A seasonally adaptive integrated and dynamic anomaly correction temperature prediction system and method

Abstract

This invention relates to the field of short-term meteorological climate prediction, specifically disclosing a seasonal adaptive integration and dynamic anomaly correction temperature prediction system and method. The system includes a data acquisition module, a modeling and calculation module, and an evaluation and visualization module. The method includes: S1, fusing multi-source data to construct time-coded, lag, and cross-feature features, and generating a dynamic climate benchmark with variable weights; S2, based on the Stacking integration framework, using Ridge+LightGBM in winter, SVR and multinomial regression in summer, weighting during the transition season, and residual calibration in winter; S3, three-level anomaly evaluation, combined with benchmark calculation and visualization, with fine-tuning when the matching rate is low. The seasonal adaptive integration and dynamic anomaly correction temperature prediction system and method proposed in this invention solves the problems of difficulty in nonlinear capture, benchmark rigidity, and poor seasonal adaptation, effectively improving the accuracy of seasonal temperature prediction.

Description

Technical Field

[0001] This invention relates to the field of short-term meteorological climate prediction technology, specifically to a seasonal adaptive integration and dynamic anomaly correction temperature prediction system and method. Background Technology

[0002] In short-term meteorological climate forecasting practice, existing technologies mainly revolve around three core aspects: fusion of meteorological element characteristics, seasonal integrated modeling and calibration, and temperature anomaly assessment. Regarding the fusion of meteorological element characteristics and adjustment of the climate baseline, existing technologies often construct features using simple linear combinations of raw meteorological variables such as wind speed and soil temperature, while relying on a fixed static climate baseline for analysis. For seasonal integrated frameworks and calibration, traditional schemes typically use a single model or a simple model averaging method for predictive modeling, without specifically designing for the differences in meteorological characteristics across different seasons. Regarding temperature anomaly assessment and visualization technologies, existing technologies generally calculate simple anomaly values ​​using a static climate baseline, only achieving basic data display and lacking multi-dimensional in-depth analysis.

[0003] However, existing technologies have many obvious shortcomings. In terms of feature fusion and benchmark adjustment, simple linear combinations cannot effectively capture the complex nonlinear interactions between meteorological elements, and static climate benchmarks are difficult to reflect the dynamic trend of recent global warming, resulting in serious deficiencies in the accuracy of feature representation and the timeliness of the benchmark. In terms of seasonal integration and calibration, single models or simple averaging methods cannot adapt to the seasonal heterogeneity of meteorological elements, especially in winter, where insufficient capture of key features such as soil temperature and snow cover can lead to significantly higher prediction errors than in other seasons. In terms of anomaly assessment, simple anomaly calculations under static benchmarks cannot quantify the warming trend and the characteristic changes during seasonal transitions, and visualizations are also unable to intuitively reflect the reliability distribution of prediction results.

[0004] These problems collectively result in significant shortcomings in the existing temperature forecasting methods regarding long-term forecast accuracy, seasonal adaptability, and comprehensiveness of assessment, making it difficult to quickly, economically, and accurately meet the needs of short-term meteorological climate forecasting.

[0005] Therefore, how to construct a forecasting framework that adapts to seasonal heterogeneity and achieve accurate assessment under dynamic benchmarks has become a key and difficult issue that urgently needs to be addressed in the field of short-term meteorological climate forecasting. Summary of the Invention

[0006] To address the aforementioned problems in the prior art, this invention provides a temperature forecasting system and method with seasonal adaptive integration and dynamic anomaly correction. This effectively solves the problems of insufficient capture of nonlinear interactions of meteorological elements, rigid static benchmarks, poor seasonal adaptation, and lack of quantitative and visual anomaly assessment. It effectively improves forecast accuracy, seasonal adaptability, comprehensive assessment, and intuitive results.

[0007] To achieve the above objectives, this invention proposes a seasonal adaptive integrated and dynamic anomaly correction temperature prediction system, comprising: a data acquisition module, a modeling and calculation module, and an evaluation and visualization module;

[0008] The data acquisition module is used to integrate meteorological station observation data, reanalysis data and satellite remote sensing data, and perform spatiotemporal alignment and standardization processing;

[0009] The modeling and calculation module is used to construct time codes, lag features and cross features, generate dynamic climate benchmarks, and carry out integrated modeling and special calibration based on the seasonal adaptive Stacking integration framework.

[0010] The evaluation visualization module is used to construct a three-level dynamic anomaly correction system, calculate the prediction success rate and output multi-dimensional visualization results, and establish a real-time feedback system to trigger model fine-tuning.

[0011] A seasonally adaptive integrated and dynamic anomaly correction method for temperature forecasting includes:

[0012] S1. Cross-features of meteorological elements and generation of dynamic climate benchmarks: Multi-source data fusion and feature construction are carried out. Meteorological station observation data, reanalysis data and satellite remote sensing data are integrated and spatiotemporally aligned. Time coding, lag features and cross features are constructed. At the same time, dynamic climate benchmark modeling is designed and dynamic climate benchmarks are generated by using variable weight adjustment methods.

[0013] S2. Seasonal adaptive ensemble modeling and special calibration are carried out using deep learning fusion algorithms: a heterogeneous model of the base layer is built based on the seasonal adaptive Stacking ensemble framework. The winter model uses a combination of Ridge regression and LightGBM, the summer model uses a combination of SVR and multinomial regression, and the transition season uses weighted output. Special calibration for winter is carried out through separate residual modeling.

[0014] S3, Three-level dynamic anomaly assessment and visualization output: Construct a three-level dynamic anomaly correction system, calculate anomaly values ​​in conjunction with dynamic climate benchmarks, assess prediction success rate, output prediction results through a multi-dimensional visualization engine, and establish a real-time feedback system to trigger model fine-tuning and expand the training set time range when the prediction matching rate is low.

[0015] Preferably, in S1, the multi-source data fusion and feature construction specifically includes: collecting long-term meteorological data including temperature, wind speed, soil temperature, sea surface temperature, and snow cover rate; after standardization processing, generating periodic features, lag features, and climate anomaly features, among which the periodic features and lag features are used to characterize the annual cycle changes of temperature data; the constructed multi-source data fusion and feature also includes optional components such as soil temperature difference features, 7-day moving average of snow cover, and quarterly interaction term features, and the half-life of the rolling time window can be adjusted, set to 90 days in rapidly warming regions.

[0016] Preferably, in S1, the dynamic climate benchmark modeling and cross-feature design specifically include: updating the climate benchmark using a sliding window, automatically generating seasonal adjustment coefficients through land-sea thermal difference analysis; adding cross-features of wind speed and soil temperature difference in winter to capture the cold wave effect, and adding cross-features of sea surface temperature and annual cycle coupling in summer to enhance the high temperature prediction capability.

[0017] Preferably, in S2, the heterogeneous model of the base layer, constructed based on the seasonal adaptive Stacking ensemble framework, includes four models: Ridge regression, LightGBM, SVR, and multinomial regression. The weighted output weights for the transition season are determined inversely by the historical error of the same period.

[0018] Preferably, in S2, the winter sub-model is fused using a combination of Ridge regression and LightGBM. Specifically, the winter sub-model fusion includes: Ridge regression focusing on optimizing dominant winter features such as soil temperature and snow cover; Ridge regression employs L2 regularization with a regularization strength parameter set to 0.5; and LightGBM uses a score loss function, the specific form of which is:

[0019] ;

[0020] In the formula, For predicted values Compared with the true value The matching score, and when If the value is greater than 0.5, the prediction result is considered valid.

[0021] Preferably, in S2, the summer sub-model is fused using a combination of SVR and multinomial regression. Specifically, the summer sub-model fusion includes: SVR employing an RBF kernel tailored to sea surface temperature and solar radiation characteristics; the penalty coefficient C of the kernel parameter is set to 1.5; and the threshold of the kernel parameter insensitive loss function is set... Set to 0.1; the polynomial regression uses a second-order method to capture the quadratic relationship between temperature and sunshine duration. The input features of the polynomial regression are the coupled cross features of summer sea surface temperature and annual cycle.

[0022] Preferably, in S2, the separate residual modeling for winter-specific calibration specifically includes: adopting a two-stage bias correction strategy, first calculating the validation set residuals, then establishing a seasonally specific calibration model, using linear regression with L2 regularization for winter calibration, and dynamically updating the calibration parameters through a rolling 3-year time window with the window weight configured as exponential decay.

[0023] Preferably, the objective function of the L2-regularized linear regression is:

[0024] ;

[0025] In the formula, To verify the set residual, i.e., the difference between observed and predicted values, The feature is a one-hot encoding based on the month. The regularization coefficient is . This represents the number of samples within the rolling time window.

[0026] The exponential decay calculation formula for the sliding window weight configuration of the winter sub-model is as follows:

[0027] ;

[0028] In the formula, T is the current time, t is the historical moment within the rolling time window, τ is the half-life, and α controls the decay rate.

[0029] Preferably, in S3, the three-level dynamic anomaly assessment system specifically includes a dynamic monthly climate benchmark based on a sliding 20-year window, a correction term incorporating interannual warming trends, and a real-time monitoring module for quarterly sign matching rate. The sign matching rate is calculated through the consistency of signs between observed and predicted anomalies. The multi-dimensional visualization engine output includes a seasonal trend parallel coordinate plot and an anomaly heatmap, where the seasonal trend parallel coordinate plot is used to compare predicted and observed anomalies, and the anomaly heatmap is used to display the spatial distribution characteristics of temperature anomalies. The calculation formula for the prediction success rate assessment is:

[0030] ;

[0031] In the formula, N represents the total number of evaluation samples, i.e., the number of observed anomaly-predicted anomaly data pairs used in the prediction success rate calculation, specifically the time series sample size corresponding to the dynamic climate baseline. To observe the anomaly, To predict anomalies, For indicator functions;

[0032] The indicator function The specific form is:

[0033] ;

[0034] In the formula, For the sign function, the positive distance is 1, the negative distance is -1, and the zero distance is 0.

[0035] Therefore, this invention proposes a seasonal adaptive integration and dynamic anomaly correction temperature prediction system and method, the beneficial effects of which are as follows:

[0036] (1) Effectively capture nonlinear correlations and adapt to climate warming: By using cross features and dynamic climate benchmarks, the shortcomings of traditional linear combinations and static benchmarks are solved, and the accuracy of features and the timeliness of benchmarks are improved.

[0037] (2) Better adaptable to seasonal heterogeneity and reduced prediction error: Dedicated model combination is configured for winter and summer, and residual calibration is performed in winter to solve the problem of insufficient adaptation of a single model, especially reducing the error in winter.

[0038] (3) Further enhance the visualization of assessment to support climate services: The three-level anomaly assessment is combined with multi-dimensional visualization to quantify trend changes and provide precise technical support in conjunction with real-time fine-tuning.

[0039] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0040] Figure 1 This is a flowchart illustrating the implementation of a seasonal adaptive integration and dynamic anomaly correction temperature prediction system and method according to the present invention.

[0041] Figure 2 This is a conceptual model diagram of a seasonal adaptive integrated and dynamic anomaly correction temperature prediction system and method according to the present invention;

[0042] Figure 3 This is a monthly temperature anomaly comparison chart of the test year for a temperature prediction system and method with seasonal adaptive integration and dynamic anomaly correction according to the present invention.

[0043] Figure 4 This is a temperature anomaly heatmap of a seasonal adaptive integrated and dynamic anomaly correction temperature prediction system and method of the present invention, wherein (a) is a monthly temperature anomaly heatmap comparison chart for 2023 and (b) is a monthly temperature anomaly heatmap comparison chart for 2024. Detailed Implementation

[0044] To make the technical solutions, advantages, and objectives of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below. The described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the protection scope of this application.

[0045] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0046] like Figures 1-4 As shown, this invention provides a seasonal adaptive integration and dynamic anomaly correction temperature prediction system and method.

[0047] A seasonally adaptive integrated and dynamic anomaly correction temperature forecasting system includes: a data acquisition module, a modeling and calculation module, and an evaluation and visualization module;

[0048] The data acquisition module is used to integrate meteorological station observation data, reanalysis data and satellite remote sensing data, and to perform spatiotemporal alignment and standardization processing;

[0049] The modeling and computation module is used to construct time codes, lag features and cross features, generate dynamic climate benchmarks, and carry out integrated modeling and specialized calibration based on the seasonal adaptive Stacking integration framework.

[0050] The evaluation visualization module is used to construct a three-level dynamic anomaly correction system, calculate the prediction success rate and output multi-dimensional visualization results, and establish a real-time feedback system to trigger model fine-tuning.

[0051] A seasonally adaptive integrated and dynamic anomaly correction method for temperature forecasting includes:

[0052] S1. Cross-features of meteorological elements and generation of dynamic climate benchmarks: Multi-source data fusion and feature construction are carried out. Meteorological station observation data, reanalysis data and satellite remote sensing data are integrated and spatiotemporally aligned. Time coding, lag features and cross features are constructed. At the same time, dynamic climate benchmark modeling is designed and dynamic climate benchmarks are generated by using variable weight adjustment methods.

[0053] The multi-source data fusion and feature construction specifically includes: collecting long-term meteorological data including temperature, wind speed, soil temperature, sea surface temperature, and snow cover; after standardization, generating periodic features, lag features, and climate anomaly features; among which, periodic features and lag features are used to characterize the annual cycle variation of temperature data; the constructed multi-source data fusion and feature also includes optional components such as soil temperature difference features, 7-day moving average of snow cover, and quarterly interaction features, and the half-life of the rolling time window can be adjusted, set to 90 days in rapidly warming regions.

[0054] The dynamic climate baseline modeling and cross-feature design specifically include: updating the climate baseline using a sliding window and automatically generating seasonal adjustment coefficients through land-sea thermal difference analysis; adding cross-features of wind speed and soil temperature difference in winter to capture the cold wave effect, and adding cross-features of sea surface temperature and annual cycle in summer to enhance the high temperature prediction capability.

[0055] S2. Seasonal adaptive ensemble modeling and special calibration are carried out using deep learning fusion algorithms: a heterogeneous model of the base layer is built based on the seasonal adaptive Stacking ensemble framework. The winter model uses a combination of Ridge regression and LightGBM, the summer model uses a combination of SVR and multinomial regression, and the transition season uses weighted output. Special calibration for winter is carried out through separate residual modeling.

[0056] The base layer heterogeneous model, constructed based on the seasonal adaptive Stacking ensemble framework, includes four models: Ridge regression, LightGBM, SVR, and multinomial regression. The weighted output weights for the transitional season are determined inversely by the historical error of the same period.

[0057] In winter, a combination of Ridge regression and LightGBM was used to fuse the winter sub-model. Specifically, the winter sub-model fusion included: Ridge regression focusing on optimizing dominant winter features such as soil temperature and snow cover, employing L2 regularization with a regularization strength parameter set to 0.5; and LightGBM using a score loss function, the specific form of which is:

[0058] ;

[0059] In the formula, For predicted values Compared with the true value The matching score, and when If the value is greater than 0.5, the prediction result is considered valid.

[0060] In summer, a combination of SVR and multinomial regression was used to fuse the summer sub-model. Specifically, the summer sub-model fusion included: using an RBF kernel tailored to sea surface temperature and solar radiation characteristics for SVR, setting the penalty coefficient C of the kernel parameters to 1.5, and setting a threshold for the kernel parameter insensitivity loss function. Set to 0.1; the polynomial regression uses a second-order method to capture the quadratic relationship between temperature and sunshine duration. The input features of the polynomial regression are the coupled cross features of summer sea surface temperature and annual cycle.

[0061] The separate residual modeling for winter-specific calibration specifically includes: adopting a two-stage bias correction strategy, first calculating the validation set residuals, then establishing a seasonally specific calibration model, using linear regression with L2 regularization for winter calibration, and dynamically updating the calibration parameters through a rolling 3-year time window with the window weight configured as exponential decay.

[0062] The objective function for linear regression with L2 regularization is:

[0063] ;

[0064] In the formula, To verify the set residual, i.e., the difference between observed and predicted values, The feature is a one-hot encoding based on the month. The regularization coefficient is . This represents the number of samples within the rolling time window.

[0065] The exponential decay calculation formula for the sliding window weight configuration of the winter sub-model is as follows:

[0066] ;

[0067] In the formula, T is the current time, t is the historical moment within the rolling time window, τ is the half-life, and α controls the decay rate.

[0068] S3, Three-level dynamic anomaly assessment and visualization output: Construct a three-level dynamic anomaly correction system, calculate anomaly values ​​in conjunction with dynamic climate benchmarks, assess prediction success rate, output prediction results through a multi-dimensional visualization engine, and establish a real-time feedback system to trigger model fine-tuning and expand the training set time range when the prediction matching rate is low.

[0069] The three-tiered dynamic anomaly assessment system specifically includes a dynamic monthly climate baseline based on a sliding 20-year window, a correction term incorporating interannual warming trends, and a real-time monitoring module for quarterly sign-matching rates. The sign-matching rate is calculated based on the consistency of signs between observed and predicted anomalies. The multi-dimensional visualization engine outputs a seasonal trend parallel coordinate plot and anomaly heatmap. The seasonal trend parallel coordinate plot is used to compare predicted and observed anomalies, while the anomaly heatmap is used to display the spatial distribution characteristics of temperature anomalies. The formula for calculating the prediction success rate is as follows:

[0070] ;

[0071] In the formula, N represents the total number of evaluation samples, i.e., the number of observed anomaly-predicted anomaly data pairs used in the prediction success rate calculation, specifically the time series sample size corresponding to the dynamic climate baseline. To observe the anomaly, To predict anomalies, For indicator functions;

[0072] Indicator Function The specific form is:

[0073] ;

[0074] In the formula, For the sign function, the positive distance is 1, the negative distance is -1, and the zero distance is 0.

[0075] Example 1

[0076] like Figures 1-4As shown, the present invention uses a seasonal adaptive integration and dynamic anomaly correction temperature prediction system and method, and conducts practical application tests in temperature prediction in East China.

[0077] The system first loads the data, standardizes it, and then constructs necessary features such as time encoding, cross features, and lag features to generate a 42-dimensional feature vector. After time series cross-validation, a seasonal adaptive ensemble model is trained. Required parameters include: periodic coefficient and calibration model. Optional components include: soil temperature difference features, 7-day moving average of snow cover, and quarterly interaction term features.

[0078] The specific implementation steps are as follows:

[0079] Data preparation and feature generation: Meteorological data from East China from 2000 to 2023 were loaded, covering observations from 50 meteorological stations, ERA5 reanalysis, and MODIS satellite remote sensing data. After standardization, time-coded, periodic, lag, and winter-summer cross-features were constructed to generate a 42-dimensional feature vector. Optional components enabled soil temperature difference and 7-day moving average snow cover features, with the dynamic baseline half-life set to 120 days.

[0080] Model training and calibration: Time-series cross-validation was employed, with the test set from 2014 to 2024 and the training set from 2000 to 2013. A seasonally adaptive stacking ensemble model was trained, with necessary parameters including the periodicity coefficient. The calibration model regularization coefficient was set to 0.1. The winter sub-model was fused using a combination of Ridge regression and LightGBM. Specifically, Ridge regression focused on optimizing dominant winter features such as soil temperature and snow cover, employing L2 regularization with a regularization strength parameter of 0.5. LightGBM used a fractional loss function with a target fraction parameter of 0.5 to enhance the nonlinear expression and robust fitting of lagged features. Figure 3 The monthly average forecast bias was compared before and after calibration. Overall, the forecast bias was significantly reduced after calibration, especially in July.

[0081] In summer, a combination of SVR and multinomial regression was used to fuse the summer sub-model. The summer sub-model fusion specifically included: SVR using an RBF kernel targeting sea surface temperature and solar radiation characteristics, with the penalty coefficient C of the kernel parameter set to 1.5 and the threshold ε of the kernel parameter insensitivity loss function set to 0.1; and multinomial regression using a second-order method to capture the quadratic relationship between temperature and sunshine duration, with features including sea surface temperature.

[0082] Prediction and Evaluation: Outputs monthly temperature forecast sequences and anomalies from 2013 to 2024. Level 3 evaluation shows: overall anomaly sign accuracy for the test set is 81.8%, and monthly anomaly sign accuracy is 66.7%; April and September show the best accuracy at 73%, compared to... Figure 4 The monthly temperature forecast series and anomaly values ​​for 2023-2024 are generally consistent.

[0083] Therefore, this invention provides a temperature prediction system and method with seasonal adaptive integration and dynamic anomaly correction. By constructing cross features, dynamic climate benchmarks and seasonal adaptive integrated modeling, it effectively solves the problems of insufficient nonlinear capture, benchmark rigidity and poor seasonal adaptation in existing technologies, significantly improves the accuracy and seasonal adaptability of temperature prediction, and enhances the quantification and visualization of anomaly assessment, providing reliable technical support for short-term meteorological climate services.

[0084] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A seasonally adaptive integrated temperature forecasting system with dynamic anomaly correction, characterized in that, include: Data acquisition module, modeling and calculation module, and evaluation and visualization module; The data acquisition module is used to integrate meteorological station observation data, reanalysis data and satellite remote sensing data, and perform spatiotemporal alignment and standardization processing; The modeling and calculation module is used to construct time codes, lag features and cross features. At the same time, it designs dynamic climate benchmark modeling, uses a variable weight adjustment method to generate dynamic climate benchmark, and constructs a heterogeneous model of the basic layer based on the seasonal adaptive Stacking integration framework. The winter model uses a combination of Ridge regression and LightGBM, the summer model uses a combination of SVR and multinomial regression, and the transition season uses weighted output. Furthermore, the winter model is specifically calibrated through separate residual modeling. The separate residual modeling for winter-specific calibration specifically includes: adopting a two-stage bias correction strategy, first calculating the validation set residuals, then establishing a seasonally specific calibration model, using linear regression with L2 regularization for winter calibration, and dynamically updating the calibration parameters through a rolling 3-year time window with the window weight configured as exponential decay. The evaluation visualization module is used to construct a three-level dynamic anomaly correction system, calculate the prediction success rate and output multi-dimensional visualization results, and establish a real-time feedback system to trigger model fine-tuning. The three-level dynamic anomaly assessment system specifically includes a dynamic monthly climate benchmark based on a sliding 20-year window, a correction term incorporating interannual warming trends, and a real-time monitoring module for quarterly sign matching rate. The sign matching rate is calculated by the consistency of the signs between observed and predicted anomalies. The multi-dimensional visualization engine outputs a seasonal trend parallel coordinate plot and anomaly heatmap. The seasonal trend parallel coordinate plot is used to compare predicted and observed anomalies, while the anomaly heatmap is used to display the spatial distribution characteristics of temperature anomalies. The system also includes dynamic climate baseline modeling and seasonal-specific cross-feature design. Specifically, dynamic climate baseline modeling and cross-feature design include: updating the climate baseline using a sliding window and automatically generating seasonal adjustment coefficients through land-sea thermal difference analysis; adding cross-features of wind speed and soil temperature difference in winter to capture the cold wave effect; and adding cross-features of sea surface temperature and annual cycle coupling in summer to enhance high temperature prediction capabilities.

2. A seasonal adaptive integrated and dynamic anomaly correction method for temperature forecasting, characterized in that, include: S1. Cross-features of meteorological elements and generation of dynamic climate benchmarks: Multi-source data fusion and feature construction are carried out. Meteorological station observation data, reanalysis data and satellite remote sensing data are integrated and spatiotemporally aligned. Time coding, lag features and cross features are constructed. At the same time, dynamic climate benchmark modeling is designed and dynamic climate benchmarks are generated by using variable weight adjustment methods. The dynamic climate baseline modeling and cross-feature design specifically include: updating the climate baseline using a sliding window and automatically generating seasonal adjustment coefficients through land-sea thermal difference analysis; adding cross-features of wind speed and soil temperature difference in winter to capture the cold wave effect, and adding cross-features of sea surface temperature and annual cycle in summer to enhance the high temperature prediction capability. S2. Seasonal adaptive ensemble modeling and special calibration are carried out using deep learning fusion algorithms: a heterogeneous model of the base layer is built based on the seasonal adaptive Stacking ensemble framework. The winter model uses a combination of Ridge regression and LightGBM, the summer model uses a combination of SVR and multinomial regression, and the transition season uses weighted output. Special calibration for winter is carried out through separate residual modeling. The separate residual modeling for winter-specific calibration specifically includes: adopting a two-stage bias correction strategy, first calculating the validation set residuals, then establishing a seasonally specific calibration model, using linear regression with L2 regularization for winter calibration, and dynamically updating the calibration parameters through a rolling 3-year time window with the window weight configured as exponential decay. S3, Three-level dynamic anomaly assessment and visualization output: Construct a three-level dynamic anomaly correction system, calculate anomaly values ​​in conjunction with dynamic climate benchmarks, assess prediction success rate, output prediction results through a multi-dimensional visualization engine, and establish a real-time feedback system to trigger model fine-tuning when the prediction matching rate is low and expand the training set time range. The three-level dynamic anomaly assessment system specifically includes a dynamic monthly climate benchmark based on a sliding 20-year window, a correction term incorporating interannual warming trends, and a real-time monitoring module for quarterly sign matching rate. The sign matching rate is calculated by the consistency of the signs between observed and predicted anomalies. The multi-dimensional visualization engine outputs a seasonal trend parallel coordinate plot and anomaly heatmap. The seasonal trend parallel coordinate plot is used to compare predicted and observed anomalies, while the anomaly heatmap is used to display the spatial distribution characteristics of temperature anomalies.

3. The temperature prediction method based on seasonal adaptive integration and dynamic anomaly correction according to claim 2, characterized in that, In S1, the multi-source data fusion and feature construction specifically includes: collecting long-term meteorological data including temperature, wind speed, soil temperature, sea surface temperature, and snow cover; after standardization, generating periodic features, lag features, and climate anomaly features; among which, periodic features and lag features are used to characterize the annual cycle variation of temperature data; the constructed multi-source data fusion and feature also includes optional components such as soil temperature difference features, 7-day moving average of snow cover, and quarterly interaction features, and the half-life of the rolling time window can be adjusted, set to 90 days in rapidly warming regions.

4. The temperature prediction method based on seasonal adaptive integration and dynamic anomaly correction according to claim 2, characterized in that, In S2, the heterogeneous models of the base layer are constructed based on the seasonal adaptive Stacking ensemble framework, including four models: Ridge regression, LightGBM, SVR, and multinomial regression. The weighted output weights for the transition season are determined by the inverse of the historical same-period error.

5. The temperature prediction method based on seasonal adaptive integration and dynamic anomaly correction according to claim 2, characterized in that, In S2, the winter sub-model is fused using a combination of Ridge regression and LightGBM. Specifically, the winter sub-model fusion includes: Ridge regression focusing on optimizing dominant winter features such as soil temperature and snow cover, employing L2 regularization with a regularization strength parameter set to 0.5; and LightGBM using a score loss function, the specific form of which is: ； In the formula, For predicted values Compared with the true value The matching score, and when If the value is greater than 0.5, the prediction result is considered valid.

6. The temperature prediction method based on seasonal adaptive integration and dynamic anomaly correction according to claim 2, characterized in that, In S2, the summer sub-model is fused using a combination of SVR and multinomial regression. Specifically, the summer sub-model fusion includes: SVR employing an RBF kernel tailored to sea surface temperature and solar radiation characteristics, with the kernel parameter penalty coefficient C set to 1.5, and a threshold for the kernel parameter insensitive loss function. Set to 0.1; the polynomial regression uses a second-order method to capture the quadratic relationship between temperature and sunshine duration. The input features of the polynomial regression are the coupled cross features of summer sea surface temperature and annual cycle.

7. The temperature prediction method based on seasonal adaptive integration and dynamic anomaly correction according to claim 2, characterized in that, The objective function of the L2-regularized linear regression is: ； In the formula, To verify the set residual, i.e., the difference between observed and predicted values, The feature is a one-hot encoding based on the month. The regularization coefficient is . This represents the number of samples within the rolling time window. The exponential decay calculation formula for the sliding window weight configuration of the winter sub-model is as follows: ； In the formula, T is the current time, t is the historical moment within the rolling time window, τ is the half-life, and α controls the decay rate.

8. The temperature prediction method based on seasonal adaptive integration and dynamic anomaly correction according to claim 2, characterized in that, The formula for calculating the prediction success rate is as follows: ； In the formula, N represents the total number of evaluation samples, i.e., the number of observed anomaly-predicted anomaly data pairs used in the prediction success rate calculation, specifically the time series sample size corresponding to the dynamic climate baseline. To observe the anomaly, To predict anomalies, For indicator functions; The indicator function The specific form is as follows: ； In the formula, For the sign function, the positive distance is 1, the negative distance is -1, and the zero distance is 0.

Images