Multivariate Meteorological Element Data Prediction Method Based on Fine-Tuning Training of Deep Neural Network

Abstract

A multi - meteorological element data prediction method based on fine - tuning training of a deep neural network, including pre - processing a multi - meteorological data set to obtain a preliminary prediction result, and splitting the preliminary prediction result by meteorological elements or by time steps; respectively matching each time series with a general lightweight fine - tuning module for fine - tuning, and fully extracting features through a multi - layer perceptron composed of multiple linear layers; multiplying the output of the linear layer by the weight V obtained by layer normalization to obtain the prediction result corresponding to this lightweight fine - tuning module; finally, combining the results of each lightweight fine - tuning module. By introducing advanced time - series analysis techniques and a dynamic adjustment mechanism, the present invention can monitor the model performance in real time and automatically adjust when necessary, thereby reducing the accumulation of prediction errors, maintaining the long - term stability and reliability of the model, being applicable to the prediction of any meteorological element expressed in time series, and allowing users to independently select a suitable pre - trained deep - learning model.

Description

Technical Field

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

[0002] In the field of meteorology, time series prediction technology plays a crucial role. Facing the uncertainties brought about by global climate change, the need to accurately predict meteorological changes has become particularly urgent, which is of great importance in many fields such as agricultural production, disaster prevention, and resource planning. By deeply analyzing and studying historical meteorological data, time series models for predicting future weather changes can be constructed, and these models are key tools for understanding and coping with climate change.

[0003] Currently, deep learning models have made significant progress in the field of meteorological spatio-temporal prediction. These models can predict future trends and patterns by analyzing historical data. They use complex algorithms and a large amount of computing resources to deeply analyze time series data and spatial distribution data, showing good prediction effects. However, although these models have powerful prediction capabilities in theory, their performance is largely limited by the quality and quantity of available data. The finiteness of data is the main bottleneck encountered by deep learning models when processing spatio-temporal data. The incompleteness of data may cause the model to fail to capture all important features and patterns, thus affecting the prediction accuracy. The sparsity or uneven distribution of data may make the model insufficient in predicting certain meteorological elements or time periods. In addition, spatio-temporal data often has the characteristics of high dimension and non-linearity, which further increases the difficulty of model training.

[0004] As the model performance approaches saturation, researchers and engineers have begun to seek new breakthroughs and try to expand the training set through data augmentation techniques, or use methods such as transfer learning to improve the generalization ability of meteorological models. At the same time, some researchers are also committed to developing more efficient algorithms and model architectures in order to achieve better prediction effects under limited data conditions. Due to the finiteness of data, large-scale high-performance models are often unable to be deployed, which limits the learning ability and generalization ability of the models. Summary of the Invention

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

[0006] A prediction method for multivariate meteorological element data based on fine-tuning training of a deep neural network is characterized by including the following steps:

[0007] 1) Preprocessing of multivariate meteorological data sets

[0008] In the multi - variable meteorological data set X ∈ R T×K×N where T is the time step, K is the spatial range, N is the number of meteorological elements, and the data X ∈ R T×K×N is the spatial range K in which the time step is T and contains N meteorological elements. Its predicted spatial range K in the future S time steps and contains N meteorological elements of the predicted data Ῡ ∈ R S×K×N ; The pre - processing includes removing outliers and missing values in the sequence data and normalizing the data;

[0009] 2) Obtain the preliminary prediction results

[0010] For the historical data X= { x 1 ,...,x T}∈ R T×K×N where x i ∈ R K×N and i = 1, …, T select any pre - trained deep learning model for preliminary training to obtain the preliminary prediction results Ŷ ={ y 1 ,...,y S}∈ R S×K×N where y j ∈ R K×N and j = 1, …, S ;

[0011] 3) Split the preliminary prediction results

[0012] Split the preliminary prediction results into N time series according to meteorological elements, or split them into S time series according to time steps;

[0013] 4) Fine - tune and train each time series

[0014] Match each time series separately N or S a general lightweight fine-tuning module for fine-tuning; each lightweight fine-tuning module is equipped with Adam (Adaptive Moment Estimation) optimizer;

[0015] The training process of each lightweight fine-tuning module is as follows:

[0016] S1: First, perform mean-variance normalization on each time series to obtain a normalized series Ŷ var_norm ;

[0017] S2: At the same time, perform layer normalization on each time series using LayerNorm to obtain the weight V;

[0018] S3: Feed the normalized series obtained in S1 Ŷ var_norm into a multi-layer perceptron composed 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 features. Finally, the series is compressed back to a 1-dimensional channel, completing a complete cycle of channel change, and obtaining the result of the linear layer Ŷ var_hidden ;

[0019] After each linear layer, introduce the GELU (Gaussian Error Linear Unit) activation function;

[0020] S4: Finally, multiply the output of the linear layer Ŷ var_hidden by the weight V obtained from layer normalization to obtain the prediction result corresponding to this lightweight fine-tuning module;

[0021] 5) Generate prediction results

[0022] Combine the results of each lightweight fine-tuning module to obtain the prediction result of multivariate meteorological element data Ῡ ∈ R S×K×N .

[0023] The meteorological element prediction method based on deep neural network fine-tuning training is characterized in that the deep learning model described in step 2) includes but is not limited to belonging to Transformer the model and is used for long-term time series prediction decomposition Autoformer the model, and has interpretability and high-efficiency prediction DLinearThe model, the MICN model that captures data features at different time scales through a multi-layer convolutional network, and the Unet model.

[0024] In the meteorological element prediction method based on fine-tuning training of a deep neural network, it is characterized in that in 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 functions, and the optimization of the model inference process.

[0025] In the meteorological element prediction method based on fine-tuning training of a deep neural network, it is characterized in that in step 4), the mean-variance normalization in S1 is specifically as follows: calculate the mean of all decomposed data, calculate the variance of the decomposed data, for each data point in each time series, subtract its mean and then divide by the standard deviation. After such processing, the data of each feature will be transformed into a distribution with a mean of 0 and a standard deviation of 1 to obtain a normalized sequence. Ŷ vari_norm . Through mean-variance normalization, not only can the influence brought by different dimensions and magnitudes be eliminated, but also the convergence speed of the algorithm and the generalization ability of the model can be improved.

[0026] In the meteorological element prediction method based on fine-tuning training of a deep neural network, it is characterized in that S1-S4 in the training process of the lightweight fine-tuning module all include the following steps:

[0027] Forward propagation: Calculate the predicted value of the model;

[0028] Error calculation: Evaluate the prediction performance through MSE;

[0029] Backward propagation: Calculate the gradient information according to MSE;

[0030] Parameter update: Update the weights and biases of the multi-layer perceptron to reduce the prediction error and improve the prediction ability of the model.

[0031] In the current field of deep learning, the performance of multi-variable spatio-temporal meteorological prediction models mostly depends on large models. However, in practical applications, due to hardware device limitations, users often cannot choose the latest large models that require higher computing resources. Even so, users' demand for improving prediction performance has never decreased. To maximize the utilization of existing resources and models, it is particularly important to develop a meteorological element prediction method based on fine-tuning training of deep neural networks. To address this challenge, the present invention proposes an innovative solution, aiming to meet users' continuous pursuit of performance improvement by providing a meteorological element prediction method based on fine-tuning training of deep neural networks. Such a method can select a suitable pre-trained deep learning model according to the performance of the user's device and obtain better prediction performance. And it can maximize the utilization of existing models and resources to achieve more accurate predictions. Obviously, the present invention can predict any meteorological element that can be expressed through time series.

[0032] The core advantages of the present invention lie in its flexibility and adaptability. First of all, users can independently select a suitable base model according to their own needs and resource situations. This means that in both resource-constrained environments and scenarios pursuing high performance, users can find a suitable starting point. Secondly, the present invention allows users to freely design the fine-tuning module, which provides users with great creative space. Users can customize and adjust model parameters according to their professional knowledge and understanding of specific tasks to obtain the best prediction results.

[0033] In addition, another remarkable 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 the potential capabilities of the base model, further optimizing the performance of the model on specific tasks.

[0034] More importantly, the present invention can effectively improve the effect that the prediction error increases with time. In weather change prediction, the prediction error often accumulates over time, leading to a decline in the long-term prediction ability of the model. By introducing advanced time series analysis techniques and dynamic adjustment mechanisms, the present invention can monitor the model performance in real time and automatically make adjustments 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

[0035] Figure 1 The overall flowchart of the present invention.

[0036] Figure 2 Schematic diagram of the spatio-temporal meteorological prediction fine-tuning framework.

[0037] Figure 3 The average MSE results of five meteorological elements in the embodiment.

[0038] Figure 4 Average MSE results of five meteorological elements at 20 time steps of the embodiment.

[0039] Figure 5 Spatio-temporal comparison diagrams of the predicted surface two-meter temperature T2M of 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.

[0040] Figure 6 Spatio-temporal comparison diagrams of the predicted surface two-meter temperature T2M of the embodiment before and after fine-tuning at the twentieth 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.

[0041] Figure 7 Spatio-temporal comparison diagrams of the predicted ten-meter zonal wind speed V10 of 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.

[0042] Figure 8 Spatio-temporal comparison diagrams of the predicted ten-meter zonal wind speed V10 of the embodiment before and after fine-tuning at the twentieth 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. Detailed implementation manners

[0043] As Figure 1 shown, the meteorological element prediction method based on fine-tuning training of a deep neural network includes the following steps:

[0044] 1) Preprocessing of the multi-source meteorological dataset

[0045] In the multi-source meteorological dataset X ∈ R T×K×N its elements can be expressed as multi-source meteorological data presented in a time series, and each time series includes N meteorological elements, with a spatial range K , and a time step of T ; the preprocessing includes removing outliers and missing values in the time series data to ensure the integrity and accuracy of the data, and performing normalization processing on the data.

[0046] 2) Obtaining preliminary prediction results

[0047] 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 .

[0048] Here, different deep learning models with different design structures can be selected according to requirements. For example Autoformer A decomposition model for long-term time series prediction Transformer model that gradually separates long-term trend information from the predicted hidden variables through an embedding decomposition block and an autocorrelation mechanism. 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 cycles, achieving O ( L log L ) computational complexity. Or DLinear achieve high-efficiency prediction. Since each branch has only one linear layer, it consumes less memory and fewer parameters than existing Transformer ones, has a faster inference speed, and DLinear is interpretable. After training, the weights of the seasonal and trend branches can be visualized, thus enhancing the insight into the predicted values. One can also choose MICN , which captures data features at different time scales through a multi-layer convolutional network and effectively extracts complex time series information. This model integrates and interacts information at different scales through an interactive fusion module, enhancing the understanding of the dependencies between different time periods. MICN While improving the accuracy, it maintains a low computational overhead and the number of parameters. One can also choose Unet model. Users can select the above or other suitable trained deep learning models as the base model according to the specific device needs.

[0049] In the preliminary training, the model configuration of the deep learning model can use its original settings, including the settings of model initialization parameters, the definition of the loss function, and the optimization of the model inference process. These factors jointly determine the performance of the base model and the accuracy of the first-stage prediction results.

[0050] 3) Preliminary prediction result segmentation

[0051] Two segmentation schemes are designed:

[0052] Variable-by-variable fine-tuning scheme

[0053] The preliminary prediction results are segmented into NSeparate time series. Each time series focuses on capturing and predicting the characteristic changes of a specific meteorological element in the future S for a certain number of time steps. 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 prediction.

[0054] The segmented N time series are respectively corresponding to N a number of fine-tuning modules. Each fine-tuning module is specifically responsible for processing the time series of the corresponding meteorological element and making further refined adjustments to it. This modular design allows the model to be specifically optimized for the characteristics and change patterns of different meteorological elements, thus improving the overall prediction performance.

[0055] Each fine-tuning module is equipped with Adam the Adam (Adaptive Moment Estimation) optimizer. Adam Adam is an adaptive learning rate optimization algorithm that combines RMSprop and Momentum the advantages of two optimization algorithms. Adam The Adam optimizer can automatically adjust the learning rate and adapt to the optimization path of each parameter according to the first-order moment estimate and second-order moment estimate of the gradient. This adaptive feature makes Adam Adam perform well when dealing with time series data, especially in the face of non-stationarity and noise.

[0056] Time-step by time-step fine-tuning scheme

[0057] The preliminary prediction results are decomposed into S a number of time series. Similar to the per-variable fine-tuning scheme, each time series in the time-step by time-step fine-tuning scheme focuses on representing the characteristic changes of N a number of meteorological elements at a specific time step. The decomposed S time series are respectively corresponding to S a number of fine-tuning modules. Each fine-tuning module is equipped with Adam the Adam optimizer.

[0058] 4) Fine-tuning training

[0059] Match each time series with a general lightweight fine-tuning module. Taking per-variable fine-tuning as an example:

[0060] S1: First, perform mean-variance normalization on it to calculate the mean of the meteorological elements. The mean is the sum of all future sequence data points divided by the number of data points, which represents the central tendency of the future of this meteorological element. Then, calculate the variance of the meteorological elements. Variance is an index to measure the degree of data dispersion, which describes the magnitude of the difference between data points and the mean. Then, for each data point in this time series, subtract the mean of this feature and divide by the standard deviation of this feature. After such processing, the data of each feature will be transformed into a distribution with a mean of 0 and a standard deviation of 1, obtaining the normalized features. Through mean-variance normalization, not only can the influence brought by different dimensions and magnitudes be eliminated, but also the convergence speed of the algorithm and the generalization ability of the model can be improved.

[0061] S2: At the same time, use LayerNorm to perform layer normalization on each time series to obtain the weights V .

[0062] S3: Feed the normalized sequence obtained after S1 into a linear layer, and the role of this layer is to change the feature dimension of the data. Specifically, the number of channels of the sequence starts from the initial 1 channels and changes to h1 channels. This step greatly enriches the representation ability of the data. Subsequently, in order to further deepen the abstraction level of the features, the number of channels changes again. This time it changes to h2 channels, enabling the model to capture more subtle and complex features. After sufficient feature extraction, the number of channels returns to h1 channels. Finally, the sequence is compressed back to 1 dimensional channels, completing a complete cycle of channel changes to obtain the result of the linear layer Ŷ var_hidden .

[0063] After each linear layer, the GELU (Gaussian Error Linear Unit) activation function is introduced, which can dynamically adjust the activation degree according to the statistical characteristics of the input. GELU The introduction of the activation function enables the model to introduce non-linearity 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 non-linear expression ability, providing a more powerful feature basis for subsequent tasks.

[0064] S4: Finally, multiply the output of the linear layer by the weight V obtained by layer normalization to optimize the performance of the model and improve the prediction accuracy; obtain the prediction result corresponding to this lightweight fine-tuning module;

[0065] The output result of the linear layer and the weight obtained by layer normalizationV When multiplying, a re-weighting of features is actually carried out. This weighting takes into account the distribution of each time step among different samples, enabling 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 introduce noise.

[0066] 5) Generate prediction results

[0067] The results of each lightweight fine-tuning module are combined. Finally, through this fine-tuning of features, the prediction results of multivariate meteorological element data are obtained. Ῡ ∈ R S×K×N 。

[0068] After being processed by the fine-tuning module, the prediction results of each time series will be combined to form the final prediction output. This output not only contains a detailed prediction of the future changes of each meteorological element but also takes into account the mutual influence and overall trend among meteorological elements, providing strong data support for decision-making.

[0069] For the obtained preliminary prediction results, the present invention constructs a fine-tuning module with a multi-layer perceptron as the core. In each iteration, the model first performs forward propagation to calculate the predicted values, and then calculates the MSE to evaluate the prediction performance. Next, the backpropagation algorithm will update the weights and biases in the multi-layer perceptron according to the gradient information calculated by the MSE to reduce the prediction error and improve the prediction ability of the model.

[0070] The fine-tuning module of the present invention iterates continuously during the training process. Each iteration includes forward propagation, error calculation, backpropagation, and parameter update until the module reaches a predetermined early stopping number. Finally, when all the modules stop training, the fine-tuning is completed.

[0071] Whenever predicting, only the historical data X = {x 1 ,..., x T} ∈ R T×K×N needs to be given. First, the preliminary prediction results are obtained, and a fine-tuning framework for each variable or each time step is carried out to obtain the final prediction results. Ῡ= { y 1 ,..., y S} ∈ R S×K×N 。

[0072] Embodiment

[0073] Next, in combination with the overall flowchart in the present invention, taking the reanalysis data of the European Centre for Medium-Range Weather Forecasts (ECMWF) as an example ERA5The 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.

[0074] 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 to December 31, 2015 This embodiment mainly includes the following 5 steps, and the step diagram is shown in Figure 1 .

[0075] 1) ERA5 reanalysis data preprocessing

[0076] The ERA5 reanalysis data have a sampling time d ,longitude lon ,latitude lat The grid data.

[0077] 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.

[0078] 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 the environmental changes in the next five days, i.e., data at 20 time points (20 time points, 5 environmental parameters, width of 160 spatial grids, height of 160 spatial grids).

[0079] Preprocessing includes abnormal data filtering and interpolation filling:

[0080] For the data in X = {x 1, x 2} ∈ R 2×(160×160)×5 , replace the outliers with the average value;

[0081] Normalize the data using the mean and variance.

[0082] 2) Obtain the preliminary prediction results

[0083] In model selection, the Unet model is adopted. In this embodiment, Unet the number of input channels of the model is set to 10, which corresponds to 5 surface meteorological elements at 2 historical moments. The number of prediction channels is 100, which corresponds to 5 surface meteorological elements at 20 historical moments. Through Unet obtain the preliminary prediction results Ῡ= { y 1 ,..., y 20} ∈ R 20×(160×160)×5 .

[0084] After training, freeze the weights of the original deep learning model according to the conventional operation Unet for subsequent fine-tuning.

[0085] 3) Split the preliminary prediction results

[0086] In this example, fine-tuning is performed on the next 20 time steps, that is, 20 fine-tuning modules are set. The input data dimension of each fine-tuning module is （5，160，160） , representing the values of the five surface meteorological elements within the entire prediction area at a single time step.

[0087] The preliminary prediction results, that is, the prediction outputs 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 5 characteristic changes of the meteorological elements within the prediction range at a specific time step. The 20 spatio-temporal series obtained by decomposition correspond to 20 fine-tuning modules respectively. Each fine-tuning module is equipped with AdamOptimizer. After being processed by the fine-tuning module, the prediction result Y of each spatio-temporal sequence stepi ∈ R (160×160)×5 will be merged to form the final prediction output Ῡ = { y 1 ,..., y 20 ∈ R 20×(160×160)×5 .

[0088] 4) Fine-tuning training

[0089] The input data for each fine-tuning module is Y stepi ∈ R (160×160)×5 , and the process is as follows:

[0090] S1: First, perform mean-variance normalization on its result Y stepi ∈ R (160×160)×5 to calculate the mean of meteorological elements. Then calculate the variance. After that, for each data point in this time series, subtract the mean of this feature from it and divide by the standard deviation of this feature. After such processing, the data of each feature will be transformed into a distribution with a mean of 0 and a standard deviation of 1 to obtain the normalized feature Ŷ stepi_norm .

[0091] S2: At the same time, for the given sequence Ŷ stepi use LayerNorm to perform layer normalization to obtain the weight V .

[0092] S3: During the process of processing the normalized sequence, first send Ŷ stepi_norm into a linear layer, and the function of this layer is to significantly reduce the feature dimension of the data. Specifically, first flatten the sequence, and the number of channels obtained starts from the initial 5×160 ×160 channels. The first linear layer directly reduces the number of channels to 256 , and then, in order to further deepen the abstraction level of the features, the number of channels changes again, this time reducing to 128 channels. After sufficient feature extraction, the number of channels increases to 256 channels. Finally, the sequence is expanded to 5×160×160 dimensional channels to obtain the result of the linear layer Ŷ stepi_hidden ;

[0093] After each linear layer, we introduce GELUThe (Gaussian Error Linear Unit) activation function.

[0094] S4: Finally, the output of the linear layer Ŷ Vari_hidden is multiplied by the weights obtained from layer normalization V to get the product.

[0095] When the output result of the linear layer is multiplied by the weights obtained from layer normalization V it is actually a kind of reweighting of features. This weighting takes into account the distribution of each time step among different samples, enabling the model to pay more attention to those time steps that have a greater impact on the prediction result, while suppressing those features that may introduce noise.

[0096] 5) Obtaining the final prediction result

[0097] The results of each lightweight fine-tuning module are combined, and finally through this fine adjustment of features, the prediction result of multivariate meteorological element data is obtained Ῡ= { y 1 ,..., y 20} ∈ R 20×(160×160)×5 , as Figure 2 shown.

[0098] For the obtained preliminary prediction result Ῡ , the present invention constructs a fine-tuning module with a multi-layer perceptron as the core;

[0099] In each iteration, the model first performs forward propagation to calculate the predicted value, and then calculates the MSE to evaluate the prediction performance. Next, the backpropagation algorithm will update the weights and biases in the fine-tuning module according to the gradient information calculated by the MSE to reduce the prediction error and improve the prediction ability of the model;

[0100] The fine-tuning module of the present invention iterates continuously during the training process. Each iteration includes forward propagation, error calculation, backpropagation, and parameter update until a certain time step corresponding to the module reaches a predetermined early stopping number. Finally, when all modules stop training, the model training ends.

[0101] This embodiment is tested on a total of 1443 test samples in 2016. The average MSE results of five meteorological elements are shown in Figure 3 . The horizontal axis represents 5 surface meteorological elements, and the vertical axis represents the MSE, which can significantly reduce the MSE of the five surface meteorological elements; the average MSE results of the five meteorological elements at 20 time steps are shown in Figure 4 . The horizontal axis represents 20 time steps, and the vertical axis represents the MSE, which can significantly reduce the overall prediction result; the fine-tuning spatio-temporal results of T2M and V10 are shown inFigures 5 - 8 , where (a) is the prediction result of the original deep learning model Unet, (b) is the result after fine-tuning with the original deep learning model frozen, and (c) is the true value. It can be seen that the prediction result is significantly closer to the real situation after fine-tuning. In particular, the method of the present invention has a more obvious improvement effect on Figures 5 - 8 the boxed area in

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