A dynamic-statistical prediction method and system for summer seasonal precipitation and air temperature

Abstract

This invention relates to the field of meteorological forecasting technology, specifically a dynamic-statistical forecasting method and system for summer pentad precipitation and temperature. The method includes: acquiring reanalysis gridded real-time data, historical return circulation factors, and future predicted circulation factors for each pentad in a target area over a historical period; using singular value decomposition (SVD) to screen for the optimal predictor factors based on the reanalysis gridded real-time data and constructing a dynamic-statistical forecasting model; constructing a correction model based on the historical return circulation factors and the optimal predictor factors; using the correction model to correct the future predicted circulation factors to obtain the corrected future predicted circulation factors; and inputting the corrected future predicted circulation factors into the dynamic-statistical forecasting model to obtain the final forecast results for summer pentad precipitation and temperature. This invention solves the problem that existing dynamic-statistical forecasting methods for summer pentad precipitation and temperature are not scientifically sound and accurate. The correction of the optimal predictor factors and predicted circulation factors greatly improves the scientific validity and accuracy of the forecasts.

Description

Technical Field

[0001] This invention relates to the field of meteorological forecasting technology, specifically a dynamic-statistical forecasting method and system for summer pentad precipitation and temperature. Background Technology

[0002] Accurate forecasting of summer precipitation and temperature is a crucial aspect of meteorological operations, especially in core agricultural areas like river basins. The pentad-by-pentad changes in meteorological elements directly impact agricultural production, flood control, and drought relief decisions, demanding higher forecast accuracy. Currently, relevant forecasting methods are mainly divided into two categories: dynamic model forecasting and statistical model forecasting, both of which have significant technical limitations. While dynamic models such as CPSv3 can depict the physical evolution of atmospheric circulation, their output circulation factors suffer from systematic biases and spatial distribution errors, leading to a significant decrease in forecast accuracy at the extended pentad scale (10-60 days) as forecast lead time increases. Traditional statistical models often rely on empirical selection of circulation factors, lacking robust quantitative verification of factor predictive capabilities, making them prone to overfitting and exhibiting poor generalization ability. Existing dynamic-statistical fusion methods are mostly simple weighted summaries of prediction results, failing to fully explore the coupled modal statistical relationships between atmospheric circulation factors and precipitation and temperature elements. Furthermore, they lack effective corrections for systematic biases in dynamic model circulation factors, making it difficult to balance the physical rationality and statistical accuracy of predictions at the pentad scale. Consequently, they cannot meet the actual operational needs for high-precision predictions of summer pentad precipitation and temperature.

[0003] There is an urgent need for a more scientific and accurate dynamic-statistical forecasting method for summer pentad precipitation and temperature. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a dynamic-statistical prediction method and system for summer pentad precipitation and temperature, solving the problem that existing dynamic-statistical prediction methods for summer pentad precipitation and temperature are not scientific or accurate enough.

[0005] To achieve the above objectives, this invention provides a dynamic-statistical prediction method for summer pentad precipitation and temperature. The method includes: acquiring reanalysis gridded real-time data, historical return circulation factors, and future predicted circulation factors for each pentad in a target area over a historical period; using singular value decomposition to screen the optimal predictor factors and construct a dynamic-statistical prediction model based on the reanalysis gridded real-time data; constructing a correction model based on the historical return circulation factors and the optimal predictor factors; correcting the future predicted circulation factors using the correction model to obtain corrected future predicted circulation factors; and inputting the corrected future predicted circulation factors into the dynamic-statistical prediction model to obtain the final prediction results for summer pentad precipitation and temperature.

[0006] This invention acquires multi-source meteorological data for each pentad of a target area, combines singular value decomposition (SVD) to screen optimal predictive factors, and constructs a dynamic-statistical prediction model. It uncovers the coupling patterns between circulation factors and precipitation and temperature, improving the scientific rigor of factor selection and the statistical rationality of the model. By constructing a correction model using historical return circulation factors and the optimal predictive factors, it provides precise model support for correcting circulation factor biases. The correction model corrects future predicted circulation factors, eliminating systematic biases in the dynamic model and improving the accuracy of circulation factors. The corrected circulation factors are then input into the prediction model to obtain the final results. This invention integrates the advantages of dynamic physics and statistical correction, significantly improving the accuracy and scientific rigor of summer pentad precipitation and temperature predictions.

[0007] Optionally, the step of using singular value decomposition to screen the optimal predictor factors and construct a dynamic-statistical prediction model based on the reanalysis grid data includes: sequentially extracting the actual circulation factor, actual precipitation field, actual temperature field, and the actual precipitation field and actual temperature field of the validation year from the reanalysis grid data using the leave-one-out method; performing singular value decomposition on the actual circulation factor, the actual precipitation field, and the actual temperature field and then screening them to obtain the optimized left-field spatial mode, optimized left-field time coefficient, optimized right-field spatial mode, and optimized right-field time coefficient corresponding to precipitation and temperature, respectively. Coefficients; based on the optimized left-field spatial mode, the optimized left-field time coefficient, the optimized right-field spatial mode, and the optimized right-field time coefficient, the precipitation prediction field and temperature prediction field corresponding to the actual circulation factor are obtained; according to the precipitation prediction field, the temperature prediction field, the actual precipitation field of the verification year, and the actual temperature field of the verification year, multiple optimal prediction factors are selected from the actual circulation factors; based on the optimal prediction factors, single-factor dynamic-statistical sub-models are constructed respectively; based on multiple single-factor dynamic-statistical sub-models, a dynamic-statistical prediction model is constructed.

[0008] This invention employs a leave-one-out method to iteratively extract the element fields of the training and validation sets, achieving independent training and validation of the entire historical sample. This avoids the problems of insufficient sample utilization and inadequate model generalization caused by fixed training and validation periods. It makes the results of factor selection and model construction more closely aligned with actual business prediction scenarios. By performing singular value decomposition on the real-world element fields of the training set, it accurately uncovers the coupling modal patterns of circulation factors with precipitation and temperature, laying a scientific physical and statistical foundation for predictive field reconstruction. Based on the coupling modes, it obtains the precipitation and temperature prediction fields corresponding to the validation year, and selects the optimal predictive factors based on the validation year's real-world field. This makes the basis for factor selection more closely aligned with real prediction results, avoiding the one-sidedness of experience-based factor selection. Based on the optimal predictive factors, it constructs independent single-factor sub-models and then integrates them into a total model, fully integrating the predictive advantages of multiple factors. Combined with a grid-by-grid and pentad-by-pentad fusion strategy, it effectively reduces the uncertainty of single-factor modeling and significantly improves the robustness, adaptability, and global generalization ability of the dynamic-statistical prediction model.

[0009] Optionally, the step of performing singular value decomposition and filtering on the actual circulation factor, the actual precipitation field, and the actual temperature field to obtain the optimized left-field spatial mode, optimized left-field time coefficient, optimized right-field spatial mode, and optimized right-field time coefficient corresponding to precipitation and temperature includes: performing singular value decomposition with the actual circulation factor as the left field and the actual precipitation field as the right field to obtain multiple sets of first singular values ​​and first coupled mode data corresponding to precipitation; performing singular value decomposition with the actual circulation factor as the left field and the actual temperature field as the right field to obtain multiple sets of second singular values ​​and second coupled mode data corresponding to temperature; and filtering with the first singular value and the second singular value as the target to obtain multiple sets of optimized left-field spatial mode, optimized left-field time coefficient, optimized right-field spatial mode, and optimized right-field time coefficient corresponding to precipitation and temperature.

[0010] This invention performs singular value decomposition (SVD) with the actual circulation factor as the left field and the precipitation and temperature fields as the right fields, respectively. This accurately separates the coupling modes and singular values ​​corresponding to precipitation and temperature, enabling targeted mining of the coupling relationship between circulation factors and hydrothermal elements. By selecting and optimizing modes and coefficients with the maximum singular value as the objective, the core coupling features with the highest contribution are selected, and invalid and redundant information is eliminated. This improves the accuracy and effectiveness of mode selection, avoids interference from low-contribution modes in subsequent modeling, and enhances the pertinence and scientific nature of mining the coupling laws between circulation and precipitation and temperature.

[0011] Optionally, obtaining the precipitation prediction field and temperature prediction field corresponding to the actual circulation factor based on the optimized left-field spatial mode, the optimized left-field time coefficient, the optimized right-field spatial mode, and the optimized right-field time coefficient includes: performing linear fitting on the optimized left-field time coefficient and the optimized right-field time coefficient to obtain multiple linear regression relationships corresponding to the precipitation and the temperature, respectively; calculating the predicted right-field time coefficients corresponding to the precipitation and the temperature based on the linear regression relationships and using the optimized left-field time coefficients, respectively; and reconstructing the precipitation prediction field and temperature prediction field corresponding to the actual circulation factor based on the predicted right-field time coefficients and the optimized right-field spatial mode.

[0012] This invention establishes a quantitative statistical relationship between circulation and precipitation and temperature by linearly fitting the time coefficients of the optimized left and right fields, clarifying the linear correlation law between the elements. Based on this regression relationship, the predicted time coefficient of the right field is calculated, realizing the accurate deduction from circulation factors to the time characteristics of hydrothermal elements. By combining the predicted time coefficient with the optimized right field spatial mode to reconstruct the predicted field, the spatiotemporal characteristics are fused and restored, accurately restoring the spatial distribution characteristics of precipitation and temperature, and improving the accuracy and rationality of the predicted field reconstruction.

[0013] Optionally, the step of selecting multiple optimal prediction factors from the actual circulation factors based on the precipitation prediction field, the temperature prediction field, the actual precipitation field of the verification year, and the actual temperature field of the verification year includes: obtaining a comprehensive score of the actual circulation factors based on the actual precipitation field of the verification year, the precipitation prediction field, the actual temperature field of the verification year, and the temperature prediction field; and selecting multiple optimal prediction factors from the actual circulation factors based on the comprehensive score.

[0014] This invention uses the verified year's actual precipitation and temperature fields obtained through leave-one-out iteration as a benchmark, and combines them with the corresponding predicted fields generated by the model to conduct factor screening. This ensures that the data source for factor evaluation is highly consistent with the actual prediction scenario, effectively reflecting the true predictive ability of each actual circulation factor. By comparing the verified year's actual field with the predicted field, a comprehensive score of the actual circulation factor is obtained, realizing a quantitative assessment of the predictive ability of factors in both precipitation and temperature dimensions. This avoids the one-sidedness of single-dimensional evaluation. Based on the comprehensive score, the optimal factor is selected from the actual circulation factors. The quantitative indicators replace the traditional experience-based selection, improving the objectivity and scientific nature of factor screening.

[0015] Optionally, obtaining the comprehensive score of the actual circulation factor based on the actual precipitation field, the predicted precipitation field, the actual temperature field, and the predicted temperature field of the verification year includes: calculating a first root mean square error, a first time correlation coefficient, and a first anomaly correlation coefficient based on the actual precipitation field and the predicted precipitation field of the verification year; calculating a precipitation dimension score of the actual circulation factor based on the first root mean square error, the first time correlation coefficient, and the first anomaly correlation coefficient; calculating a second root mean square error, a second time correlation coefficient, and a second anomaly correlation coefficient based on the actual temperature field and the predicted temperature field of the verification year; calculating a temperature dimension score of the actual circulation factor based on the second root mean square error, the second time correlation coefficient, and the second anomaly correlation coefficient; and forming a comprehensive score based on the precipitation dimension score and the temperature dimension score.

[0016] This invention evaluates the predictive capabilities of precipitation and temperature using three indicators: root mean square error, time correlation coefficient, and anomaly correlation coefficient. This unifies the dual-element evaluation system, eliminating scoring biases caused by inconsistencies in evaluation dimensions across different elements. By quantifying numerical deviation through root mean square error, temporal consistency through time correlation coefficient, and spatial anomaly correlation coefficient, the predictive capability of circulation factors is comprehensively assessed from multiple dimensions, avoiding the limitations of single-indicator evaluation. Precipitation and temperature dimension scores are calculated separately based on the three types of indicators, and then combined with the dual-dimensional scores to form a comprehensive score, thus improving the scientific quantification of the overall predictive capability of circulation factors.

[0017] Optionally, the step of calculating the first root mean square error, the first time correlation coefficient, and the first anomalous correlation coefficient based on the actual precipitation field of the verification year and the precipitation prediction field includes: calculating the first single-year root mean square error, the first single-year time correlation coefficient, and the first single-year anomalous correlation coefficient based on the actual precipitation field of the verification year and the corresponding precipitation prediction field for each year; and calculating the average values ​​of the first single-year root mean square error, the first single-year time correlation coefficient, and the first single-year anomalous correlation coefficient to obtain the first root mean square error, the first time correlation coefficient, and the first anomalous correlation coefficient.

[0018] This invention combines the leave-one-out iterative verification logic to calculate the annual precipitation evaluation index for each verification year, achieving independent full-sample verification of each actual circulation factor. This avoids the random interference of data in a fixed verification period, allowing the index calculation to better reflect the true predictive performance of the factors. By calculating the arithmetic mean of the root mean square error, time correlation coefficient, and anomaly correlation coefficient for each verification year, the index results reflecting the global precipitation prediction capability of the factors are obtained. This effectively eliminates the index fluctuations caused by annual climate anomalies and differences in prediction difficulty, making the final global index more robust and objective. This calculation method abandons the traditional precipitation-specific scoring and maintains consistency with the index calculation logic of the temperature dimension, achieving synergy in the dual-element index calculation system and improving the scientificity and rigor of the entire factor evaluation system.

[0019] Optionally, constructing a dynamic-statistical prediction model based on multiple single-factor dynamic-statistical sub-models includes: combining the single-factor dynamic-statistical sub-models; and based on the result of the combination, introducing a grid-by-grid, event-by-event arithmetic averaging strategy to form a dynamic-statistical prediction model.

[0020] This invention combines multiple single-factor dynamic statistical sub-models, fully integrates the predictive advantages of different optimal predictors, reduces the uncertainty of single-factor modeling, introduces a grid-by-grid and event-by-event arithmetic averaging strategy, achieves smooth optimization on spatial and temporal scales, reduces the impact of local anomalies and random errors, and improves the stability, robustness and generalization ability of the dynamic statistical prediction model.

[0021] Optionally, the step of constructing a correction model based on the historical return circulation factor and the optimal predictor factor includes: extracting the actual two-dimensional spatial grid field of each of the optimal predictor factors for each epoch; extracting the return two-dimensional spatial grid field from the historical return circulation factor that has the same time, the same elements, and the same spatial range as the optimal predictor factor; constructing correction training samples using the return two-dimensional spatial grid field as input features and the actual two-dimensional spatial grid field as output features; and training the correction model based on U-Net using the correction training samples.

[0022] This invention constructs spatiotemporally consistent correction training samples by extracting the real field of the optimal predictor and the matching historical return circulation factor field, ensuring sample quality and matching accuracy. Using the return field as input and the real field as output training samples, the learning objective of bias correction is clearly defined. Based on U-Net, a correction model is constructed and trained, making full use of its spatial feature extraction advantages to accurately learn the systematic bias law of dynamic patterns, thereby improving the overall accuracy and spatial consistency of circulation factor bias correction.

[0023] Another aspect of the present invention provides a dynamic-statistical prediction system for summer pentad precipitation and temperature, comprising: a processor, an input device, an output device, and a memory, wherein the processor, the input device, the output device, and the memory are interconnected, wherein the memory is used to store a computer program, the computer program including program instructions, and the processor is configured to call the program instructions to execute the dynamic-statistical prediction method for summer pentad precipitation and temperature as described in any of the preceding aspects of the present invention.

[0024] The present invention provides a dynamic-statistical prediction system for summer pentad precipitation and temperature, which is compact in structure, stable in performance, highly integrated and simple in composition. It can stably execute the dynamic-statistical prediction method for summer pentad precipitation and temperature provided in the preceding part of the present invention, further improving the overall applicability and practical application capability of the present invention. Attached Figure Description

[0025] Figure 1 This is a flowchart of a dynamic-statistical prediction method for summer pentad precipitation and temperature according to an embodiment of the present invention; Figure 2 This is a schematic diagram of a dynamic-statistical prediction system for summer precipitation and temperature according to an embodiment of the present invention. Figure 3 The RMSE score for predicting the average temperature of the next six pentads 11-20 days in advance, as per the embodiments of the present invention; Figure 4 The RMSE score for predicting the average temperature of the next three pentads 11-40 days in advance, as per the embodiments of the present invention; Figure 5 The TCC score for predicting the average temperature of the next six pentads 11-20 days in advance, as per the embodiments of the present invention; Figure 6 The TCC score for predicting the average temperature of the next three pentads 11-40 days in advance, as per the embodiments of the present invention; Figure 7 The ACC score for predicting the average temperature of the next six pentads 11-20 days in advance, as per the embodiments of the present invention; Figure 8 The ACC score for predicting the average temperature of the next three pentads 11-40 days in advance, as per the embodiments of the present invention; Figure 9 The RMSE score for predicting precipitation in the next six pentads 11-20 days in advance, as per the embodiments of the present invention; Figure 10 The RMSE score for predicting precipitation in the next three pentads 11-40 days in advance, as per the embodiments of the present invention; Figure 11 The TCC score for predicting precipitation 11-20 days in advance for the next six pentads is used in this embodiment of the invention. Figure 12 The TCC score for predicting precipitation in the next three pentads 11-40 days in advance, as per the embodiments of the present invention; Figure 13 The ACC score for predicting precipitation in the next six pentads 11-20 days in advance, as per the embodiments of the present invention; Figure 14 The ACC score for predicting precipitation in the next three pentads 11-40 days in advance is used as an example of the present invention. Detailed Implementation

[0026] Specific embodiments of the present invention will now be described in detail. It should be noted that the embodiments described herein are for illustrative purposes only and are not intended to limit the invention. In the following description, numerous specific details are set forth in order to provide a thorough understanding of the invention. However, it will be apparent to those skilled in the art that these specific details are not necessary to practice the invention. In other instances, well-known circuits, software, or methods have not been specifically described to avoid obscuring the invention.

[0027] Throughout this specification, references to "an embodiment," "an embodiment," "an example," or "an example" mean that a particular feature, structure, or characteristic described in connection with that embodiment or example is included in at least one embodiment of the invention. Therefore, the phrases "in an embodiment," "in an embodiment," "an example," or "an example" appearing in various places throughout the specification do not necessarily refer to the same embodiment or example. Furthermore, specific features, structures, or characteristics can be combined in one or more embodiments or examples in any suitable combination and / or sub-combination. Moreover, those skilled in the art will understand that the illustrations provided herein are for illustrative purposes and are not necessarily drawn to scale.

[0028] Please see Figure 1 In order to solve the problems in the prior art, in an alternative embodiment, such as Figure 1 The dynamic-statistical prediction method for summer pentad precipitation and temperature shown includes the following steps: Step S1: Obtain real-time grid data of each pentad in the target area during historical periods, historical return circulation factors, and future predicted circulation factors.

[0029] In this embodiment, for the target area, reanalysis gridded real-time data, historical return circulation factors from climate models, and future predicted circulation factors for each year from May to September, 1991 to 2023, were collected. All raw daily resolution data were processed into pentad resolution data. The reanalysis gridded real-time data included two types of data sources: CRA40 and NOAA (after converting the CRA40 and NOAA reanalysis gridded real-time data to pentad resolution, a total of 990 pentads of valid data were obtained for each year from May to September, 1991 to 2023). The elements covered included precipitation, average temperature, geopotential height of different pressure layers, zonal wind, meridional wind, specific humidity, OLR, etc. Each element was matched with a corresponding key spatial range and a spatial resolution of 0.25° (spatial range: watershed 38°-55°N, 115°-136°E, spatial resolution 0.25°), geopotential height (20 The key areas selected for circulation factors are 0 hPa, 500 hPa, and 850 hPa (15°-80°N, 20°-160°E), zonal winds (200 hPa and 850 hPa, 10°S-70°N, 20°-160°E), meridional winds (200 hPa and 850 hPa, 10°S-70°N, 20°-160°E), and specific humidity (700 hPa, 20°S-50°N, 50°-170°E). Both historical and future predicted circulation factors are derived from the CPSv3 climate model with a spatial resolution of 0.45° (corresponding to T266 Gaussian grid points). The future predicted circulation factors are updated daily with a prediction lead time of 60 days. The elements of both types of circulation factors are consistent with the circulation elements used for prediction in the reanalysis grid data and match the corresponding key spatial ranges.

[0030] Further analysis revealed that all elements of the gridded real-time data (real-time circulation factors, precipitation field, and temperature field) had a spatial resolution of 0.25°. The original resolution of the historical returns and future prediction circulation factors of the CPSv3 model was 0.45°. These were interpolated to 0.25° using bilinear interpolation, achieving resolution consistency with the real-time data. The interpolation maintained the spatial range and pentad time resolution of the elements, ensuring meteorological and physical significance and spatial continuity.

[0031] The historical return circulation factor is derived from historical backcalculation data of the CPSv3 climate model, while the future predicted circulation factor is derived from real-time prediction data of the CPSv3 model, updated daily, with a prediction lead time of 60 days. The elements are the same as the historical return circulation factor, and the spatial resolution is consistent.

[0032] Step S2: Based on the reanalysis grid data, the optimal predictor is selected using singular value decomposition and a dynamic-statistical prediction model is constructed.

[0033] The process of selecting the optimal predictor based on the reanalysis grid data and constructing a dynamic-statistical prediction model using singular value decomposition includes the following sub-steps: Step S201: Using the leave-one-out method, extract the actual circulation factor, actual precipitation field, actual temperature field, and actual precipitation field of the verification year and actual temperature field of the verification year sequentially from the actual data of the reanalysis grid points.

[0034] In this embodiment, the reanalysis grid data of the target area during the historical period is divided into natural years. The total number of years of historical data is N years. The leave-one-out method is used to iteratively extract the training set and validation set feature fields from the N years of data. During the iteration process, k ranges from 1 to N. In each iteration, the data of the kth year is removed as an independent validation set. The validation year's actual precipitation field and validation year's actual temperature field are extracted from the validation set. The remaining N-1 years of data are integrated into the training set, and the actual circulation factor, actual precipitation field, and actual temperature field of the training set are extracted. All extracted feature fields maintain the same 0.25° spatial resolution, target area spatial range, and pentad time resolution as the original reanalysis grid data. The feature fields of the training set and the validation set are precisely matched at the pentad time nodes and spatial grid points. After extracting the feature field corresponding to a single k value, the next round of k value iteration is performed until all years from 1 to N are traversed, resulting in N sets of one-to-one corresponding training set feature fields (actual circulation factor, actual precipitation field, actual temperature field) and validation set feature fields (actual precipitation field of the validation year, actual temperature field of the validation year).

[0035] Step S202: Perform singular value decomposition and screening on the actual circulation factor, the actual precipitation field and the actual temperature field to obtain the optimized left field spatial mode, optimized left field time coefficient, optimized right field spatial mode and optimized right field time coefficient corresponding to precipitation and temperature, respectively.

[0036] Specifically, singular value decomposition and screening are performed on the actual circulation factor, the actual precipitation field, and the actual temperature field to obtain the optimized left-field spatial mode, optimized left-field time coefficient, optimized right-field spatial mode, and optimized right-field time coefficient corresponding to precipitation and temperature, respectively: Step S20201: Using the actual circulation factor as the left field and the actual precipitation field as the right field, singular value decomposition is performed to obtain multiple sets of first singular values ​​and first coupled mode data corresponding to the precipitation.

[0037] In this embodiment, each extracted real-time circulation factor is taken as the left field and the real-time precipitation field as the right field. Singular value decomposition analysis is performed on the two element fields. Through this decomposition process, multiple sets of first singular values ​​corresponding to the precipitation forecast are obtained, as well as first coupled mode data containing the left field spatial mode, left field time coefficient, right field spatial mode, and right field time coefficient. Each first singular value corresponds one-to-one with each set of first coupled mode data. The first coupled mode data includes the left field spatial mode, left field time coefficient, right field spatial mode, and right field time coefficient corresponding to precipitation.

[0038] Step S20202: Using the actual circulation factor as the left field and the actual temperature field as the right field, singular value decomposition is performed to obtain multiple sets of second singular values ​​and second coupled mode data corresponding to the temperature.

[0039] In this embodiment, each extracted real-time circulation factor is taken as the left field and the real-time temperature field as the right field. Singular value decomposition analysis is performed on the two element fields. Through this decomposition process, multiple sets of second singular values ​​corresponding to the temperature prediction are obtained, as well as second coupled mode data containing the left field spatial mode, left field time coefficient, right field spatial mode, and right field time coefficient. Each second singular value corresponds one-to-one with each set of second coupled mode data. The second coupled mode data includes the left field spatial mode, left field time coefficient, right field spatial mode, and right field time coefficient corresponding to the temperature.

[0040] Step S20203 involves filtering based on maximizing the first and second singular values ​​to obtain multiple optimized left-field spatial modes, optimized left-field time coefficients, optimized right-field spatial modes, and optimized right-field time coefficients corresponding to precipitation and temperature, respectively. In this embodiment, the core filtering objective is to maximize the first singular value corresponding to precipitation and the second singular value corresponding to temperature. The top 5 sets of first and second coupled mode data with the highest singular value contribution (or a cumulative variance contribution rate exceeding 80%) are selected, and the optimized left-field spatial modes, left-field time coefficients, right-field spatial modes, and right-field time coefficients corresponding to precipitation and temperature are extracted from them.

[0041] It should be noted that steps S201 to S20303 above are all performed within the framework of the leave-one-out method: for each iteration, singular value decomposition, modality screening, and linear regression fitting are performed using the current training set (year N-1), and precipitation and temperature prediction fields are generated for the current validation set (year k) based on this. By traversing all years, it is ensured that the prediction results for each year come from independent models that do not include data for that year, fundamentally avoiding overfitting and guaranteeing the objectivity and generalization ability of factor selection.

[0042] Step S203: Based on the optimized left-field spatial mode, the optimized left-field time coefficient, the optimized right-field spatial mode, and the optimized right-field time coefficient, obtain the precipitation prediction field and temperature prediction field corresponding to the actual circulation factor.

[0043] The specific steps for obtaining the precipitation prediction field and temperature prediction field corresponding to the actual circulation factor based on the optimized left-field spatial mode, the optimized left-field time coefficient, the optimized right-field spatial mode, and the optimized right-field time coefficient include the following sub-steps: Step S20301: Perform linear fitting on the optimized left field time coefficient and the optimized right field time coefficient to obtain multiple linear regression relationships corresponding to the precipitation and the temperature, respectively.

[0044] In this embodiment, the optimized left and right field time coefficients obtained from SVD decomposition are used as the full training samples. A linear fitting method is used to construct linear statistical laws corresponding to precipitation and temperature, respectively. Linear fitting operations are carried out for the two types of prediction objects, precipitation and temperature, respectively. The linear regression method is used to perform a one-to-one linear fitting between each group of optimized left field time coefficients corresponding to precipitation and the optimized right field time coefficients of the same group. At the same time, a one-to-one linear fitting is performed between each group of optimized left field time coefficients corresponding to temperature and the optimized right field time coefficients of the same group. Based on the above group fitting process, multiple linear regression relationships corresponding to precipitation and multiple linear regression relationships corresponding to temperature are obtained that match the number of optimized modal coefficient groups.

[0045] Step S20302: Based on the linear regression relationship, the predicted right field time coefficients corresponding to the precipitation and the temperature are calculated using the optimized left field time coefficient.

[0046] In this embodiment, calculations are performed separately for two types of prediction objects: precipitation and temperature. For the precipitation dimension, each set of optimized left-field time coefficients is substituted into the matching linear regression relationship for precipitation, and the calculation is completed through the regression formula to obtain the precipitation prediction right-field time coefficients corresponding to each set of coefficients. For the temperature dimension, the same calculation logic is used, and each set of optimized left-field time coefficients is substituted into the matching linear regression relationship for temperature to obtain the temperature prediction right-field time coefficients corresponding to each set of coefficients, ensuring the inter-group correspondence between the prediction right-field time coefficients and the linear regression relationship and optimized left-field time coefficients.

[0047] Step S20303: Based on the predicted right field time coefficient and the optimized right field spatial mode, the precipitation prediction field and temperature prediction field corresponding to the actual circulation factor are reconstructed.

[0048] In this embodiment, field reconstruction operations are performed separately for precipitation and temperature. First, the time coefficient of each predicted right field corresponding to precipitation is multiplied by the optimized right field spatial mode of the same group grid by grid to obtain the spatial distribution data corresponding to each precipitation mode. Then, the spatial distribution data of all precipitation modes are accumulated and integrated to obtain the precipitation prediction field corresponding to the actual circulation factor. Using the same reconstruction logic, the time coefficient of each predicted right field corresponding to temperature is multiplied by the optimized right field spatial mode of the same group grid by grid to obtain the spatial distribution data corresponding to each temperature mode. After accumulating and integrating the spatial distribution data of all temperature modes, the temperature prediction field corresponding to the actual circulation factor is obtained. The spatiotemporal resolution of the two reconstructed fields is consistent with the original element field.

[0049] Step S204: Select multiple optimal prediction factors from the actual circulation factors based on the precipitation prediction field, the temperature prediction field, the actual precipitation field of the verification year, and the actual temperature field of the verification year.

[0050] The process of selecting multiple optimal prediction factors from the actual circulation factors based on the precipitation prediction field, the temperature prediction field, the actual precipitation field of the verification year, and the actual temperature field of the verification year specifically includes the following sub-steps: Step S20401: Based on the actual precipitation field of the verification year, the predicted precipitation field, the actual temperature field of the verification year, and the predicted temperature field, obtain the comprehensive score of the actual circulation factor.

[0051] The process of obtaining a comprehensive score for the actual circulation factor based on the actual precipitation field, the predicted precipitation field, the actual temperature field, and the predicted temperature field of the verification year specifically includes the following sub-steps: Step S2040101: Calculate the first root mean square error, the first time correlation coefficient, and the first anomaly correlation coefficient based on the actual precipitation field of the verification year and the precipitation prediction field.

[0052] The calculation of the first root mean square error, the first time correlation coefficient, and the first anomaly correlation coefficient based on the actual precipitation field of the verification year and the predicted precipitation field specifically includes the following sub-steps: Step S204010101: Calculate the root mean square error of the first single year, the time correlation coefficient of the first single year, and the anomalous correlation coefficient of the first single year based on the actual precipitation field of the verification year and the corresponding precipitation prediction field for each year.

[0053] In this embodiment, for the validation set corresponding to each k value in the leave-one-out iterative process, the actual precipitation field of the validation year and the corresponding precipitation prediction field generated by the model for that year are used as the basis for calculation. Meteorological indicators are calculated on a grid-by-grid basis within the target area, ensuring that the spatial grid and pentad time series of the two fields are completely and accurately matched. Based on the grid-level pentad precipitation data, three single-year precipitation evaluation indicators are calculated sequentially: the root mean square error (RMSE) of the first single year, the time correlation coefficient (TCC) of the first single year, and the anomaly correlation coefficient (ACC) of the first single year. The root mean square error reflects the prediction... The numerical deviation between measured precipitation and actual precipitation is determined. The time correlation coefficient reflects the linear correlation of the precipitation time series, and the anomaly correlation coefficient reflects the spatial correlation of the precipitation anomaly field. The grid-level calculation results of the three indicators are arithmetically averaged over the target area to obtain the root mean square error, time correlation coefficient, and anomaly correlation coefficient of the first single year corresponding to the validation set of year k. The indicators are calculated for all validation years from 1 to N in this way to obtain N sets of first single-year precipitation evaluation indicators corresponding to each actual circulation factor, covering all historical natural years without missing or duplicates.

[0054] Step S204010102: Calculate the average values ​​of the first single-year root mean square error, the first single-year time correlation coefficient, and the first single-year abnormal correlation coefficient to obtain the first root mean square error, the first time correlation coefficient, and the first abnormal correlation coefficient.

[0055] In this embodiment, for a single real-world circulation factor, all N sets of first-year root mean square errors, first-year time correlation coefficients, and first-year outlier correlation coefficients obtained in N iterations of the leave-one-out method are summarized to ensure that the index sequence completely covers all validation years and has no outliers. The arithmetic mean of the N first-year root mean square errors is obtained as the first root mean square error of the real-world circulation factor. The arithmetic mean of the N first-year time correlation coefficients is obtained as the first time correlation coefficient of the real-world circulation factor. The arithmetic mean of the N first-year outlier correlation coefficients is obtained as the first outlier correlation coefficient of the real-world circulation factor. These three average values ​​are the global precipitation prediction evaluation indicators for the real-world circulation factor over the entire historical time period, comprehensively reflecting its overall precipitation prediction ability and generalization stability. The global precipitation index calculation for all candidate real-world circulation factors is completed using this method.

[0056] Step S2040102: Calculate the precipitation dimension score of the actual circulation factor based on the first root mean square error, the first time correlation coefficient, and the first anomaly correlation coefficient.

[0057] In this embodiment, the first root mean square error, first time correlation coefficient, and first anomaly correlation coefficient of the real-time circulation factor are first normalized. The first root mean square error is a negative indicator (the smaller the value, the better the prediction effect), and it is mapped to the 0-1 interval using reverse normalization. The first time correlation coefficient and the first anomaly correlation coefficient are positive indicators (the larger the value, the better the prediction effect), and they are mapped to the 0-1 interval using positive normalization to eliminate the dimensional differences between different indicators. Then, a grid search method is introduced to determine the optimal weight combination of the three normalized indicators. The historical period is divided into independent sub-training sets and sub-validation sets. All possible weight combinations are traversed with a step size of 0.1, and the comprehensive precipitation evaluation score under each weight is calculated. The weight combination that maximizes the regional average ACC of precipitation prediction in the sub-validation set is selected as the final weight. Finally, the three normalized global precipitation indicators are weighted and summed using the optimal weight to obtain the precipitation dimension score of the real-time circulation factor. The score value is mapped to the 0-1 interval, and the higher the value, the stronger the predictive ability of the factor for precipitation.

[0058] Step S2040103: Calculate the second root mean square error, the second time correlation coefficient, and the second anomaly correlation coefficient based on the actual temperature field of the verification year and the predicted temperature field.

[0059] In this embodiment, the calculation of evaluation indicators for the temperature dimension adopts the same calculation logic and process as that for the precipitation dimension. For the k-th year validation set corresponding to each k value in the leave-one-out iterative process, the actual temperature field of the validation year and the corresponding temperature prediction field generated by the model for that year are used as the calculation basis. The second single-year root mean square error, the second single-year time correlation coefficient, and the second single-year anomalous correlation coefficient are calculated for each grid point in the target area. After the grid-level results are arithmetically averaged for the target area, the single-year temperature evaluation index of the k-th year validation set is obtained. After the single-year temperature index calculation for all N validation years is completed, the N second single-year root mean square errors, N second single-year time correlation coefficients, and N second single-year anomalous correlation coefficients are arithmetically averaged to obtain the global temperature prediction evaluation index of the actual circulation factor over the entire historical period, namely the second root mean square error, the second time correlation coefficient, and the second anomalous correlation coefficient, which comprehensively reflect its overall predictive ability and generalization stability for temperature.

[0060] Step S2040104: Calculate the temperature dimension score of the real-time circulation factor based on the second root mean square error, the second time correlation coefficient, and the second anomaly correlation coefficient.

[0061] In this embodiment, the calculation rules for the temperature dimension score are completely consistent with those for the precipitation dimension score, ensuring the consistency of the two-factor evaluation system. First, the second root mean square error, the second time correlation coefficient, and the second anomaly correlation coefficient are normalized. The second root mean square error is a negative indicator and is normalized in reverse, while the second time correlation coefficient and the second anomaly correlation coefficient are positive indicators and are normalized in positive. All indicators are mapped to the 0-1 interval. Then, the same grid search method as for the precipitation dimension is used to determine the optimal weight combination. That is, the sub-training set and sub-validation set division method of the precipitation dimension weight optimization is used. The weight combination is traversed with a step size of 0.1, and the weight combination that maximizes the regional average ACC of the pentad temperature prediction in the sub-validation set is selected as the final weight. Finally, the optimal weight is used to perform a weighted summation of the three normalized global temperature indicators to obtain the temperature dimension score of the real-time circulation factor. The score value is mapped to the 0-1 interval, and the higher the value, the stronger the predictive ability of the factor for temperature.

[0062] Step S2040105: A comprehensive score is formed based on the precipitation dimension score and the temperature dimension score.

[0063] In this embodiment, based on the calculated precipitation and temperature dimension scores, the optimal weights of the two in the comprehensive score are determined using a grid search method. The historical time period is divided into independent internal training and internal validation sets. In the internal training set, the weight combinations of precipitation and temperature dimension scores are traversed with a certain step size. A weighted comprehensive score is calculated for each weight combination, and the optimal predictive factor is selected accordingly. A dynamic statistical prediction model is then constructed. This model is applied to the internal validation set to calculate a comprehensive evaluation index of the prediction results. The weight combination that optimizes this index is selected as the final weight. The precipitation and temperature dimension scores are then weighted and summed using this optimal weight to obtain the comprehensive score for each actual circulation factor.

[0064] Step S20402: Based on the comprehensive score, select multiple optimal prediction factors from the real-world circulation factors.

[0065] In this embodiment, the comprehensive scores of all participating real-time circulation factors are first summarized to ensure that each real-time circulation factor has unique and complete comprehensive score data without missing, duplicate, or abnormal situations. This comprehensive score comprehensively reflects the overall performance of each factor in precipitation and temperature prediction. Then, all real-time circulation factors are sorted in descending order according to their comprehensive score values. The sorting process strictly follows the size relationship of the score values ​​to ensure the accuracy and objectivity of the sorting results. Finally, from the real-time circulation factor sequence that has been sorted in descending order, the top-ranked real-time circulation factors are selected (three in this embodiment).

[0066] Step S205: Based on the optimal predictor, construct single-factor dynamic-statistical sub-models respectively.

[0067] In this embodiment, based on multiple selected optimal predictive factors, corresponding single-factor dynamic-statistical sub-models are constructed. For each optimal predictive factor, its pentad time series within a historical period is extracted. Combined with the linear regression relationship between the optimized left-field time coefficient and the optimized right-field time coefficient obtained through singular value decomposition, a statistical mapping relationship is established between this factor and the summer pentad precipitation and temperature of the target area. The historical circulation field of each optimal predictive factor is projected onto the corresponding optimized left-field spatial mode to obtain the corresponding time coefficient. This coefficient is then substituted into the fitted linear regression equation to calculate the predicted right-field time coefficient. Finally, it is reconstructed with the optimized right-field spatial mode to generate the precipitation and temperature prediction fields under the sole effect of this factor. This modeling process is executed independently for each of the multiple optimal predictive factors, ultimately forming three independent single-factor dynamic-statistical sub-models. Each sub-model can output corresponding pentad precipitation and temperature prediction results based on the input circulation factor, laying the foundation for subsequent multi-model integrated prediction.

[0068] Step S206: Construct a dynamic-statistical prediction model based on multiple single-factor dynamic-statistical sub-models.

[0069] The construction of the dynamic-statistical prediction model based on multiple single-factor dynamic-statistical sub-models specifically includes the following sub-steps: Step S20601: Combine the single-factor dynamic-statistical sub-models.

[0070] In this embodiment, the three independently constructed single-factor dynamic-statistical sub-models are combined and integrated in parallel, keeping the structure, operation logic and prediction ability of each sub-model intact. This allows the three sub-models to be deployed collaboratively under a unified framework, forming a multi-sub-model combination system that can receive input synchronously and output precipitation and temperature prediction results independently.

[0071] Step S20602: Based on the results of the combination, a grid-by-grid and week-by-week arithmetic averaging strategy is introduced to form a dynamic-statistical prediction model.

[0072] In this embodiment, based on the multi-sub-model collaborative system formed by combining three single-factor dynamic-statistical sub-models, a grid-by-grid and pentad-by-pentad precipitation and temperature prediction fields independently output by each sub-model are introduced. A grid-by-grid and pentad-by-pentad arithmetic mean fusion strategy is adopted. The precipitation prediction values ​​output by the three sub-models at the same spatial grid point and pentad are arithmetically averaged to obtain the fused precipitation prediction value for that grid point and pentad. At the same time, the temperature prediction values ​​output by the three sub-models at the same spatial grid point and pentad are arithmetically averaged to obtain the fused temperature prediction value for that grid point and pentad. By combining the multi-sub-model combination architecture with the fusion rules of the grid-by-grid and pentad-by-pentad arithmetic mean results, a complete and unified dynamic-statistical prediction model is finally formed. This model can output a fused and optimized global pentad-by-pentad precipitation and temperature prediction fields.

[0073] Step S3: Construct a correction model based on the historical return circulation factor and the optimal predictor factor.

[0074] The specific steps for constructing the correction model based on the historical return circulation factor and the optimal predictor factor include the following: Step S301: Extract the actual two-dimensional spatial grid field for each of the optimal prediction factors.

[0075] In this embodiment, for each optimal predictor, the actual two-dimensional spatial grid field corresponding to each time-series event within the complete historical period is independently extracted from the corresponding reanalysis grid point real-world data, according to the target spatial range, two-dimensional spatial grid division method, and event-series time resolution consistent with the previous modeling. This ensures that the extracted actual two-dimensional spatial grid fields are uniformly aligned and accurately matched in terms of spatial grid nodes and event time sequence arrangement, providing standardized and consistent real-world input data for subsequent field projection and prediction calculations.

[0076] Step S302: Extract a two-dimensional spatial grid field of returns from the historical return circulation factor that has the same time, the same elements, and the same spatial range as the optimal prediction factor.

[0077] In this embodiment, for each optimal predictor, a matching two-dimensional spatial grid field is precisely extracted from the corresponding historical return circulation factor dataset, strictly following the time series period, meteorological element type, and target research spatial range that are completely consistent with the optimal predictor, while maintaining the same pentad time resolution and two-dimensional spatial grid division rules. This ensures that each extracted two-dimensional spatial grid field corresponds one-to-one with and is completely aligned with the actual two-dimensional spatial grid field of the optimal predictor in terms of time nodes, element attributes, spatial grid, and pentad arrangement. This provides a standardized and unified comparative data basis for the subsequent return verification and effect evaluation of the dynamic-statistical prediction model.

[0078] Step S303: Using the returned two-dimensional spatial grid field as input features and the actual two-dimensional spatial grid field as output features, construct correction training samples.

[0079] In this embodiment, for each optimal predictor, the reward two-dimensional spatial grid field that is perfectly aligned with its time, element, spatial range and grid resolution is used as the input feature of the correction model. At the same time, the real two-dimensional spatial grid field corresponding to the same epoch, the same spatial grid point and the same element is used as the target output feature. The input features and output features are paired and combined according to the one-to-one correspondence between each epoch and each spatial grid point to form multiple sets of mutually matched correction training samples. The paired samples corresponding to all optimal predictors together constitute a complete correction training sample set.

[0080] Step S304: Train the correction model using the correction training samples based on U-Net.

[0081] In this embodiment, U-Net is used as the basic network structure to construct the correction model. The correction training samples are used as the model training data, the two-dimensional spatial grid field of the feedback is used as the input data of the correction model, and the corresponding real two-dimensional spatial grid field is used as the real label for supervised learning. The loss value between the correction result output by the model and the real field is calculated through forward propagation. Then, the weights and bias parameters of the U-Net network are continuously updated iteratively using the backpropagation algorithm to gradually reduce the error between the model output and the real label. The model's ability to fit and correct the bias of the feedback field is continuously optimized until the model reaches the preset convergence condition. Finally, the training of the correction model is completed, and a trained correction model with the function of correcting prediction field errors is obtained.

[0082] During model building, the original model structure is not adjusted; only parameters such as the loss function and optimizer are modified. Following common parameters, the root mean square error loss function is used, the Adam optimizer is employed, and the learning rate is 0.01, halved every 10 training epochs.

[0083] It is symmetrical from left to right and consists of three parts: the encoder path on the left side of the network, the decoder path on the right side, and the jump connection path.

[0084] Encoder Path: The encoder path, also known as the shrinking network, achieves the goal of reducing image size and doubling the number of channels through four consecutive downsampling operations. Each downsampling operation consists of two consecutive 3x3 convolutions, a ReLU function, and one max pooling operation. Through this continuous downsampling process, the network can acquire shallow feature information of the image.

[0085] Decoder Path: The decoder path, also known as the dilated network, achieves the goal of increasing the image size and halving the number of channels through four consecutive upsampling operations. Similarly, each downsampling operation consists of two consecutive 3x3 deconvolutions to enlarge the image size. Through this continuous upsampling process, the network can acquire deep feature information of the image.

[0086] Skip connection path: U-Net establishes skip connection paths between the encoder and decoder in each layer to connect the two, helping the model to stitch together image features of the same size in the same layer.

[0087] Finally, the network performs a 1x1 convolution to output the final segmented image.

[0088] Therefore, we generally refer to U-Net as a semantic segmentation network based on an Encoder-Decoder architecture. In this architecture, the Encoder is responsible for feature extraction, and the Decoder restores the original resolution. It's simple yet effective, so U-Net performs well even on small sample datasets.

[0089] Compared to SegNet, a semantic segmentation network that is also based on the Encoder-Decoder architecture, U-Net's innovation lies in the establishment of skip connection paths, which helps the model successfully fuse deep and shallow feature information, achieving effective fusion of coarse-grained and fine-grained features.

[0090] Therefore, U-Net's biggest advantage lies in its ability to skip connection paths, providing the network with more refined image features and helping the network achieve pixel-level semantic segmentation.

[0091] At the same time, the U-Net network has a simple structure, so it runs fast and is less prone to overfitting when trained on small datasets.

[0092] Step S4: Correct the future predicted circulation factor using the correction model to obtain the corrected future predicted circulation factor.

[0093] In this embodiment, the future predicted circulation factor to be predicted is preprocessed according to the same spatial grid resolution, feature format and input specifications as the correction training sample, and then input into the U-Net correction model that has been trained. The correction model intelligently identifies, fits and corrects the systematic bias and spatial distribution error of the future predicted circulation factor itself, eliminates the systematic bias of the original predicted circulation field, and finally outputs the optimized and corrected circulation field data with improved accuracy, thus obtaining the corrected future predicted circulation factor that meets the needs of subsequent precipitation and temperature prediction.

[0094] Step S5: Input the corrected future prediction circulation factor into the dynamic-statistical prediction model to obtain the final prediction results of summer pentad precipitation and temperature.

[0095] In this embodiment, the corrected future circulation factors obtained after model correction are input into the constructed dynamic-statistical prediction model according to the data format, spatial grid, and temporal resolution requirements that match the model input. This model is based on the combined single-factor dynamic-statistical sub-model. Through a grid-by-grid and pentad-by-pentad arithmetic mean fusion strategy, the corrected circulation factors are sequentially subjected to factor projection, regression calculation, prediction field reconstruction, and multi-sub-model result fusion processing. Finally, the spatial grid precipitation prediction field and temperature prediction field for the target area in summer pentad are output, resulting in accurate and reliable final prediction results for summer pentad precipitation and temperature.

[0096] like Figure 2 As shown, in another aspect, the present invention also provides a dynamic-statistical prediction system for summer pentad precipitation and temperature, comprising: a processor, an input device, an output device, and a memory, wherein the processor, the input device, the output device, and the memory are interconnected, wherein the memory is used to store a computer program, the computer program including program instructions, and the processor is configured to call the program instructions to execute the relevant steps of a relevant embodiment of the dynamic-statistical prediction method for summer pentad precipitation and temperature of the present invention.

[0097] This invention provides a dynamic-statistical forecasting system for summer pentad precipitation and temperature. The functional components can be integrated into a single processing unit, exist as individual physical entities, or be integrated into a single unit. The integrated components can be implemented in hardware or as software functions.

[0098] The improved model of this invention has the following effects: The prediction results were subjected to grid verification. The actual data was CRA40, and the verification dimensions were divided into the following two types: Period 1 uses forecast data for every 6 pentads from May to September 2025 (rolling every 2 pentads), with an advance of 11-20 days (for example, the average temperature forecast for a fixed period from August 11 to September 10, 2025, based on daily forecasts from July 22 to July 31, 2025, is used for verification; the logic is the same for other verification periods).

[0099] The second period uses forecast data for every three pentads from May to September 2025 (rolling every two pentads), with an advance of 11-40 days (selecting 11, 15, 20, 25, 30, 35, and 40 days in advance for display; for example, using the average temperature forecast for a fixed period from July 11 to July 25, 2025, predicted 35 days in advance on June 6, 2025, for verification; the logic for other verification periods is the same).

[0100] Note: Cases where some models did not generate data or where the original model had no data are not included in the statistics. All scores are rounded to two decimal places.

[0101] RMSE score for predicting the average temperature of the next 6 pentads 11-20 days in advance: The score was compared with the original CPSv3 model's prediction of average temperature, and the results are as follows: Figure 3 As shown. RMSE examines the root mean square error; a smaller score indicates better data quality. It can be seen that the overall score of the data using the model is significantly lower than the original model, indicating that the improved dynamic-statistical prediction model (for ease of description, the improved model of this invention is abbreviated as MJO) is superior to the original model in terms of average temperature prediction. The average score of the original model for different lead times was 3.72, while the model score was 3.27. Looking at each lead time individually, the overall prediction effect was not significantly different. The overall trend was that as the lead time increased, the score slowly increased, and the prediction effect deteriorated. The model performed best with a lead time of 16 days, achieving a score of 3.14.

[0102] RMSE score for predicting the average temperature of the next three pentads 11-40 days in advance: The score was compared with the original CPSv3 model's prediction of average temperature, and the results are as follows: Figure 4 As shown. RMSE examines the root mean square error; a smaller score indicates better data quality. It can be seen that the overall score of the data using the model is significantly lower than the original model, indicating that the MJO-based dynamic-statistical prediction model is superior to the original model in predicting average temperature. The original model's average score for different lead times was 3.58, while the model's score was 3.05. However, for each lead time, the overall prediction performance was not significantly different. The overall trend was that the score slowly increased with increasing lead time, indicating a worse prediction performance. The model performed best with a lead time of 25 days, achieving a score of 2.96.

[0103] TCC score for predicting the average temperature of the next 6 pentads 11-20 days in advance: The score was compared with the original CPSv3 model's prediction of average temperature, and the results are as follows: Figure 5 As shown, the TCC score examines the time correlation coefficient, and it can be seen that the model performs better overall. All scores of the model are higher than those of the original model, with an average score of 0.38 compared to 0.31 for the original model. The overall score is relatively unstable, and the pattern of change over time is not obvious. Among them, the model with a lead time of 16 days has the highest score, at 0.43.

[0104] TCC score for predicting the average temperature of the next three pentads 11-40 days in advance: The score was compared with the original CPSv3 model's prediction of average temperature, and the results are as follows: Figure 6 As shown, the TCC score examines the time correlation coefficient, and it can be seen that the model performs relatively better. The model's score is higher than the original model in most cases, with an average score of 0.17 compared to the original model's 0.12. The model score is highest at 0.26 when the scoring lead time is 35 days.

[0105] ACC score for predicting the average temperature of the next six pentads 11-20 days in advance: The score was compared with the original CPSv3 model's prediction of average temperature, and the results are as follows: Figure 7 As shown, the ACC score examines the abnormal correlation coefficient. It can be seen that the score fluctuates with the change in lead time. The model's performance is generally better than the original model, with almost all scores higher than the original model. Among them, the model scores the highest at 0.14 for a lead time of 11 days. The average score of the model for all lead times is 0.06, slightly higher than the original model's 0.01.

[0106] ACC score for predicting the average temperature of the next three pentads 11-40 days in advance: The score was compared with the original CPSv3 model's prediction of average temperature, and the results are as follows: Figure 8 As shown, the ACC score examines the abnormal correlation coefficient. It can be seen that the score fluctuates with the change in lead time. The model performs better than the original model in all aspects, with all scores higher than the original model and the trend being correct. Among them, the model score is the highest at 0.10 for a lead time of 30 days. The average score of all lead times in the model is 0.05, slightly higher than the original model's 0.01.

[0107] Precipitation RMSE score for the next 6 pentads 11-20 days in advance: The results were compared with the original CPSv3 model's precipitation predictions, and the scores were as follows: Figure 9As shown, RMSE examines the root mean square error; a smaller score indicates better data quality. It can be seen that the overall score of the data using the model is significantly lower than the original model, indicating that the MJO-based dynamic-statistical prediction model is superior to the original model in precipitation prediction. The average score of the original model for different lead times was 7.54, while the model scored 5.39. However, for each lead time, the overall prediction performance was not significantly different. The overall trend was that the score slowly increased with increasing lead time, indicating a worse prediction performance. The model performed best with a lead time of 17 days, scoring 5.13.

[0108] RMSE score for predicting precipitation over the next three pentads 11-40 days in advance: The results were compared with the original CPSv3 model's precipitation predictions, and the scores were as follows: Figure 10 As shown, RMSE examines the root mean square error; a smaller score indicates better data quality. It can be seen that the overall score of the data using the model is significantly lower than the original model, indicating that the MJO-based dynamic-statistical prediction model is superior to the original model in precipitation prediction. The original model's average score for different lead times was 7.20, while the model's score was 4.96. However, for each lead time, the overall prediction performance was not significantly different, with the model showing the best prediction performance at 35 days, at 4.77.

[0109] Precipitation forecast for the next 6 pentads 11-20 days in advance (TCC score): The results were compared with the original CPSv3 model's precipitation predictions, and the scores were as follows: Figure 11 As shown, the TCC score examines the time correlation coefficient. It can be seen that the model performs better overall. Almost all of the model's scores are higher than the original model, and most of them are trend-correct. The average score is 0.06, while the original model is 0.01. The model with a scoring lead time of 17 days has the highest score, at 0.10.

[0110] Precipitation forecast for the next three pentads 11-40 days in advance (TCC score): The results were compared with the original CPSv3 model's precipitation predictions, and the scores were as follows: Figure 12 As shown, the TCC score examines the time correlation coefficient. It can be seen that the model performs better overall. Almost all of the model's scores are higher than the original model, and most of them are correct in trend. The average score is 0.04, while the original model is -0.02. The model with a scoring lead time of 20 days has the highest score, at 0.09.

[0111] ACC score for predicting precipitation 11-20 days in advance for the next 6 pentads: The results were compared with the original CPSv3 model's precipitation predictions, and the scores were as follows: Figure 13As shown, the ACC score examines the abnormal correlation coefficient. It can be seen that the overall performance of the model is generally average, with the opposite trend occurring. The absolute value of the model score is significantly higher than that of the original model, but the trend is opposite.

[0112] ACC score for predicting precipitation over the next three pentads 11-40 days in advance: The results were compared with the original CPSv3 model's precipitation predictions, and the scores were as follows: Figure 14 As shown, the ACC score examines the abnormal correlation coefficient. It can be seen that the overall performance of the original model and the model is generally average, with opposite trends. The absolute value of the model score is significantly higher than that of the original model, but the trend is opposite.

[0113] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.

Claims

1. A dynamic-statistical prediction method for summer pentad precipitation and temperature, characterized in that, The method includes: Obtain real-time grid data, historical return circulation factors, and future predicted circulation factors for each pentad in the target area during historical periods; Based on the reanalysis grid data, the optimal predictor was selected using singular value decomposition and a dynamic-statistical prediction model was constructed. A correction model is constructed based on the historical return circulation factor and the optimal predictor factor; The future predicted circulation factor is corrected using the correction model to obtain the corrected future predicted circulation factor; The corrected future circulation factor is input into the dynamic-statistical prediction model to obtain the final prediction results of summer pentad precipitation and temperature.

2. The dynamic-statistical prediction method for summer pentad precipitation and temperature according to claim 1, characterized in that, The step of using singular value decomposition to screen the optimal predictor and construct a dynamic-statistical prediction model based on the reanalysis grid point real-world data includes: The leave-one-out method was used to extract the actual circulation factor, actual precipitation field, actual temperature field, and actual precipitation field and actual temperature field of the validation year from the actual data of the reanalysis grid points in sequence; Singular value decomposition and screening are performed on the actual circulation factor, the actual precipitation field and the actual temperature field to obtain the optimized left field spatial mode, optimized left field time coefficient, optimized right field spatial mode and optimized right field time coefficient corresponding to precipitation and temperature, respectively. Based on the optimized left-field spatial mode, the optimized left-field time coefficient, the optimized right-field spatial mode, and the optimized right-field time coefficient, the precipitation prediction field and temperature prediction field corresponding to the actual circulation factor are obtained; Based on the precipitation prediction field, the temperature prediction field, the actual precipitation field of the verification year, and the actual temperature field of the verification year, multiple optimal prediction factors are selected from the actual circulation factors. Based on the optimal predictor, single-factor dynamic-statistical sub-models are constructed respectively; A dynamic-statistical prediction model is constructed based on multiple single-factor dynamic-statistical sub-models.

3. The dynamic-statistical prediction method for summer pentad precipitation and temperature according to claim 2, characterized in that, The process of performing singular value decomposition and filtering on the actual circulation factor, the actual precipitation field, and the actual temperature field to obtain the optimized left-field spatial mode, optimized left-field time coefficient, optimized right-field spatial mode, and optimized right-field time coefficient corresponding to precipitation and temperature, respectively, includes: Singular value decomposition is performed using the actual circulation factor as the left field and the actual precipitation field as the right field to obtain multiple sets of first singular values ​​and first coupled mode data corresponding to the precipitation. Using the actual circulation factor as the left field and the actual temperature field as the right field, singular value decomposition is performed to obtain multiple sets of second singular values ​​and second coupled mode data corresponding to the temperature. By selecting the maximum of the first singular value and the second singular value, multiple sets of optimized left-field spatial modes, optimized left-field time coefficients, optimized right-field spatial modes, and optimized right-field time coefficients corresponding to the precipitation and the temperature are obtained respectively.

4. The dynamic-statistical prediction method for summer pentad precipitation and temperature according to claim 2, characterized in that, The process of obtaining the precipitation prediction field and temperature prediction field corresponding to the actual circulation factor based on the optimized left-field spatial mode, the optimized left-field time coefficient, the optimized right-field spatial mode, and the optimized right-field time coefficient includes: By performing linear fitting on the optimized left field time coefficient and the optimized right field time coefficient, multiple linear regression relationships corresponding to the precipitation and the temperature are obtained respectively. Based on the linear regression relationship, the optimized left field time coefficient is used to calculate the predicted right field time coefficients for precipitation and temperature, respectively. Based on the predicted right-field time coefficient and the optimized right-field spatial mode, the precipitation prediction field and temperature prediction field corresponding to the actual circulation factor are reconstructed.

5. The dynamic-statistical prediction method for summer pentad precipitation and temperature according to claim 2, characterized in that, The step of selecting multiple optimal prediction factors from the actual circulation factors based on the precipitation prediction field, the temperature prediction field, the actual precipitation field of the verification year, and the actual temperature field of the verification year includes: Based on the actual precipitation field of the verification year, the predicted precipitation field, the actual temperature field of the verification year, and the predicted temperature field, a comprehensive score for the actual circulation factor is obtained. Based on the comprehensive score, several optimal predictive factors are selected from the real-world circulation factors.

6. The dynamic-statistical prediction method for summer pentad precipitation and temperature according to claim 5, characterized in that, The comprehensive score of the actual circulation factor obtained based on the verified year's actual precipitation field, the predicted precipitation field, the verified year's actual temperature field, and the predicted temperature field includes: The first root mean square error, the first time correlation coefficient, and the first anomaly correlation coefficient are calculated based on the actual precipitation field of the verification year and the precipitation prediction field. The precipitation dimension score of the real circulation factor is calculated based on the first root mean square error, the first time correlation coefficient, and the first anomaly correlation coefficient. The second root mean square error, the second time correlation coefficient, and the second anomaly correlation coefficient are calculated based on the actual temperature field of the verification year and the temperature prediction field. The temperature dimension score of the real-time circulation factor is calculated based on the second root mean square error, the second time correlation coefficient, and the second anomaly correlation coefficient. A comprehensive score is formed based on the precipitation dimension score and the temperature dimension score.

7. The dynamic-statistical prediction method for summer pentad precipitation and temperature according to claim 6, characterized in that, The calculation of the first root mean square error, the first time correlation coefficient, and the first anomaly correlation coefficient based on the verified year's actual precipitation field and the predicted precipitation field includes: The root mean square error, temporal correlation coefficient, and anomalous correlation coefficient for the first year are calculated based on the actual precipitation field and the corresponding precipitation prediction field for each year of the verification year, respectively. The first root mean square error, the first time correlation coefficient, and the first abnormal correlation coefficient are obtained by averaging the first single-year root mean square error, the first single-year time correlation coefficient, and the first single-year abnormal correlation coefficient, respectively.

8. The dynamic-statistical prediction method for summer pentad precipitation and temperature according to claim 2, characterized in that, The construction of the dynamic-statistical prediction model based on multiple single-factor dynamic-statistical sub-models includes: Combine the single-factor dynamic-statistical sub-models; Based on the results of the combination, a grid-by-grid and week-by-week arithmetic averaging strategy is introduced to form a dynamic-statistical prediction model.

9. The dynamic-statistical prediction method for summer pentad precipitation and temperature according to claim 1, characterized in that, The construction of the correction model based on the historical return circulation factor and the optimal predictor factor includes: Extract the real-time two-dimensional spatial grid field for each of the optimal predictors; Extract a two-dimensional spatial grid field of returns from the historical return circulation factor that has the same time, the same elements, and the same spatial range as the optimal prediction factor; Using the returned two-dimensional spatial grid field as input features and the actual two-dimensional spatial grid field as output features, a correction training sample is constructed. The correction model is trained using the correction training samples based on U-Net.

10. A dynamic-statistical prediction system for summer pentad precipitation and temperature, characterized in that, include: The system includes a processor, an input device, an output device, and a memory, all interconnected. The memory stores a computer program, which includes program instructions. The processor is configured to invoke the program instructions to execute a dynamic-statistical prediction method for summer pentad precipitation and temperature as described in any one of claims 1 to 9.

Images