A method for predicting monthly or seasonal scale precipitation based on interannual increment

Abstract

The application discloses a kind of monthly or seasonal scale precipitation prediction methods based on interannual increment, which comprises the following steps: constructing potential predictor dataset;Monthly or seasonal average precipitation anomaly percentage interannual increment single-factor prediction model is constructed;The deviation of monthly or seasonal average precipitation anomaly percentage interannual increment predicted by single-factor prediction model is revised based on singular value decomposition method;After the average precipitation anomaly percentage interannual increment predicted by each single-factor model is revised, the average precipitation anomaly percentage of the same period observed in the previous year is added to obtain the average precipitation anomaly percentage predicted by single-factor model, and the preferred set of multi-factor prediction model is obtained by confidence level threshold;The selection month of predictor and the target month of precipitation prediction are adjusted to realize the rolling prediction of monthly or seasonal precipitation.The application optimizes the selection of predictor, prediction object, multi-model set and rolling prediction scheme, and improves the accuracy of monthly or seasonal scale precipitation prediction.

Description

Technical Field

[0001] This invention relates to the field of precipitation prediction technology, specifically to a method for predicting monthly or seasonal precipitation based on interannual increments. Background Technology

[0002] In recent years, my country has experienced significant inter-monthly climate anomalies, leading to severe meteorological disasters, such as rapid shifts between drought and flood in different summer months and extreme cold and warm events between different winter months. Domestic and international climate dynamic models have low predictive efficiency for seasonal precipitation, and accurate prediction of inter-monthly precipitation is even more difficult. Therefore, there is an urgent need to improve the predictive capability for inter-monthly and seasonal precipitation.

[0003] In short-term climate prediction, the traditional prediction object is the anomaly of climate variables (the difference between a climate variable and its multi-year average), but the multi-year average of climate exhibits certain interdecadal variations. The interannual increment prediction method uses the interannual increment of climate quantities (the current year's value minus the previous year's value) as the prediction object, making climate predictions based on the interannual increment variation patterns of climate variables (Fan et al., 2008). The interannual increment method can effectively utilize the previous year's observational information and the quasi-biennial variation patterns of climate variables. Compared with traditional anomalies, the interannual increment prediction method can not only effectively amplify the signals of climate quantities and predictive factors, but also better capture the interdecadal signals of climate variables, thereby improving the effectiveness of climate prediction. The interannual increment method has now been widely applied to climate prediction across multiple time scales and fields (such as summer precipitation, winter temperature, North Atlantic Oscillation, East Asian monsoon, frequency of extreme precipitation, frequency of typhoon activity, and haze), achieving good predictive results.

[0004] Commonly used methods for short-term climate prediction include statistical methods, dynamical models, and combined dynamical and statistical methods. Statistical methods are computationally efficient and easy to implement, but the statistical relationships are unstable. Dynamical models are primarily effective in the tropics, with lower predictive power for mid- and high-latitude climates. Traditional combined dynamical and statistical methods rely solely on the results of dynamical models, offering limited improvement in predictive capability. The selection of key regions for predictive factors in climate prediction is largely subjective, based on statistical relationships, making it difficult to implement rolling predictions from the model.

[0005] Among existing short-term climate prediction methods, such as the short-term climate prediction method based on statistical downscaling technology with patent publication number CN113219555A, this method mainly focuses on using field information coupling methods to establish prediction models, but mentions less about automatic selection algorithms for key areas of prediction factors, ensemble schemes of multi-factor models, and rolling prediction technology schemes. Summary of the Invention

[0006] To overcome the shortcomings and deficiencies of existing technologies, this invention provides a method for predicting monthly or seasonal precipitation based on interannual increments. Based on key climate factors influencing monthly-seasonal precipitation, this invention automatically determines key regions of predictive factors using image closure operations in mathematical morphology. It combines the patterns of previous climate variability with the predictive effectiveness of concurrent dynamic models. Through interannual increments, coupled field information prediction, and optimal ensemble analysis, it achieves real-time rolling prediction of monthly or seasonal average precipitation, combining dynamics and statistics. Optimizations have been made in the selection of predictive factors, prediction objects, multi-model ensemble schemes, and rolling prediction schemes, thereby improving the accuracy of monthly-seasonal precipitation prediction.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] This invention provides a method for predicting monthly or seasonal precipitation based on interannual increments, comprising the following steps:

[0009] Construct a dataset of potential predictors;

[0010] Construct a single-factor prediction model for the interannual increment of the percentage anomaly in monthly or seasonal average precipitation;

[0011] The singular value decomposition method was used to correct the deviation of the annual increment of the monthly or seasonal average precipitation anomaly percentage predicted by the single-factor prediction model.

[0012] The average precipitation anomaly percentage predicted by each single-factor model is obtained by correcting the interannual increment of the average precipitation anomaly percentage observed in the same period of the previous year and then adding it to the average precipitation anomaly percentage predicted by the single-factor model.

[0013] The monthly or seasonal average precipitation anomaly percentage and the correlation coefficient of the observation time are selected during the modeling stage. The results of single-factor models are screened by setting a confidence level threshold and then averaged equally to obtain the final prediction result. If there is no single-factor model that passes the significance test, the equal weight results of multiple single-factor models are used as the final prediction result.

[0014] Adjusting the selected months for predictive factors and the target months for precipitation forecasting enables rolling monthly or seasonal precipitation forecasts.

[0015] As a preferred technical solution, a potential predictor dataset is constructed, specifically including:

[0016] Calculate the interannual increment of the monthly or seasonal average precipitation anomaly percentage for each meteorological station or grid point. Extract the time coefficients corresponding to the modes that satisfy the set proportion of variance explanation for the interannual increment of the monthly or seasonal average precipitation anomaly percentage through empirical orthogonal function decomposition. Select key regions for predictors based on the time coefficients corresponding to the selected modes. Rank the potential predictors for each climate variable according to the area size of the key regions. Compare the predictive efficacy of the potential predictors in the key regions and select the predictors.

[0017] As a preferred technical solution, a potential predictor dataset is constructed, specifically including:

[0018] Calculate the time correlation coefficients between grid points of the interannual increment fields of each selected predictor and predictor variable, the observed monthly or seasonal climate variables, and the climate model-predicted contemporaneous climate variables.

[0019] The time correlation coefficient is assigned the explained variance of the corresponding EOF mode by grid points with set confidence level thresholds, and the grid points without set confidence level thresholds are assigned 0, so as to obtain the spatial distribution of the explained variance for each mode. The spatial field of the explained variance for each mode of the predictor is accumulated and summed.

[0020] The values ​​greater than 0 in the cumulative explained variance space field are sorted from smallest to largest. A threshold for the key region is selected, and the morphological closing operation method is used to filter the grid points in the cumulative explained variance space field that are greater than or equal to the threshold to determine the key regions of the predictor.

[0021] By using morphological closing operations to screen grid points in the cumulative explained variance space field that are greater than or equal to a threshold, key regions of predictors are determined. Based on the size of the selected key regions, the main key regions are selected as potential predictors. The predictive efficacy of the potential predictors in the key regions is compared, and the predictors are selected.

[0022] As a preferred technical solution, a single-factor prediction model for the interannual increment of the percentage anomaly in monthly or seasonal average precipitation is constructed, specifically including:

[0023] A single-factor prediction model for the interannual increment of the percentage of monthly or seasonal average precipitation anomalies was established using the field information coupling method.

[0024] Field reconstruction is performed by extracting predictors and predictor factors that satisfy a set ratio to explain the variance using empirical orthogonal function decomposition.

[0025] The optimal coupling variation of the spatial field of predictor and predictor after field reconstruction is extracted using the singular value decomposition method.

[0026] Construct the product matrix of the predictor and the predictor, and perform singular value decomposition on the product matrix;

[0027] For the time coefficients corresponding to the extracted optimal coupling variation, a single-factor downscaling prediction model is established using the linear regression method.

[0028] As a preferred technical solution, the time coefficient corresponding to the extracted optimal coupling variation type specifically includes:

[0029]

[0030] Where k is the total number of SVD modes, Y(t,x) represents the interannual increment of PAP, X(t,x) represents the predictor factor, and R i (x), W i (x) represent the interannual increment of the average precipitation anomaly percentage and the singular vector of the i-th mode of the predictor, respectively. i (t) and O i (t) represents the time dilation coefficient of the i-th SVD mode of the interannual increment of the average precipitation anomaly percentage and the predictor factor;

[0031] A single-factor downscaling prediction model is established using linear regression, specifically including:

[0032] For each mode, the corresponding time dilation coefficient S i (t) and time dilation factor O i (t) is subjected to univariate linear regression, based on the univariate linear regression coefficient α. i and residual ε i The time dilation coefficients of each SVD mode of the i-th mode at the predicted time t+Δt are obtained and expressed as:

[0033]

[0034] Based on the singular vector R of the i-th mode of the interannual increment of the average precipitation anomaly percentage i Based on the prediction results of (x) and the time dilation coefficients of each SVD mode, the interannual increment of the predicted average precipitation anomaly percentage at time t+Δt is reconstructed and expressed as:

[0035]

[0036] in, It represents the interannual increment of the average precipitation anomaly predicted at time t+Δt.

[0037] The present invention also provides a monthly or seasonal precipitation prediction system based on interannual increments, comprising: a dataset construction module, a single-factor prediction model construction module, a prediction model error correction module, a prediction model optimization module, and a rolling prediction module;

[0038] The dataset construction module is used to construct a potential predictor dataset;

[0039] The single-factor prediction model building module is used to build a single-factor prediction model for the annual increment of the monthly or seasonal average precipitation anomaly percentage.

[0040] The prediction model error correction module is used to correct the deviation of the annual increment of the monthly or seasonal average precipitation anomaly percentage predicted by the single-factor prediction model based on the singular value decomposition method.

[0041] The prediction model selection module is used to correct the interannual increment of the average precipitation anomaly percentage predicted by each single-factor model and add it to the average precipitation anomaly percentage observed in the same period of the previous year to obtain the average precipitation anomaly percentage predicted by the single-factor model.

[0042] The monthly or seasonal average precipitation anomaly percentage and the correlation coefficient of the observation time are selected during the modeling stage. The results of single-factor models are screened by setting a confidence level threshold and then averaged equally to obtain the final prediction result. If there is no single-factor model that passes the significance test, the equal weight results of multiple single-factor models are used as the final prediction result.

[0043] The rolling forecast module is used to adjust the selected months of the forecast factors and the target months of precipitation forecasting, so as to realize rolling forecasts of monthly or seasonal precipitation.

[0044] As a preferred technical solution, the dataset construction module is used to construct a potential predictor dataset, specifically including:

[0045] Calculate the interannual increment of the monthly or seasonal average precipitation anomaly percentage for each meteorological station or grid point. Extract the time coefficients corresponding to the modes that satisfy the set proportion of variance explanation for the interannual increment of the monthly or seasonal average precipitation anomaly percentage through empirical orthogonal function decomposition. Select key regions for predictors based on the time coefficients corresponding to the selected modes. Rank the potential predictors for each climate variable according to the area size of the key regions. Compare the predictive efficacy of the potential predictors in the key regions and select the predictors.

[0046] As a preferred technical solution, the dataset construction module is used to construct a potential predictor dataset, specifically including:

[0047] Calculate the time correlation coefficients between grid points of the interannual increment fields of each selected predictor and predictor variable, the observed monthly or seasonal climate variables, and the climate model-predicted contemporaneous climate variables.

[0048] The time correlation coefficient is assigned the explained variance of the corresponding EOF mode by grid points with set confidence level thresholds, and the grid points without set confidence level thresholds are assigned 0, so as to obtain the spatial distribution of the explained variance for each mode. The spatial field of the explained variance for each mode of the predictor is accumulated and summed.

[0049] The values ​​greater than 0 in the cumulative explained variance space field are sorted from smallest to largest. A threshold for the key region is selected, and the morphological closing operation method is used to filter the grid points in the cumulative explained variance space field that are greater than or equal to the threshold to determine the key regions of the predictor.

[0050] By using morphological closing operations to screen grid points in the cumulative explained variance space field that are greater than or equal to a threshold, key regions of predictors are determined. Based on the size of the selected key regions, the main key regions are selected as potential predictors. The predictive efficacy of the potential predictors in the key regions is compared, and the predictors are selected.

[0051] As a preferred technical solution, the single-factor prediction model construction module is used to construct a single-factor prediction model for the interannual increment of the monthly or seasonal average precipitation anomaly percentage, specifically including:

[0052] A single-factor prediction model for the interannual increment of the percentage of monthly or seasonal average precipitation anomalies was established using the field information coupling method.

[0053] Field reconstruction is performed by extracting predictors and predictor factors that satisfy a set ratio to explain the variance using empirical orthogonal function decomposition.

[0054] The optimal coupling variation of the spatial field of predictor and predictor after field reconstruction is extracted using the singular value decomposition method.

[0055] Construct the product matrix of the predictor and the predictor, and perform singular value decomposition on the product matrix;

[0056] For the time coefficients corresponding to the extracted optimal coupling variation, a single-factor downscaling prediction model is established using the linear regression method.

[0057] As a preferred technical solution, the time coefficient corresponding to the extracted optimal coupling variation type specifically includes:

[0058]

[0059] Where k is the total number of SVD modes, Y(t,x) represents the interannual increment of PAP, X(t,x) represents the predictor factor, and R i (x), W i (x) represent the interannual increment of the average precipitation anomaly percentage and the singular vector of the i-th mode of the predictor, respectively. i (t) and O i (t) represents the time dilation coefficient of the i-th SVD mode of the interannual increment of the average precipitation anomaly percentage and the predictor factor;

[0060] A single-factor downscaling prediction model is established using linear regression, specifically including:

[0061] For each mode, the corresponding time dilation coefficient S i (t) and time dilation factor O i (t) is subjected to univariate linear regression, based on the univariate linear regression coefficient α. i and residual ε i The time dilation coefficients of each SVD mode of the i-th mode at the predicted time t+Δt are obtained and expressed as:

[0062]

[0063] Based on the singular vector R of the i-th mode of the interannual increment of the average precipitation anomaly percentage i Based on the prediction results of (x) and the time dilation coefficients of each SVD mode, the interannual increment of the predicted average precipitation anomaly percentage at time t+Δt is reconstructed and expressed as:

[0064]

[0065] in, It represents the interannual increment of the average precipitation anomaly predicted at time t+Δt.

[0066] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0067] (1) In short-term climate prediction, the traditional prediction object is the anomaly. Climate anomalies are affected by the climate state time period. This invention uses the interannual increment of climate variables for modeling, which not only eliminates the influence of climate state on the prediction results, but also amplifies the prediction signals of the prediction quantity and prediction factors, improves the prediction ability of monthly or seasonal precipitation, and can also capture the interdecadal signals of climate variables.

[0068] (2) Compared with selecting predictors based on a single modal coefficient, this invention selects predictors based on the cumulative explained variance of the time coefficients corresponding to multiple empirical orthogonal function modes of precipitation at the monthly or seasonal scale and the related fields of climate variables, which fully considers the key climate factors affecting precipitation.

[0069] (3) Existing technologies mainly select key regions for predictors based on the regions that pass the significance test in the spatial field of the time coefficient corresponding to a single empirical orthogonal function mode and the time correlation coefficient of climate variables. The selection of the key region boundary is subjective and it is difficult to achieve rolling prediction. Compared with the existing subjective selection method of key regions for predictors, this invention uses the cumulative explained variance of the correlation field of the time coefficient corresponding to the EOF mode before 90% of the inter-month precipitation as the basis, and uses the image closing operation in mathematical morphology to automatically determine the key regions of predictors, thereby realizing the automatic selection of key regions for predictors and realizing rolling prediction of the prediction model that combines dynamics and statistics.

[0070] (4) Traditional multi-model ensemble schemes are mostly equal-weighted averages. This invention adopts a multi-factor optimal ensemble scheme based on the prediction performance of each site or grid point in the modeling stage, which fully considers the regional differences in the prediction performance of different factors and the interdecadal changes in the relationship between prediction factors and prediction quantities. Attached Figure Description

[0071] Figure 1 This is a flowchart illustrating the monthly or seasonal precipitation prediction method of the present invention.

[0072] Figure 2 This is a schematic diagram illustrating the implementation process of field information coupling in this invention. Detailed Implementation

[0073] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0074] Example 1

[0075] like Figure 1 As shown, this embodiment provides a method for predicting monthly or seasonal precipitation based on interannual increments, including the following steps:

[0076] S1: Establish a dataset of potential predictors;

[0077] In the computational modeling phase, the annual increment of the monthly or seasonal average precipitation anomaly percentage (PAP) for each meteorological station or grid point is calculated. The time coefficients corresponding to the modes that explain the variance of the annual increment of the monthly or seasonal average PAP are extracted using empirical orthogonal function (EOF) decomposition. Based on the selected time coefficients corresponding to the modes, key regions for predictors are selected. Specific steps include:

[0078] (a) Calculate the time correlation coefficient (TCC) between each selected predictor and predictor variable time coefficient (PC) and each grid point of the interannual increment field of previously observed monthly or seasonal climate variables (such as sea surface temperature, sea level pressure, 500 hPa geopotential height field, sea ice concentration, soil moisture, snow cover, etc.) and climate model predicted contemporaneous climate variables.

[0079] (b) The time correlation coefficient (TCC) is assigned to the explanatory variance of the corresponding EOF mode through grid points at the 90% confidence level, and grid points that do not pass the 90% confidence level are assigned 0, so as to obtain the spatial distribution of the explanatory variance corresponding to each mode. The spatial field of the explanatory variance corresponding to each mode of the predictor is accumulated and summed. The region with high accumulated explanatory variance indicates the region where the predictor has the most significant impact on multiple main modes of precipitation.

[0080] (c) Calculate and sort the values ​​greater than 0 in the cumulative explained variance space field from smallest to largest, and select the value at the 90th percentile as the threshold for selecting key regions. Use morphological closing operations to optimize and filter grid points in the cumulative explained variance space field that are greater than or equal to the threshold, and determine the key regions for predictors. Based on the size of the selected key regions, select the top three key regions by area for each climate variable. Based on the area size of the selected key regions, for each climate variable, select the top three key regions by area (this parameter can be adjusted according to actual conditions) as potential predictors. By comparing the predictive efficacy of the three key region potential predictors in the modeling stage, dynamically select the predictor with the highest predictive efficacy.

[0081] Morphological closing operations process images through the dilation and erosion operations of structuring elements. The advantage of using this method to select key regions of predictors is that it can not only effectively eliminate small-scale noise and preserve large-scale structural features, but also achieve automatic selection of key regions and enable monthly rolling prediction of the prediction model.

[0082] S2: Construct a single-factor prediction model for the interannual increment of monthly or seasonal precipitation PAP;

[0083] Based on the potential predictor dataset established in step S1, such as Figure 2 As shown, firstly, a single-factor prediction model for the interannual increment of monthly or seasonal average precipitation PAP is established using the field information coupling method. The EOF decomposition is used to extract the modes of the predicted quantity and the top 90% explained variance of the predictor for field reconstruction to achieve filtering. Secondly, the optimal coupling variation of the spatial fields of the predicted factor (X) and predicted quantity (Y) after field reconstruction is extracted using the singular value decomposition (SVD) method. A product matrix A = Y × X of the predicted quantity and the predictor is constructed, and SVD decomposition of matrix A is performed: A = U × S × V, where matrix U is the left singular matrix of A, matrix V is the right singular matrix of A, and the diagonal elements of matrix S are the singular values ​​of A, and A is the singular value matrix. Finally, for the time coefficients corresponding to the extracted optimal coupling variation (Formulas (1) and (2)), a single-factor downscaling prediction model is established using linear regression and the singular value decomposition (SVD) method (Formulas (3) and (4)). The downscaling prediction method can obtain a stable statistical relationship between the large-scale circulation system and local meteorological elements.

[0084]

[0085] Where k is the total number of SVD modes, and Y(t,x) represents the interannual increment of PAP (ΔY). i X(t,x) represents the predictor, and Rt represents the predictor. i (x), W i (x) represent the interannual increment of PAP (ΔY) i S and the singular vector of the i-th mode of the predictor. i (t) and O i (t) represents ΔY i The time dilation coefficient of the i-th SVD mode of the predictor;

[0086] For each mode corresponding to S i (t) and O i (t) is subjected to univariate linear regression, based on the univariate linear regression coefficient α. i and residual ε i The time dilation coefficients of each SVD mode of the i-th mode at the predicted time t+Δt are obtained and expressed as:

[0087]

[0088] Combining the time dilation coefficient prediction results of each SVD mode established by formula (3) with the mode reconstruction of formula (1), the interannual increment of precipitation PAP at time t+Δt is obtained. Represented as:

[0089]

[0090] S3: Predictive model error correction;

[0091] This invention employs the Singular Value Decomposition (SVD) method to correct the bias of the single-factor prediction model for PAP interannual increments constructed in step S2. SVD utilizes the synergistic relationship between the observed and predicted fields to correct biases for different modes and their corresponding time coefficients, with the best correction effect achieved by directly using the observed modes to replace the predicted modes. Since the constructed downscaling prediction model retains the first 99% of the modes in the reconstructed PAP interannual increment field, this invention primarily focuses on correcting the first 99% of the modes predicted by the downscaling model using the SVD method.

[0092] S4: Optimal set of multi-factor prediction models;

[0093] The interannual increment of PAP predicted by each single-factor model is corrected and then added to the PAP observed in the same period of the previous year to obtain the PAP predicted by the single-factor model. For each site or grid point, this invention selects the equal-weighted average of the monthly or seasonal average PAP returned during the modeling phase and the single-factor model results with the observed TCC passing the 90% confidence level as the final prediction result for that site; if there is no single-factor model that passes the significance test, the equal-weighted results of multiple single-factor models are used as the final prediction result for that site / grid point. This ensemble method fully considers the predictive power of different factors at each site / grid point.

[0094] S5: Real-time rolling forecasts are achieved by combining monthly or seasonal precipitation dynamics and statistics.

[0095] This invention, based on the variation patterns of predicted quantities, employs morphological closing operations to select key regions for predictive factors. It applies interannual increment methods and field information coupling methods to combine previous observational information with the high predictive efficiency of concurrent dynamic models, achieving rolling predictions that combine monthly or seasonal precipitation dynamics and statistics. Taking the prediction of June PAP at Chinese stations from December to February as an example: First, single-factor prediction models for different reporting months are established based on the monthly predictive factors from December to February and the June predictive factors predicted by the climate model in February. Specifically, the December prediction results are based solely on the December predictive factors, the January prediction results are the ensemble average of the December and January prediction results, and so on. Second, by performing a best-fit aggregation of the single-factor prediction models, the rolling predictions of June PAP at Chinese stations in December, January, and February can be obtained. By adjusting the selected months for predictive factors and the target months for precipitation prediction, rolling predictions of monthly or seasonal precipitation can be achieved.

[0096] This example provides a forecast instance for China's Pacific Precipitation Apparent (PAP) in June, July, and August 2024, reported starting in March (the start time of the consultation at the National Climate Center). The modeling period in this forecast instance is 1982-2023. Monthly precipitation data are from observations at 2400 stations in China. Predictive factor data are monthly reanalysis data from December to February, including sea surface air temperature, sea level pressure, outward longwave radiation, soil moisture, 200 hPa zonal wind, and sea ice concentration, as well as the 500 hPa geopotential height predicted from March by the US second-generation real-time climate prediction system (CFSv2) for June to August. In actual forecasts, only reanalysis data from February and earlier are available in March. The forecast performed well in the real-time monthly precipitation forecast for the summer of 2024, with prediction skill PS scores of 77, 79 and 73, respectively, all higher than the average score of 66 for the National Climate Center's operational forecasts from 1999 to 2021. The spatial anomaly correlation coefficients were 0.15, 0.26 and 0.13, respectively (the multi-year operational average ACC of the National Climate Center is ~0).

[0097] Example 2

[0098] This embodiment provides a monthly or seasonal precipitation prediction system based on interannual increments to implement the monthly or seasonal precipitation prediction method based on interannual increments in Embodiment 1 above. The system includes: a dataset construction module, a single-factor prediction model construction module, a prediction model error correction module, a prediction model optimization module, and a rolling prediction module.

[0099] In this embodiment, the dataset building module is used to build a potential predictor dataset;

[0100] In this embodiment, the single-factor prediction model building module is used to build a single-factor prediction model for the annual increment of the monthly or seasonal average precipitation anomaly percentage.

[0101] In this embodiment, the prediction model error correction module is used to correct the deviation of the annual increment of the monthly or seasonal average precipitation anomaly percentage predicted by the single-factor prediction model based on the singular value decomposition method.

[0102] In this embodiment, the prediction model selection module is used to correct the interannual increment of the average precipitation anomaly percentage predicted by each single-factor model and add it to the average precipitation anomaly percentage observed in the same period of the previous year to obtain the average precipitation anomaly percentage predicted by the single-factor model.

[0103] The monthly or seasonal average precipitation anomaly percentage and the correlation coefficient of the observation time are selected during the modeling stage. The results of single-factor models are screened by setting a confidence level threshold and then averaged equally to obtain the final prediction result. If there is no single-factor model that passes the significance test, the equal weight results of multiple single-factor models are used as the final prediction result.

[0104] In this embodiment, the rolling forecast module is used to adjust the selected month of the forecast factor and the target month of precipitation forecast to achieve rolling forecast of monthly or seasonal precipitation.

[0105] In this embodiment, the dataset construction module is used to construct a potential predictor dataset, specifically including:

[0106] Calculate the interannual increment of the monthly or seasonal average precipitation anomaly percentage for each meteorological station or grid point. Extract the time coefficients corresponding to the modes that satisfy the set proportion of variance explanation for the interannual increment of the monthly or seasonal average precipitation anomaly percentage through empirical orthogonal function decomposition. Select key regions for predictors based on the time coefficients corresponding to the selected modes. Rank the potential predictors for each climate variable according to the area size of the key regions. Compare the predictive efficacy of the potential predictors in the key regions and select the predictors.

[0107] Furthermore, the dataset building module is used to construct a dataset of potential predictors, specifically including:

[0108] Calculate the time correlation coefficients between grid points of the interannual increment fields of each selected predictor and predictor variable, the observed monthly or seasonal climate variables, and the climate model-predicted contemporaneous climate variables.

[0109] The time correlation coefficient is assigned the explained variance of the corresponding EOF mode by grid points with set confidence level thresholds, and the grid points without set confidence level thresholds are assigned 0, so as to obtain the spatial distribution of the explained variance for each mode. The spatial field of the explained variance for each mode of the predictor is accumulated and summed.

[0110] The values ​​greater than 0 in the cumulative explained variance space field are sorted from smallest to largest. A threshold for the key region is selected, and the morphological closing operation method is used to filter the grid points in the cumulative explained variance space field that are greater than or equal to the threshold to determine the key regions of the predictor.

[0111] By using morphological closing operations to screen grid points in the cumulative explained variance space field that are greater than or equal to a threshold, key regions of predictors are determined. Based on the size of the selected key regions, the main key regions are selected as potential predictors. The predictive efficacy of the potential predictors in the key regions is compared, and the predictors are selected.

[0112] In this embodiment, the single-factor prediction model construction module is used to construct a single-factor prediction model for the interannual increment of the monthly or seasonal average precipitation anomaly percentage, specifically including:

[0113] A single-factor prediction model for the interannual increment of the percentage of monthly or seasonal average precipitation anomalies was established using the field information coupling method.

[0114] Field reconstruction is performed by extracting predictors and predictor factors that satisfy a set ratio to explain the variance using empirical orthogonal function decomposition.

[0115] The optimal coupling variation of the spatial field of predictor and predictor after field reconstruction is extracted using the singular value decomposition method.

[0116] Construct the product matrix of the predictor and the predictor, and perform singular value decomposition on the product matrix;

[0117] For the time coefficients corresponding to the extracted optimal coupling variation, a single-factor downscaling prediction model is established using the linear regression method.

[0118] In this embodiment, the time coefficient corresponding to the extracted optimal coupling variation type specifically includes:

[0119]

[0120] Where k is the total number of SVD modes, Y(t,x) represents the interannual increment of PAP, X(t,x) represents the predictor factor, and Ri (x), W i (x) represent the interannual increment of the average precipitation anomaly percentage and the singular vector of the i-th mode of the predictor, respectively. i (t) and O i (t) represents the time dilation coefficient of the i-th SVD mode of the interannual increment of the average precipitation anomaly percentage and the predictor factor;

[0121] A single-factor downscaling prediction model is established using linear regression, specifically including:

[0122] For each mode, the corresponding time dilation coefficient S i (t) and time dilation factor O i (t) is subjected to univariate linear regression, based on the univariate linear regression coefficient α. i and residual ε i The time dilation coefficients of each SVD mode of the i-th mode at the predicted time t+Δt are obtained and expressed as:

[0123]

[0124] Based on the singular vector R of the i-th mode of the interannual increment of the average precipitation anomaly percentage i Based on the prediction results of (x) and the time dilation coefficients of each SVD mode, the interannual increment of the predicted average precipitation anomaly percentage at time t+Δt is reconstructed and expressed as:

[0125]

[0126] in, This represents the interannual increment of the average precipitation anomaly percentage predicted at time t+Δt. The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention should be considered equivalent substitutions and are included within the protection scope of the present invention.

Claims

1. A method for predicting monthly or seasonal precipitation based on interannual increments, characterized in that, Includes the following steps: Construct a dataset of potential predictors; Construct a single-factor prediction model for the interannual increment of the percentage anomaly in monthly or seasonal average precipitation, specifically including: A single-factor prediction model for the interannual increment of the percentage of monthly or seasonal average precipitation anomalies was established using the field information coupling method. Field reconstruction is performed by extracting predictors and predictor factors that satisfy a set ratio to explain the variance using empirical orthogonal function decomposition. The optimal coupling variation of the spatial field of predictor and predictor after field reconstruction is extracted using the singular value decomposition method. Construct the product matrix of the predictor and the predictor, and perform singular value decomposition on the product matrix; For the time coefficients corresponding to the extracted optimal coupling variation, a single-factor downscaling prediction model is established using the linear regression method. The time coefficients corresponding to the extracted optimal coupling variation type specifically include: ； ； Where k is the total number of SVD modes, This represents the annual increment of PAP. Indicates predictor factor, , Let represent the interannual increment of the average precipitation anomaly percentage and the singular vector of the i-th mode of the predictor, respectively. and The time dilation coefficient of the i-th SVD mode represents the interannual increment of the average precipitation anomaly percentage and the predictor factor; A single-factor downscaling prediction model is established using linear regression, specifically including: The time dilation coefficient for each mode and time dilation factor Perform univariate linear regression, based on the univariate linear regression coefficients and residual The prediction was obtained. The time dilation coefficients of each SVD mode at time i are expressed as: ； Based on the singular vector of the i-th mode of the interannual increment of the average precipitation anomaly percentage Based on the predicted time dilation coefficients of each SVD mode, the reconstruction is obtained. The interannual increment of the average precipitation anomaly percentage predicted at any given time is expressed as: ； in, express Interannual increment of the average precipitation anomaly percentage predicted at any given time; The singular value decomposition method was used to correct the deviation of the annual increment of the monthly or seasonal average precipitation anomaly percentage predicted by the single-factor prediction model. The average precipitation anomaly percentage predicted by each single-factor model is obtained by correcting the interannual increment of the average precipitation anomaly percentage observed in the same period of the previous year and then adding it to the average precipitation anomaly percentage predicted by the single-factor model. The monthly or seasonal average precipitation anomaly percentage and the correlation coefficient of the observation time are selected during the modeling stage. The results of single-factor models are screened by setting a confidence level threshold and then averaged equally to obtain the final prediction result. If there is no single-factor model that passes the significance test, the equal weight results of multiple single-factor models are used as the final prediction result. Adjusting the selected months for predictive factors and the target months for precipitation forecasting enables rolling monthly or seasonal precipitation forecasts.

2. The method for predicting monthly or seasonal precipitation based on interannual increments according to claim 1, characterized in that, Constructing a potential predictor dataset specifically includes: Calculate the interannual increment of the monthly or seasonal average precipitation anomaly percentage for each meteorological station or grid point. Extract the time coefficients corresponding to the modes that satisfy the set proportion of variance explanation for the interannual increment of the monthly or seasonal average precipitation anomaly percentage through empirical orthogonal function decomposition. Select key regions for predictors based on the time coefficients corresponding to the selected modes. Rank the potential predictors for each climate variable according to the area size of the key regions. Compare the predictive efficacy of the potential predictors in the key regions and select the predictors. The key regions of the predictor are selected based on the time coefficients corresponding to the selected modes, specifically including: Calculate the time correlation coefficients between grid points of the interannual increment fields of each selected predictor and predictor variable, the observed monthly or seasonal climate variables, and the climate model-predicted contemporaneous climate variables. The time correlation coefficient is assigned the explained variance of the corresponding EOF mode by grid points with set confidence level thresholds, and the grid points without set confidence level thresholds are assigned 0, so as to obtain the spatial distribution of the explained variance for each mode. The spatial field of the explained variance for each mode of the predictor is accumulated and summed. The values ​​greater than 0 in the cumulative explained variance space field are calculated and sorted from smallest to largest. A threshold for the key region is selected, and the morphological closing operation method is used to filter the grid points in the cumulative explained variance space field that are greater than or equal to the threshold to determine the key regions of the predictor.

3. A monthly or seasonal precipitation prediction system based on interannual increments, characterized in that, include: The module includes: dataset construction module, single-factor prediction model construction module, prediction model error correction module, prediction model optimization module, and rolling prediction module. The dataset construction module is used to construct a potential predictor dataset; The single-factor prediction model construction module is used to construct a single-factor prediction model for the interannual increment of the monthly or seasonal average precipitation anomaly percentage, specifically including: A single-factor prediction model for the interannual increment of the percentage of monthly or seasonal average precipitation anomalies was established using the field information coupling method. Field reconstruction is performed by extracting predictors and predictor factors that satisfy a set ratio to explain the variance using empirical orthogonal function decomposition. The optimal coupling variation of the spatial field of predictor and predictor after field reconstruction is extracted using the singular value decomposition method. Construct the product matrix of the predictor and the predictor, and perform singular value decomposition on the product matrix; For the time coefficients corresponding to the extracted optimal coupling variation, a single-factor downscaling prediction model is established using the linear regression method. The time coefficients corresponding to the extracted optimal coupling variation type specifically include: ； ； Where k is the total number of SVD modes, This represents the annual increment of PAP. Indicates predictor factor, , Let represent the interannual increment of the average precipitation anomaly percentage and the singular vector of the i-th mode of the predictor, respectively. and The time dilation coefficient of the i-th SVD mode represents the interannual increment of the average precipitation anomaly percentage and the predictor factor; A single-factor downscaling prediction model is established using linear regression, specifically including: The time dilation coefficient for each mode and time dilation factor Perform univariate linear regression, based on the univariate linear regression coefficients and residual The prediction was obtained. The time dilation coefficients of each SVD mode at time i are expressed as: ； Based on the singular vector of the i-th mode of the interannual increment of the average precipitation anomaly percentage Based on the predicted time dilation coefficients of each SVD mode, the reconstruction is obtained. The interannual increment of the average precipitation anomaly percentage predicted at any given time is expressed as: ； in, express Interannual increment of the average precipitation anomaly percentage predicted at any given time; The prediction model error correction module is used to correct the deviation of the annual increment of the monthly or seasonal average precipitation anomaly percentage predicted by the single-factor prediction model based on the singular value decomposition method. The prediction model selection module is used to correct the interannual increment of the average precipitation anomaly percentage predicted by each single-factor model and add it to the average precipitation anomaly percentage observed in the same period of the previous year to obtain the average precipitation anomaly percentage predicted by the single-factor model. The monthly or seasonal average precipitation anomaly percentage and the correlation coefficient of the observation time are selected during the modeling stage. The results of single-factor models are screened by setting a confidence level threshold and then averaged equally to obtain the final prediction result. If there is no single-factor model that passes the significance test, the equal weight results of multiple single-factor models are used as the final prediction result. The rolling forecast module is used to adjust the selected months of the forecast factors and the target months of precipitation forecasting, so as to realize rolling forecasts of monthly or seasonal precipitation.

4. The monthly or seasonal precipitation prediction system based on interannual increments according to claim 3, characterized in that, The dataset construction module is used to construct a potential predictor dataset, specifically including: Calculate the interannual increment of the monthly or seasonal average precipitation anomaly percentage for each meteorological station or grid point. Extract the time coefficients corresponding to the modes that satisfy the set proportion of variance explanation for the interannual increment of the monthly or seasonal average precipitation anomaly percentage through empirical orthogonal function decomposition. Select key regions for predictors based on the time coefficients corresponding to the selected modes. Rank the potential predictors for each climate variable according to the area size of the key regions. Compare the predictive efficacy of the potential predictors in the key regions and select the predictors. The key regions of the predictor are selected based on the time coefficients corresponding to the selected modes, specifically including: Calculate the time correlation coefficients between grid points of the interannual increment fields of each selected predictor and predictor variable, the observed monthly or seasonal climate variables, and the climate model-predicted contemporaneous climate variables. The time correlation coefficient is assigned the explained variance of the corresponding EOF mode by grid points with set confidence level thresholds, and the grid points without set confidence level thresholds are assigned 0, so as to obtain the spatial distribution of the explained variance for each mode. The spatial field of the explained variance for each mode of the predictor is accumulated and summed. The values ​​greater than 0 in the cumulative explained variance space field are calculated and sorted from smallest to largest. A threshold for the key region is selected, and the morphological closing operation method is used to filter the grid points in the cumulative explained variance space field that are greater than or equal to the threshold to determine the key regions of the predictor.

Images