An improved modeling method for atmospheric weighted mean temperature

Abstract

The present invention discloses an improved method for modeling the atmospheric weighted average temperature, which relates to the technical field of meteorological modeling and includes the following steps: S1, obtaining multi-source data related to the atmospheric weighted average temperature; S2, preprocessing the multi-source data; S3, constructing a least squares linear regression model to preliminarily model the atmospheric weighted average temperature; S4, constructing an improved convolutional long short-term memory network model and training it, calculating the error between the predicted value and the true value, and updating the parameters of the improved convolutional long short-term memory network model; S5, fusing the predicted values of the least squares linear regression model and the improved convolutional long short-term memory network model; S6, performing regional adaptation optimization on the fused atmospheric weighted average temperature predicted value based on the regional correction factor. The present invention combines the least squares linear regression with the improved convolutional long short-term memory network to accurately model the atmospheric weighted average temperature, and has the advantages of high accuracy, strong adaptability and outstanding regional optimization ability.

Description

Technical Field

[0001] The present invention relates to the technical field of meteorological modeling, and particularly to an improved atmospheric weighted mean temperature modeling method. Background Art

[0002] The atmospheric weighted mean temperature is a key parameter in the tropospheric delay correction of the global navigation satellite system. Its accuracy directly affects the positioning accuracy of the global navigation satellite system, the inversion effect of meteorological parameters, and the calculation performance of numerical weather prediction models. In the prior art, the calculation methods of the atmospheric weighted mean temperature mainly include empirical models, simple regression models, and some physics-based atmospheric models. Although these methods perform well under certain specific conditions, due to the limitations of the models themselves and the complexity of the calculation data, it is difficult to fully meet the requirements of different regions and complex meteorological conditions.

[0003] Empirical models are constructed based on fixed parameters and usually rely on observational data in mid-latitude regions, with poor applicability to tropical, high-latitude, and high-altitude regions. The physical assumptions of these models are usually based on static temperature and pressure distributions, ignoring the dynamic changes in the troposphere, such as strong convective activities or extreme weather conditions. In addition, it is difficult for empirical models to integrate multi-source data (such as satellite remote sensing data, ground meteorological station data, and numerical weather prediction data), and their simplicity also leads to limited ability to handle complex non-linear relationships. In practical applications, this method often introduces systematic errors, limiting the accuracy of the tropospheric delay correction of the global navigation satellite system.

[0004] Simple linear regression models are another commonly used method. These methods calculate the atmospheric weighted mean temperature by linearly fitting the relationship between meteorological characteristic variables (such as pressure, temperature) and the atmospheric weighted mean temperature. Although this method is slightly more flexible than empirical models, the limitations of the linear assumption make it difficult to capture the complex non-linear relationships between atmospheric variables, especially under dynamic meteorological conditions. In addition, with the increase in data volume and the expansion of the meteorological characteristic dimension, simple regression models are easily affected by the "curse of dimensionality", resulting in a decline in the robustness and prediction accuracy of the models. At the same time, most existing regression models lack an error correction mechanism for specific regions, resulting in significant differences in adaptability under different geographical regions and climate conditions.

[0005] In recent years, the application of deep learning technology in the meteorological field has provided new possibilities for the modeling of atmospheric weighted mean temperature. Convolutional neural networks perform well in spatial feature extraction, while long short-term memory networks have advantages in processing time series data. However, these technologies still face many problems in practical applications. On the one hand, traditional convolutional neural network and long short-term memory network models often fail to fully utilize the spatio-temporal coupling characteristics contained in meteorological data, resulting in limited model prediction ability. On the other hand, existing deep learning models usually lack a mechanism for dynamically weighting the importance of key regions and time steps in data, and are unable to effectively capture complex spatio-temporal features. In addition, the "black box" nature of deep learning models increases the difficulty of model interpretability and limits its application in fields with high precision requirements.

[0006] Existing technologies also have deficiencies in the fusion and feature extraction of multi-source meteorological data. Meteorological data usually includes ground meteorological station observations, satellite remote sensing, and numerical weather prediction output data. These data often have inconsistent temporal and spatial resolutions, and data missing, noise, and outliers also pose challenges to modeling. Traditional data preprocessing methods lack systematicness and standardization, making it difficult to effectively utilize the information in multi-source data, resulting in input features insufficient to support the requirements of complex models. In addition, current modeling technologies usually fail to fully consider the temporal, spatial, and vertical profile features in meteorological data, and do not fully exploit the potential of multi-dimensional meteorological data.

[0007] In terms of regional adaptability, existing atmospheric weighted mean temperature models lack a dynamic optimization mechanism for different geographical regions and climate conditions, and usually rely on static empirical corrections or offline optimizations. This method is difficult to adapt to real-time changing meteorological conditions, resulting in significant differences in prediction accuracy among regions. For regions with significant climate differences, such as tropical and high-latitude regions, the errors of existing models are more obvious. In addition, the regional adaptation of the model often ignores the quantification and correction of systematic biases, further limiting the actual application effect of the model.

[0008] In summary, the existing technologies mainly have the following deficiencies in the modeling of atmospheric weighted mean temperature: the accuracy of empirical models and simple regression models is limited and difficult to adapt to complex meteorological conditions; the ability of multi-source data fusion and feature extraction is insufficient, and the multi-dimensional information in meteorological data is not effectively utilized; the feature extraction and time series modeling ability of deep learning models is limited, and the complex relationships in the data are not fully captured; the model lacks regional adaptability and fails to effectively optimize for different geographical regions and climate conditions. Therefore, there is an urgent need for an improved method that combines statistical modeling and deep learning technology, which can fully utilize multi-source data, enhance feature extraction ability, and dynamically optimize the regional correction mechanism to improve the accuracy, applicability, and stability of atmospheric weighted mean temperature modeling. Summary of the Invention

[0009] An object of the present invention is to propose an improved modeling method for atmospheric weighted mean temperature. The present invention combines the least squares linear regression model with an improved convolutional long short-term memory network, makes full use of multi-source meteorological data, accurately models the atmospheric weighted mean temperature by integrating statistical modeling and deep learning techniques, and proposes a dynamic regional correction and optimization mechanism, which has the advantages of high calculation accuracy, strong adaptability, and outstanding dynamic optimization ability for regional errors.

[0010] An improved modeling method for atmospheric weighted mean temperature according to an embodiment of the present invention includes the following steps:

[0011] S1. Obtain multi-source data related to the atmospheric weighted mean temperature;

[0012] S2. Preprocess the multi-source data to generate a multi-dimensional meteorological feature input data set;

[0013] S3. Construct a least squares linear regression model, and based on the multi-dimensional meteorological feature input data set, use the least squares linear regression model to preliminarily model the atmospheric weighted mean temperature, fit the relationship between the atmospheric weighted mean temperature and meteorological feature variables, and generate the predicted value of the least squares linear regression model;

[0014] S4. Construct an improved convolutional long short-term memory network model, use the multi-dimensional meteorological feature input data set as a training sample to train the convolutional long short-term memory network model, calculate the error between the predicted value and the true value according to the loss function, update the parameters of the improved convolutional long short-term memory network model, and generate the predicted value of the improved convolutional long short-term memory network model;

[0015] S5. Integrate the predicted values of the least squares linear regression model and the improved convolutional long short-term memory network model to generate a fused atmospheric weighted mean temperature predicted value;

[0016] S6. Based on the fused atmospheric weighted mean temperature predicted value, perform regional adaptation optimization on the fused atmospheric weighted mean temperature predicted value in combination with a regional correction factor.

[0017] Optionally, the multi-source data includes the temperature, humidity, air pressure observed by a ground meteorological station and satellite remote sensing data.

[0018] Optionally, the preprocessing includes time and space alignment, data cleaning, standardization processing, and feature variable extraction, and the multi-dimensional meteorological feature input data set includes time features, space features, and vertical profile features.

[0019] Optionally, the S3 specifically includes:

[0020] S31. Obtain the air pressure and temperature data in the multi-dimensional meteorological feature input data set;

[0021] S32. Construct a least - squares linear regression model to fit the atmospheric weighted mean temperature and the non - linear relationship between air pressure and temperature data:

[0022] ;

[0023] ;

[0024] ;

[0025] Among them, represents the atmospheric weighted mean temperature, , , , , , and represent the coefficients of the regression model, which are determined by data fitting. P represents air pressure, T represents temperature, DOY represents the day of the year, and represent the day - number terms;

[0026] S33. Use the least - squares method to perform parameter fitting on historical meteorological data to obtain the optimal model parameters;

[0027] S34. Calculate the mean squared error MSE of the least - squares linear regression model to evaluate the performance of the least - squares linear regression model:

[0028] ;

[0029] Among them, MSE represents the mean squared error of the improved polynomial regression model, represents the true value of the i - th sample, represents the predicted value of the i - th sample, N represents the number of samples, represents the mean of the true values;

[0030] S35. Output the predicted value of the improved polynomial regression model .

[0031] Optionally, the specific steps of S4 include:

[0032] S41. Construct an improved convolutional long - short - term memory network model, including:

[0033] An input layer that receives a multi - dimensional meteorological feature input data set, and sets the dimension of the input tensor to N×Y×M. In the dimension of the input tensor, N represents the number of samples, Y represents the time step, and M represents the number of meteorological feature variables;

[0034] An improved convolutional layer that introduces a spatial attention mechanism to enhance the feature extraction ability for important regions of spatial correlation. The output dimension of the improved convolutional layer is N×Y×C, where C in the output dimension of the improved convolutional layer represents the number of channels;

[0035] An improved long short-term memory network layer that introduces a gated attention mechanism to enhance the ability to allocate weights to important information at different time steps. The output dimension of the improved long short-term memory network layer is N×H, where H in the output dimension of the improved long short-term memory network layer represents the number of hidden units;

[0036] An output layer that generates an atmospheric weighted average temperature prediction value, with an output dimension of N×1;

[0037] S42. Input the multi-dimensional meteorological feature input dataset into the improved convolutional layer for feature extraction. The features Z extracted by the convolutional layer introduce spatial attention weights:

[0038] ;

[0039] Among them, represents the enhanced features output by the convolutional layer, represents the spatial attention weight matrix, and respectively represent the convolutional kernel weights and biases of the spatial attention mechanism, Sigmoid represents the activation function, represents the convolutional calculation, represents the element-wise product calculation, and Z represents the features extracted by the improved convolutional layer;

[0040] S43. Add time attention weights to the input gate, forget gate, and output gate formulas of the traditional long short-term memory network layer to form an improved long short-term memory network, and input the enhanced features into the improved long short-term memory network:

[0041] ;

[0042] ;

[0043] ;

[0044] ;

[0045] ;

[0046] ;

[0047] Among them, 、 and respectively represent the activation values of the input gate, forget gate, and output gate, with dimension H, and respectively represent the cell state and hidden state, with dimension H, 、 、 and represent the weight matrices input to each gate and state, with dimension H×C, 、 、 and represent the bias terms, with dimension H, 、 、 and represent the weight matrices from the hidden state to each gate and state, with dimension H×H, represents the temporal attention weights, with dimension Y, represents the temporal attention weight matrix, with dimension Y×(C + H), represents the temporal attention bias term, with dimension Y, represents the concatenation of the input at time step t and the hidden state of the previous time step, represents the input at time step t, and tanh represents the hyperbolic tangent activation function, represents the ReLU activation function;

[0048] S44. Define the mean squared error loss function and update the parameters using the Adam algorithm:

[0049] ;

[0050] wherein, represents the mean squared error loss function, represents the true value of the i-th sample, represents the predicted value of the i-th sample, and N represents the number of samples, represents the parameters of the improved convolutional long short-term memory network model;

[0051] S45. The output layer generates the predicted value of the improved convolutional long short-term memory network model:

[0052] ;

[0053] wherein, represents the predicted value of the improved convolutional long short-term memory network model, represents the output layer weights, represents the output layer bias.

[0054] Optionally, the S6 specifically includes:

[0055] S61. Define the regional correction factor , the systematic deviation of the prediction error within the quantization region i:

[0056] ;

[0057] where, represents the correction factor for region i, represents the number of samples within region i, represents the true atmospheric weighted average temperature of the j-th sample, represents the predicted value of the fused atmospheric weighted average temperature for the j-th sample;

[0058] S62. Divide region i according to geographical location and meteorological conditions, and each region contains the range of latitude and longitude coordinates corresponding to the samples:

[0059] ;

[0060] ;

[0061] where, and respectively represent the latitude and longitude of the j-th sample, , , and represent the upper and lower bounds of the geographical range of region i;

[0062] S63. Use the regional correction factor to correct the predicted value of the fused atmospheric weighted average temperature to generate an optimized predicted value adapted to the region:

[0063] ;

[0064] where, represents the predicted value of the atmospheric weighted average temperature optimized for the region adaptation, represents the correction factor for region i to which sample j belongs, represents the predicted value of the fused atmospheric weighted average temperature for the j-th sample;

[0065] S64. In real-time prediction, dynamically update the regional correction factor , and perform correction in combination with new observation data:

[0066] ;

[0067] where, represents the updated regional correction factor, represents the historical regional correction factor, represents the correction amount based on the new observation data, Indicates the update rate.

[0068] The beneficial effects of the present invention are as follows:

[0069] First, during the modeling process, the present invention uses the least squares linear regression model to initially model the atmospheric weighted mean temperature. The introduction of the least squares method provides a good physical interpretability for modeling the linear relationship between meteorological variables and the atmospheric weighted mean temperature, and evaluates the model performance through the mean square error, making the model have higher robustness and controllability.

[0070] Second, in order to further improve the ability to capture non-linear relationships, the present invention designs a deep learning model based on an improved convolutional long short-term memory network. The convolutional layer enhances the ability to extract spatial features by introducing a spatial attention mechanism to dynamically identify important regions in multi-source meteorological data, significantly improving the adaptability of the model to complex geographical conditions. At the same time, the improved long short-term memory network layer uses a gated attention mechanism to dynamically assign weights to time series data, thereby accurately capturing the time-dependent characteristics of the atmospheric weighted mean temperature. This cascaded deep learning structure fully integrates spatial and temporal characteristics, providing a powerful tool for non-linear modeling of the atmospheric weighted mean temperature.

[0071] In addition, the present invention further adopts a model fusion strategy, combining the prediction results of the least squares linear regression model with the prediction results of the improved convolutional long short-term memory network model to generate more accurate and stable predicted values of the atmospheric weighted mean temperature. By integrating the interpretability of traditional statistical models and the non-linear modeling ability of deep learning models, the limitations of single models are overcome, achieving high-precision prediction results under different meteorological conditions.

[0072] Finally, regarding the regional adaptability problem, the present invention innovatively proposes a dynamic optimization mechanism based on regional correction factors. By quantifying the systematic deviation of the prediction error within the region and combining a real-time update strategy, the present invention can dynamically adjust the prediction results of the model to adapt to changes in different geographical regions and meteorological conditions. Description of the Drawings

[0073] The drawings are used to provide a further understanding of the present invention, and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention, but do not constitute a limitation to the present invention. In the drawings:

[0074] Figure 1 is a flowchart of an improved method for modeling the atmospheric weighted mean temperature proposed by the present invention;

[0075] Figure 2 is a schematic structural diagram of an improved convolutional long short-term memory network model for an improved method for modeling the atmospheric weighted mean temperature proposed by the present invention. Detailed implementation manners

[0076] Now, the present invention will be further described in detail with reference to the accompanying drawings. These drawings are all simplified schematic diagrams, only illustrating the basic structure of the present invention in a schematic manner, so they only show the components related to the present invention.

[0077] Refer to Figure 1 and Figure 2 , an improved atmospheric weighted mean temperature modeling method, comprising the following steps:

[0078] S1. Obtain multi-source data related to the atmospheric weighted mean temperature;

[0079] S2. Preprocess the multi-source data to generate a multi-dimensional meteorological feature input data set;

[0080] S3. Construct a least squares linear regression model, and based on the multi-dimensional meteorological feature input data set, use the least squares linear regression model to preliminarily model the atmospheric weighted mean temperature, fit the relationship between the atmospheric weighted mean temperature and the meteorological feature variables, and generate the predicted value of the least squares linear regression model;

[0081] S4. Construct an improved convolutional long short-term memory network model, use the multi-dimensional meteorological feature input data set as a training sample to train the convolutional long short-term memory network model, calculate the error between the predicted value and the true value according to the loss function, update the parameters of the improved convolutional long short-term memory network model, and generate the predicted value of the improved convolutional long short-term memory network model;

[0082] S5. Fuse the predicted values of the least squares linear regression model and the improved convolutional long short-term memory network model to generate a fused atmospheric weighted mean temperature predicted value;

[0083] S6. Based on the fused atmospheric weighted mean temperature predicted value, combine the regional correction factor to perform regional adaptation optimization on the fused atmospheric weighted mean temperature predicted value.

[0084] In this embodiment, the multi-source data includes the temperature, humidity, air pressure observed by a ground meteorological station and satellite remote sensing data.

[0085] In this embodiment, the preprocessing includes time and space alignment, data cleaning, normalization processing and feature variable extraction, and the multi-dimensional meteorological feature input data set includes time features, space features and vertical profile features.

[0086] In this embodiment, the S3 specifically includes:

[0087] S31. Obtain the air pressure and temperature data in the multi-dimensional meteorological feature input data set;

[0088] S32. Construct a least squares linear regression model to fit the non-linear relationship between the atmospheric weighted mean temperature and the barometric pressure and temperature data:

[0089] ;

[0090] ;

[0091] ;

[0092] Among them, represents the atmospheric weighted mean temperature, , , , , , and represent the coefficients of the regression model, which are determined by data fitting. P represents the barometric pressure, T represents the temperature, DOY represents the day of the year, and represent the day terms;

[0093] S33. Use the least squares method to perform parameter fitting on historical meteorological data to obtain the optimal model parameters;

[0094] S34. Calculate the mean square error MSE of the least squares linear regression model to evaluate the performance of the least squares linear regression model:

[0095] ;

[0096] Among them, MSE represents the mean square error of the improved polynomial regression model, represents the true value of the i-th sample, represents the predicted value of the i-th sample, N represents the number of samples, represents the mean of the true values;

[0097] S35. Output the predicted value of the improved polynomial regression model .

[0098] In this embodiment, the specific steps of S4 include:

[0099] S41. Construct an improved convolutional long short-term memory network model, including:

[0100] An input layer that receives a multi-dimensional meteorological feature input data set, and sets the dimension of the input tensor to N×Y×M. In the dimension of the input tensor, N represents the number of samples, Y represents the time step, and M represents the number of meteorological feature variables;

[0101] An improved convolutional layer introduces a spatial attention mechanism to enhance the feature extraction ability for important regions of spatial correlation. The output dimension of the improved convolutional layer is N×Y×C, where C in the output dimension of the improved convolutional layer represents the number of channels;

[0102] An improved long short-term memory network layer introduces a gated attention mechanism to enhance the ability to allocate weights to important information at different time steps. The output dimension of the improved long short-term memory network layer is N×H, where H in the output dimension of the improved long short-term memory network layer represents the number of hidden units;

[0103] An output layer generates an atmospheric weighted average temperature prediction value, and the output dimension is N×1;

[0104] S42. Input the multi-dimensional meteorological feature input dataset into the improved convolutional layer for feature extraction. The features Z extracted by the convolutional layer introduce spatial attention weights:

[0105] ;

[0106] Among them, represents the enhanced features output by the convolutional layer, represents the spatial attention weight matrix, and respectively represent the convolutional kernel weights and biases of the spatial attention mechanism. Sigmoid represents the activation function, represents the convolutional calculation, represents the element-wise product calculation, and Z represents the features extracted by the improved convolutional layer;

[0107] S43. Add time attention weights to the input gate, forget gate, and output gate formulas of the traditional long short-term memory network layer to form an improved long short-term memory network, and input the enhanced features into the improved long short-term memory network:

[0108] ;

[0109] ;

[0110] ;

[0111] ;

[0112] ;

[0113] ;

[0114] Among them, 、 and respectively represent the activation values of the input gate, forget gate, and output gate, with dimension H, and respectively represent the cell state and hidden state, with dimension H, 、 、 and represent the weight matrices input to each gate and state, with dimension H×C, 、 、 and represent the bias terms, with dimension H, 、 、 and represent the weight matrices from the hidden state to each gate and state, with dimension H×H, represents the temporal attention weight, with dimension Y, represents the temporal attention weight matrix, with dimension Y×(C + H), represents the temporal attention bias term, with dimension Y, represents the concatenation of the input at time step t and the hidden state of the previous time step, represents the input at time step t, and tanh represents the hyperbolic tangent activation function, represents the ReLU activation function;

[0115] S44. Define the mean squared error loss function and update the parameters using the Adam algorithm:

[0116] ;

[0117] where, represents the mean squared error loss function, represents the true value of the i-th sample, represents the predicted value of the i-th sample, and N represents the number of samples, represents the parameters of the improved convolutional long short-term memory network model;

[0118] S45. The output layer generates the predicted value of the improved convolutional long short-term memory network model:

[0119] ;

[0120] where, represents the predicted value of the improved convolutional long short-term memory network model, represents the output layer weight, represents the output layer bias.

[0121] In this embodiment, the S6 specifically includes:

[0122] S61. Define the regional correction factor , the systematic bias of the prediction error within the quantization region i:

[0123] ;

[0124] where, represents the correction factor for region i, represents the number of samples within region i, represents the true atmospheric weighted average temperature of the jth sample, represents the predicted value of the fused atmospheric weighted average temperature for the jth sample;

[0125] S62. Divide region i according to geographical location and meteorological conditions, and each region contains the range of longitude and latitude coordinates corresponding to the samples:

[0126] ;

[0127] ;

[0128] where, and respectively represent the latitude and longitude of the jth sample, , , and represent the upper and lower bounds of the geographical range of region i;

[0129] S63. Use the regional correction factor to correct the predicted value of the fused atmospheric weighted average temperature to generate an optimized predicted value adapted to the region:

[0130] ;

[0131] where, represents the predicted value of the atmospheric weighted average temperature optimized for the region, represents the correction factor for region i to which sample j belongs, represents the predicted value of the fused atmospheric weighted average temperature for the jth sample;

[0132] S64. In real-time prediction, dynamically update the regional correction factor , and perform correction in combination with new observed data:

[0133] ;

[0134] where, represents the updated regional correction factor, represents the historical regional correction factor, represents the correction amount based on the new observed data, Indicates the update rate.

[0135] Example 1:

[0136] To verify the feasibility of the present invention in implementation, the present invention was applied to an actual tropospheric delay correction scenario of a global navigation satellite system. Ground meteorological observation data and satellite remote sensing data in the central region of China were selected as the main research objects. The data collection scope covered the regions of Henan, Hubei, and Hunan provinces, and the time span was from January to December 2023. The ground observation data included air pressure, temperature, and humidity, and the remote sensing data provided vertical atmospheric profile information and cloud top temperature. The selected region had typical monsoon climate characteristics, with variable meteorological conditions throughout the year, including high temperature and high humidity in summer and low temperature and dryness in winter, which was suitable for evaluating the modeling accuracy and adaptability of the present invention.

[0137] First, the multi-source meteorological data was preprocessed, including time synchronization, spatial alignment, and outlier removal. Taking an hour as the time step, after standardizing each meteorological characteristic variable, a multi-dimensional meteorological characteristic input data set was generated. Then, a preliminary model of the atmospheric weighted mean temperature was established using the least squares linear regression model. By fitting the key meteorological variables (such as the daily variation of air pressure and the annual cycle of temperature) extracted by principal component analysis, a preliminary prediction result was obtained.

[0138] Subsequently, an improved convolutional long short-term memory network model was introduced for deep learning modeling. In the model design, the convolutional layer effectively extracted the features of key regions in the air pressure field and temperature field by introducing a spatial attention mechanism, and the long short-term memory network layer strengthened the modeling ability of the time series dependence relationship through a gated attention mechanism. To further improve the regional adaptability of the model, a regional correction factor was introduced based on fusing the predicted values of the two models. By dividing geographical regions (such as plain areas and hilly areas), the prediction errors in each region were dynamically corrected, where K in the table represents Kelvin.

[0139] Table 1 Comparison table of experimental data

[0140] ;

[0141] It can be seen from the performance comparison table of different models in the atmospheric weighted mean temperature modeling in the above experiments that the improved model proposed by the present invention has significantly better prediction accuracy and adaptability under various meteorological conditions and geographical regions. Specifically, the traditional empirical model performs relatively stably in the mid-latitude region, but in regions with complex meteorological conditions such as the tropics and polar regions, the prediction errors are large, and the mean square errors reach 4.5K and 5.0K respectively, and the overall regional average error is 4.2K. This indicates that the empirical model is restricted by its static assumptions and regional applicability limitations and is difficult to provide reliable predictions under complex conditions.

[0142] In contrast, the least squares linear regression model significantly reduces the prediction errors in the mid-latitudes and tropical regions through principal component analysis and linear fitting of meteorological characteristic variables. The mean squared error in the mid-latitudes drops to 2.8K, and the mean squared error in the tropical region is 3.8K. However, in the polar region, the model still shows certain limitations, with the mean squared error only decreasing from 5.0K to 4.5K, indicating that it is still difficult to capture the non-linear changes under complex meteorological conditions under the linear assumption.

[0143] After introducing the improved convolutional long short-term memory network model, the overall prediction performance is enhanced. The model enhances the ability to capture spatio-temporal features through the spatial attention mechanism of the convolutional layer and the temporal attention mechanism of the long short-term memory network. The mean squared errors in the mid-latitudes, tropical, and polar regions drop to 1.5K, 2.0K, and 2.8K respectively, and the overall regional average MSE is 2.1K. This performance improvement indicates that the improved deep learning model performs better than the traditional statistical model under complex meteorological conditions, especially with stronger modeling capabilities for non-linear and dynamic characteristics.

[0144] Through further model fusion and regional correction optimization, the final prediction results of the present invention achieve higher accuracy in all regions. For example, in the mid-latitudes, the corrected mean squared error of the fusion model drops from 1.3 K to 1.0K; in the tropical region, from 1.8K to 1.5K; in the polar region, from 2.5K to 2.0K. The overall regional average mean squared error is reduced to 1.5K, a 59.5% reduction compared to the least squares method model and a 28.6% reduction compared to the improved convolutional long short-term memory network model. In addition, in the tropospheric delay correction, the error of the traditional empirical model is 4.0 cm, while the fusion model of the present invention (after regional correction optimization) reduces the error to 2.1 cm, demonstrating excellent performance in practical applications.

[0145] In summary, the improved method of the present invention demonstrates excellent performance in atmospheric weighted mean temperature modeling through multi-model combination, regional correction optimization, and the introduction of deep learning technology. It not only improves the prediction accuracy but also effectively solves the adaptability problem of traditional methods in complex regions.

[0146] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, making equivalent substitutions or changes, should be covered by the protection scope of the present invention.

Claims

1. An improved atmospheric weighted average temperature modeling method, characterized in that: The steps include: S1. Obtain multi-source data related to atmospheric weighted average temperature; S2. Preprocessing the multi-source data to generate a multi-dimensional meteorological feature input data set; S3. Construct a least squares linear regression model. Based on the multidimensional meteorological characteristic input data set, use the least squares linear regression model to preliminarily model the atmospheric weighted average temperature, fit the relationship between the atmospheric weighted average temperature and the meteorological characteristic variables, and generate the predicted value of the least squares linear regression model. S4. Construct an improved convolutional long short-term memory network model, use the multidimensional meteorological feature input data set as a training sample, train the convolutional long short-term memory network model, calculate the error between the predicted value and the true value according to the loss function, update the parameters of the improved convolutional long short-term memory network model, and generate the predicted value of the improved convolutional long short-term memory network model; S5, fusion least squares linear regression model and improved convolutional long short-term memory network model prediction value, generate fusion atmospheric weighted average temperature prediction value; S6. Based on the fused atmospheric weighted average temperature prediction value, the fused atmospheric weighted average temperature prediction value is optimized for regional adaptation in combination with the regional correction factor; The S4 specifically includes: S41. Construct an improved convolutional long short-term memory network model, including: The input layer receives a multidimensional meteorological feature input data set and sets the dimension of the input tensor to N×Y×M, where N represents the number of samples, Y represents the time step, and M represents the number of meteorological feature variables; The improved convolution layer introduces a spatial attention mechanism to enhance the feature extraction capability of important areas of spatial correlation. The output dimension of the improved convolution layer is N×Y×C, where C in the output dimension of the improved convolution layer represents the number of channels. The improved long short-term memory network layer introduces a gated attention mechanism to enhance the ability to allocate weights to important information at different time steps. The output dimension of the improved long short-term memory network layer is N×H, where H in the output dimension of the improved long short-term memory network layer represents the number of hidden units. Output layer, generates the predicted value of atmospheric weighted average temperature, and the output dimension is N×1; S42, the multi-dimensional meteorological feature input data set is input into the improved convolution layer for feature extraction, and the feature Z extracted by the convolution layer introduces the spatial attention weight: ; in, represents the enhanced features of the convolutional layer output, represents the spatial attention weight matrix, and They represent the convolution kernel weight and bias of the spatial attention mechanism respectively, Sigmoid represents the activation function, represents the convolution calculation, represents the point-by-point product calculation, and Z represents the features extracted by the improved convolutional layer; S43. Adding temporal attention weights to the input gate, forget gate, and output gate formulas of the traditional long short-term memory network layer Form an improved long short-term memory network to enhance the features Input to the improved long short-term memory network: ; ; ; ; ; ; in, , and Represent the activation values ​​of the input gate, forget gate and output gate respectively, with a dimension of H. and Represent the unit state and hidden state respectively, with a dimension of H, , , and Represents the weight matrix input to each gate and state, with a dimension of H×C, , , and represents the bias term, with dimension H, , , and Represents the weight matrix from hidden state to each gate and state, with dimension H×H, represents the time attention weight, with dimension Y, represents the temporal attention weight matrix, with dimension Y×(C+H), represents the temporal attention bias term, with dimension Y, represents the concatenation of the input at time step t and the hidden state at the previous time step, represents the input of time step t, tanh represents the hyperbolic tangent activation function, ReLU activation function. S44. Define the mean square error loss function and use the Adam algorithm to update the parameters: ; in, represents the mean square error loss function, represents the true value of the i-th sample, represents the predicted value of the i-th sample, N represents the number of samples, Represents the parameters of the improved convolutional long short-term memory network model; S45, the output layer generates the predicted value of the improved convolutional long short-term memory network model: ; in, represents the predicted value of the improved convolutional long short-term memory network model, represents the output layer weight, represents the output layer bias; The S6 specifically includes: S61. Define the regional correction factor , quantifies the systematic deviation of the prediction error in region i: ; in, represents the correction factor for region i, represents the number of samples in region i, represents the real atmospheric weighted average temperature of the jth sample, represents the predicted value of the fused atmospheric weighted average temperature of the jth sample; S62. Divide region i according to geographical location and meteorological conditions, and each region contains the latitude and longitude coordinate range corresponding to the sample: ; ; in, and Respectively represent the latitude and longitude of the jth sample, , , and Indicates the upper and lower limits of the geographical range of region i; S63, use area correction factor Fusion atmospheric weighted mean temperature forecast Make corrections to generate the predicted value after regional adaptation optimization: ; in, represents the predicted value of atmospheric weighted average temperature after regional adaptation optimization, represents the correction factor of region i to which sample j belongs, represents the predicted value of the fused atmospheric weighted average temperature of the jth sample; S64. Dynamically update the regional correction factor in real-time prediction , and correct it with the new observation data: ; in, represents the updated regional correction factor, represents the historical area correction factor, represents the correction amount based on new observation data, Indicates the update rate.

2. The improved atmospheric weighted average temperature modeling method according to claim 1, characterized in that: The multi-source data include temperature, humidity, air pressure observed by ground meteorological stations and satellite remote sensing data.

3. The improved atmospheric weighted average temperature modeling method according to claim 1, characterized in that: The preprocessing includes time and space alignment, data cleaning, standardization and feature variable extraction, and the multidimensional meteorological feature input data set includes time features, space features and vertical profile features.

4. The improved atmospheric weighted average temperature modeling method according to claim 1, characterized in that: The S3 specifically includes: S31, obtaining air pressure and temperature data in a multi-dimensional meteorological feature input data set; S32. Construct a least squares linear regression model to fit the atmospheric weighted average temperature Nonlinear relationship between pressure and temperature data: ; ; ; in, represents the weighted mean temperature of the atmosphere, , , , , , and represents the coefficient of the regression model, which is determined by data fitting, P represents air pressure, T represents temperature, DOY represents the day of the year, and Indicates the number of days; S33, using the least squares method to perform parameter fitting on historical meteorological data to obtain optimal model parameters; S34. Calculate the mean square error (MSE) of the least squares linear regression model to evaluate the performance of the least squares linear regression model: ; Among them, MSE represents the mean square error of the improved polynomial regression model, represents the true value of the i-th sample, represents the predicted value of the i-th sample, N represents the number of samples, represents the true value mean; S35. Output the predicted value of the improved polynomial regression model .

Images