Figure 8The spatiotemporal comparison diagram of the predicted ten-meter latitudinal wind speed V10 before and after fine-tuning at the twentieth time step of the embodiment, wherein (a) is the result of the original deep learning model, (b) is the result of fine-tuning after freezing the original deep learning model, and (c) is the true value. DETAILED DESCRIPTION

[0023] like Figure 1 As shown, the meteorological element prediction method based on deep neural network fine-tuning training includes the following steps: 1) Preprocessing of multivariate meteorological dataset In the multivariate meteorological dataset X ∈ R T×K×N The elements can be represented as multivariate meteorological data presented in time series, each time series includes N Meteorological elements, spatial range K , the time step is T ; Preprocessing includes removing outliers and missing values ​​in time series data, ensuring the integrity and accuracy of the data, and normalizing the data.

[0024] 2) Obtain preliminary test results For historical data X ∈ R T×K×N , select any trained deep learning model for preliminary training to obtain preliminary prediction results Ŷ ={ y 1 ,...,y S}∈ R S×K×N .

[0025] Here, you can choose different deep learning models according to your needs, which include different design structures, such as Autoformer A decomposition for long-term time series forecasting Transformer The model gradually separates the long-term trend information from the predicted hidden variables by embedding decomposition blocks and autocorrelation mechanisms. It introduces an autocorrelation mechanism to replace the self-attention mechanism, discovers the similarity of subsequences based on the periodicity of the sequence, and aggregates similar subsequences from the underlying period, achieving O ( L log L ). Or DLinear Achieve high-efficiency prediction, since each branch has only one linear layer, it is better than the existing Transformer Consumes less memory and fewer parameters, has faster inference speed, and DLinear With interpretability, the weights of seasonality and trend branches can be visualized after training, thus enhancing insights into the forecasted values. MICN,The multi-layer convolutional network captures data features at different time scales and effectively extracts complex time series information.,The model integrates and interacts information at different scales through the interactive fusion module,,enhancing the understanding of the dependencies between different time periods. MICN While improving accuracy, it also keeps computational overhead and parameter count low. You can also choose Unet Model. Users can select the above or other suitable trained deep learning models as the basic model according to specific device needs.

[0026] In the initial training, the model configuration of its deep learning model can use its original settings, including the setting of model initialization parameters, the definition of loss function and the optimization of model inference process. These factors jointly determine the performance of the basic model and the accuracy of the first stage prediction results.

[0027] 3) Preliminary prediction result segmentation Two segmentation schemes are designed: Variable-by-variable fine-tuning scheme The initial prediction results are divided into N Each time series focuses on capturing and predicting a specific meteorological element in the future. S This segmentation enables the model to analyze and predict the dynamic changes of each meteorological element in more detail, thereby improving the accuracy and details of the forecast.

[0028] The split N The time series correspond to N Fine-tuning modules. Each fine-tuning module is responsible for processing the time series of the corresponding meteorological elements and further fine-tuning them. This modular design allows the model to be optimized specifically for the characteristics and change patterns of different meteorological elements, thereby improving the overall forecast performance.

[0029] Each trim module is equipped with Adam (Adaptive Moment Estimation) optimizer. Adam is an adaptive learning rate optimization algorithm that combines RMSprop and Momentum Advantages of the two optimization algorithms. Adam The optimizer can automatically adjust the learning rate to adapt the optimization path of each parameter based on the first-order moment estimate and second-order moment estimate of the gradient. This adaptive feature makes Adam It excels in processing time series data, especially in the face of non-stationarity and noise.

[0030] Time-step fine-tuning scheme The preliminary forecast results are broken down intoS Similar to the fine-tuning scheme for each meteorological element, each time series in the fine-tuning scheme for each time step focuses on representing the N The characteristic changes of each meteorological element. S The time series correspond to S Each fine-tuning module is equipped with Adam Optimizer.

[0031] 4) Fine-tuning training Match a common lightweight fine-tuning module to each time series. Take variable-by-variable fine-tuning as an example: S1: First, perform mean-variance normalization on it and calculate the mean of the meteorological element. The mean is the sum of all future sequence data points divided by the number of data points, which represents the central trend of the meteorological element in the future. Next, calculate the variance of the meteorological element. Variance is an indicator of the degree of data dispersion, which describes the difference between the data point and the mean. Then, for each data point in the time series, subtract the mean of the feature and divide it by the standard deviation of the feature. After this processing, the data of each feature will be converted to a distribution with a mean of 0 and a standard deviation of 1 to obtain a normalized feature. Mean-variance normalization can not only eliminate the impact of different dimensions and magnitudes, but also improve the convergence speed of the algorithm and the generalization ability of the model.

[0032] S2: Use for each time series simultaneously LayerNorm Perform layer normalization to obtain weights V .

[0033] S3: The normalized sequence obtained after S1 is fed into a linear layer, which changes the feature dimension of the data. Specifically, the number of channels of the sequence is increased from the initial 1 channel starts and changes to h1 This step greatly enriches the representation ability of the data. Subsequently, in order to further deepen the abstract level of features, the number of channels is changed again, this time to h2 channels, allowing the model to capture more subtle and complex features. After sufficient feature extraction, the number of channels returns to h1 Finally, the sequence is compressed back to 1 dimensional channel, completing the complete cycle of channel change and obtaining the linear layer result Ŷ var_hidden .

[0034] After each linear layer, we introduce GELU (Gaussian Error Linear Unit) activation function, which can dynamically adjust the degree of activation based on the statistical characteristics of the input. GELUThe introduction of activation functions allows the model to introduce nonlinearity after each linear transformation, which helps the model learn more complex data patterns. In this way, the results of the hidden layer not only contain rich feature information, but also have stronger nonlinear expression capabilities, providing a more powerful feature foundation for subsequent tasks.

[0035] S4: Finally, the output of the linear layer is multiplied by the weight V obtained by layer normalization to optimize the performance of the model and improve the accuracy of the prediction; the prediction result corresponding to the lightweight fine-tuning module is obtained; The output of the linear layer and the weights obtained by layer normalization V When multiplying, it is actually re-weighting the features. This weighting takes into account the distribution of each time step in different samples, allowing the model to pay more attention to those time steps that have a greater impact on the prediction results, while suppressing those features that may bring noise.

[0036] 5) Generate prediction results The results of each lightweight fine-tuning module are combined. Finally, through the fine adjustment of this feature, the prediction results of multivariate meteorological element data are obtained. Ῡ ∈ R S×K×N .

[0037] After being processed by the fine-tuning module, the forecast results of each time series will be combined to form the final forecast output. This output not only contains a detailed forecast of the future changes of each meteorological element, but also takes into account the mutual influence and overall trend between meteorological elements, providing strong data support for decision-making.

[0038] Based on the preliminary prediction results, the present invention uses the multi-layer perceptron as the core to construct a fine-tuning module. In each iteration, the model first performs forward propagation to calculate the predicted value, and then calculates the MSE to evaluate the prediction performance. Then, the back-propagation algorithm will update the weights and biases in the multi-layer perceptron based on the gradient information calculated by the MSE to reduce the prediction error and improve the prediction ability of the model.

[0039] The fine-tuning module of the present invention is continuously iterated during the training process, and each iteration includes forward propagation, error calculation, back propagation and parameter update until the module reaches a predetermined number of early stops. Finally, when all modules stop training, the fine-tuning is completed.

[0040] Whenever prediction is needed, only historical data X={x1,...,x T}∈ R T×K×N First, we get the preliminary prediction results, and then fine-tune the framework variable by variable or time step by time step to get the final prediction results. Ῡ= { y1,...,y S}∈ R S×K×N .

[0041] Example Below is the overall flow chart of the present invention, with the European Center for Medium-Range Weather Forecasts (ECMWF) Reanalysis data ERA5 The technical solution of the present invention is described clearly and completely as an embodiment. Obviously, the described embodiment is only a part of the embodiments of the present invention, not all of the embodiments, to show that the framework can be applied to spatiotemporal data in addition to time series data. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technical users in this field without creative work are within the scope of protection of the present invention.

[0042] The data used in this example comes from the European Centre for Medium-Range Weather Forecasts (ECMWF) Reanalysis data ERA5 Subset, including 5 ground meteorological elements: ground two meters low temperature T2M , 10m latitudinal wind speed U10 , 10m longitudinal wind speed V10 , mean sea level pressure MSL , 6-hour cumulative precipitation TP . The time interval is 6 hours, the spatial resolution is 0.25°, and the coverage area (N10 ~ N50, E100 ~ E140) This area covers important sea areas along the eastern coast of China, including the Yellow Sea, the East China Sea and part of the South China Sea. It has important research and application value for marine resource development, environmental protection and disaster warning. Training set time range January 1, 2007 ~ December 31, 2015 This embodiment mainly includes the following 5 steps, and the step diagram is shown in Figure 1 .

[0043] 1) ERA5 reanalysis data preprocessing The ERA5 reanalysis data have a sampling time d ,longitude lon ,latitude lat The grid data.

[0044] The research focus of this embodiment is on the prediction of five meteorological elements on the ocean surface in eastern China, specifically in the following areas: (N10 ~ N50, E100 ~ E140), First, the spatial resolution of this embodiment is 0.25 Based on this resolution, the study area is divided into 160×160 A spatial grid with a time interval of 6 hours.

[0045] The core of the prediction task is to input 12 hours of historical data, which includes the data grid space of the previous two time points, specifically 2 time points (T =2), 160 spatial grid width, 160 spatial grid height ( K =160×160), select T2M , U10 , V10 , MSL , TP As a meteorological element ( N =5). Predict environmental changes in the next five days, that is, data at 20 time points (20 time points, 5 environmental parameters, 160 spatial grid widths, 160 spatial grid heights).

[0046] Preprocessing includes abnormal data filtering and interpolation: For X={x 1, x2}∈ R 2×(160×160)×5 Replace the outliers with the mean value. The data were standardized using mean and variance.

[0047] 2) Preliminary prediction results obtained In model selection, we used Unet Model, in this embodiment, Unet The number of input channels of the model is set to 10, which corresponds to 5 ground meteorological elements at 2 historical moments. The number of prediction channels is 100, which corresponds to 5 ground meteorological elements at 20 historical moments. Unet Get the preliminary prediction results Ῡ= { y 1,...,y 20}∈ R 20×(160×160)×5 .

[0048] After the training is completed, follow the normal operation to convert the original deep learning model Unet The weights are frozen to allow for subsequent fine-tuning.

[0049] 3) Segmentation preliminary prediction results In this example, fine-tuning is performed for the next 20 time steps, that is, 20 fine-tuning modules are set, and the input data dimension of each fine-tuning module is （5，160，160） , expressed as the values ​​of five ground meteorological elements over the entire prediction area at a single time step.

[0050] The preliminary forecast results, i.e. the forecast output of the model for multiple meteorological elements at a series of time steps, are decomposed into 20 time series, namely Y stepi ∈ R (160×160)×5 Each time series focuses on representing the prediction range at a specific time step. 5The 20 time-space sequences decomposed correspond to 20 fine-tuning modules. Each fine-tuning module is equipped with Adam Optimizer. After processing by the fine-tuning module, the prediction result Y of each spatiotemporal series stepi ∈ R (160×160)×5 Will be combined to form the final prediction output Ῡ = { y 1,...,y 20 ∈ R 20×(160×160)×5 .

[0051] 4) Fine-tuning training The input data of each fine-tuning module is Y stepi ∈ R (160×160)×5 , the process is as follows: S1: first perform the result Y stepi ∈ R (160×160)×5 Perform mean-variance normalization to calculate the mean of the meteorological elements. Then calculate the variance, and then for each data point in the time series, subtract the mean of the feature and divide it by the standard deviation of the feature. After this processing, the data of each feature will be converted to a distribution with a mean of 0 and a standard deviation of 1, and a normalized feature will be obtained. Ŷ stepi_norm .

[0052] S2: Sequence given at the same time Ŷ stepi use LayerNorm Perform layer normalization to obtain weights V .

[0053] S3: In the process of processing the normalized sequence, first Ŷ stepi_norm The sequence is first flattened, and the number of channels is reduced from the initial 5×160 ×160 channels, the first linear layer directly reduces the number of channels to 256 , then, in order to further deepen the level of feature abstraction, the number of channels is changed again, this time reduced to 128 channels, and after sufficient feature extraction, the number of channels increases to 256 Finally, the sequence is expanded to 5×160×160 dimensional channel, and get the linear layer result Ŷ stepi_hidden ; After each linear layer, we introduce GELU (Gaussian Error Linear Unit) activation function.

[0054] S4: Finally, the output of the linear layer Ŷ Vari_hidden The weights obtained by layer normalization V Multiply.

[0055] The output of the linear layer and the weights obtained by layer normalization V When multiplying, it is actually re-weighting the features. This weighting takes into account the distribution of each time step in different samples, allowing the model to pay more attention to those time steps that have a greater impact on the prediction results, while suppressing those features that may bring noise.

[0056] 5) The final prediction result is obtained The results of each lightweight fine-tuning module are combined, and finally the prediction results of multivariate meteorological element data are obtained through fine-tuning of this feature. Ῡ= { y 1,...,y 20}∈ R 20×(160×160)×5 ,like Figure 2 shown.

[0057] The preliminary prediction results Ῡ ,The present invention takes the multi-layer perceptron as the core to construct a fine-tuning module; In each iteration, the model first performs forward propagation to calculate the predicted value, and then calculates the MSE to evaluate the prediction performance. Then, the backpropagation algorithm will update the weights and biases in the fine-tuning module based on the gradient information calculated by the MSE to reduce the prediction error and improve the prediction ability of the model; The fine-tuning module of the present invention is continuously iterated during the training process, and each iteration includes forward propagation, error calculation, back propagation and parameter update until the corresponding module reaches a predetermined number of early stops at a certain time step. Finally, when all modules stop training, the model training ends.

[0058] This example was tested on a total of 1443 test samples in 2016. The average MSE results of the five meteorological elements are shown in Figure 3 , the horizontal axis is five ground meteorological elements, and the vertical axis is MSE, which can significantly reduce the MSE of the five ground meteorological elements; the average MSE results of the five meteorological elements under 20 time steps can be found in Figure 4 , the horizontal axis is 20 time steps, and the vertical axis is MSE, which can significantly reduce the overall prediction results; see the fine-tuning spatiotemporal results of T2M and V10. Figure 5-Figure 8, where (a) is the prediction result of the original deep learning model Unet, (b) is the result of fine-tuning after freezing the original deep learning model, and (c) is the true value. It can be seen that the prediction result after fine-tuning is significantly close to the true situation, especially the method of the present invention. Figure 5-Figure 8 The improvement effect is more obvious in the box area.

Claims

1. A multivariate meteorological element data prediction method based on deep neural network fine-tuning training, characterized by The following steps are involved: 1) Preprocessing of multivariate meteorological dataset In the multivariate meteorological dataset R T×K×N middle, T is the time step, K is the spatial range, N is the number of meteorological elements, data X ∈ R T ×K×N is the spatial range K The time step in is T And contains N The data of meteorological elements are based on their predicted spatial range. K medium future S time steps and contains N Forecast data of meteorological elements Ῡ ∈ R S×K×N ; Preprocessing includes removing outliers and missing values ​​in the sequence data and normalizing the data; 2) Obtain preliminary prediction results For historical data X= { x 1 ,...,x T }∈ R T×K×N , select any trained deep learning model for preliminary training to obtain preliminary prediction results Ŷ ={ y 1 ,...,y S }∈ R S×K×N ; 3) Segmentation preliminary prediction results The preliminary forecast results are divided into N time series, or split into S Time series; 4) Fine-tune each time series Match each time series separately N or S A universal lightweight fine-tuning module for fine-tuning; each lightweight fine-tuning module is equipped with Adam Optimizer; The training process of each lightweight fine-tuning module is as follows: S1: First, normalize the mean and variance of each time series to obtain the normalized series Ŷ var_norm ; S2: Use LayerNorm to perform layer normalization on each time series to obtain the weight V; S3: The normalized sequence obtained from S1 Ŷ var_norm The data is fed into a multi-layer perceptron consisting of 4 linear layers to change the feature dimension of the data, so that the number of channels of the time series changes from the initial 1 channel to h1 channels, then to h2 channels, and then back to h1 channels to fully extract the features. Finally, the sequence is compressed back to a 1-dimensional channel, completing the complete cycle of channel change and obtaining the linear layer result. Ŷ var_hidden ; After each linear layer, the GELU activation function is introduced; S4: Finally, the output of the linear layer Ŷ var_hidden Multiply it by the weight V obtained by layer normalization to get the prediction result corresponding to the lightweight fine-tuning module; 5) Generate prediction results The results of each lightweight fine-tuning module are combined to obtain the prediction results of multivariate meteorological element data Ῡ ∈ R S ×K×N .

2. The multivariate meteorological element data prediction method based on deep neural network fine-tuning training as claimed in claim 1 is characterized by The deep learning model described in step 2) includes but is not limited to Transformer Model and decomposition for long-term time series forecasting Autoformer Model, with interpretability and efficient prediction DLinear Model, capturing data features at different time scales through multi-layer convolutional networks MICN Model, and encoder-decoder structure Unet Model.

3. The multivariate meteorological element data prediction method based on deep neural network fine-tuning training as claimed in claim 1 is characterized by In the step 2), the model configuration of the deep learning model adopts its original settings, including the setting of model initialization parameters, the definition of loss function and the optimization of model inference process.

4. The multivariate meteorological element data prediction method based on deep neural network fine-tuning training as claimed in claim 1 is characterized by The mean-variance normalization described in S1 in step 4) is specifically as follows: calculate the mean of all decomposed data, calculate the variance of the decomposed data, and for each data point in each time series, subtract the mean and divide it by the standard deviation.

5. The multivariate meteorological element data prediction method based on deep neural network fine-tuning training as claimed in claim 1, characterized in that: S1-S4 in the lightweight fine-tuning module training process all include the following steps: Forward propagation: calculate the model's predicted value; Error calculation: The prediction performance is evaluated by MSE; Back propagation: calculate gradient information based on MSE; Parameter Update: Update the weights and biases of the multilayer perceptron to reduce the prediction error and improve the predictive power of the model.

Images