A method for interpreting temperature forecast products from numerical models based on spatiotemporal local models

Through the method based on the local model of time and space, the parameters and models of temperature forecast products are optimized, and the uncertainty and differences in the numerical model temperature forecast are solved, and the accuracy and consistency of temperature forecasts are improved.

CN113988351BActive Publication Date: 2024-11-08HAINAN INSTITUTE OF METEOROLOGICAL SCIENCE
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202010734191.X
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2020-07-27
Publication Date
2024-11-08
Estimated Expiration
2040-07-27

AI Technical Summary

Technical Problem

There are uncertainties and forecast differences in temperature forecasts in the existing numerical modes, and there are differences in resolution, initialization and data assimilation technologies in different modes, resulting in insufficient accuracy and consistency of temperature forecasts.

Method used

Using a method based on spatial and temporal model, the area is divided into sub-regions by determining the area to be released and the numerical model temperature forecast product, obtaining temperature observation data and historical forecast products, performing sample preprocessing and quality control, and selecting releases Using models, weight function models and space-time distance models, optimize parameters to improve the release accuracy of temperature forecast products.

Benefits of technology

Through the local model of time and space, the relationship between temperature forecast products and real-life situations is more accurately modeled, which improves the release level and accuracy of numerical model temperature forecasts and reduces forecast errors.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure QLYQS_1
    Figure QLYQS_1
  • Figure QLYQS_2
    Figure QLYQS_2
  • Figure BDA0002604303030000031
    Figure BDA0002604303030000031
Patent Text Reader

Abstract

The present invention discloses a method for interpreting and using a numerical model temperature forecast product based on a spatiotemporal local model, comprising the following steps: determining the level of the area to be interpreted and the temperature forecast product, and dividing the area to be interpreted into a number of sub-areas; constructing a sample set; controlling the sample quality; for each sub-area, selecting a corresponding interpretation model, a weight function model, a spatiotemporal distance model, an evaluation index, and a sliding training period, and optimizing or setting the relevant parameters; obtaining the numerical model temperature forecast product for the forecast period, and interpolating it to the required discrete grid; interpreting the numerical model temperature forecast product for the forecast period discrete grid using the selected interpretation model, weight function model, spatiotemporal distance model, sliding training period, and optimized or set parameters. The beneficial effects of the present invention are: using the spatiotemporal local model to more accurately model the relationship between the numerical model temperature forecast product and the actual situation that changes with time and space, and further improving the interpretation level of the numerical model temperature forecast product.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of intelligent grid forecasting in meteorology, and in particular to a method for interpreting and using a numerical model temperature forecast product based on a spatiotemporal local model. Background Art

[0002] Temperature is one of the most basic meteorological elements. Temperature has an important impact on the growth and development of natural vegetation and crops, human production and life, etc. Too low temperature can easily cause disasters such as chilling damage, freezing damage, and frost damage, while too high temperature can easily cause disasters such as high temperature heat damage and dry hot winds. Therefore, accurate forecasting of temperature is one of the main contents of weather forecasting and climate forecasting. With the development of earth detection technology (such as satellite remote sensing, meteorological radar), dynamic meteorology, information and communication technology, supercomputer technology, numerical computing technology, data assimilation technology, etc., the spatial and temporal scale and timeliness of numerical models for temperature forecasting are constantly expanding, and the resolution and accuracy are constantly improving. Numerical model temperature forecast products have become the basis for temperature forecasting.

[0003] However, due to the dynamic framework of numerical models, the physical processes are only approximations of the real atmosphere, the inevitable errors in the initial values ​​of the models obtained by data assimilation, the chaotic characteristics of atmospheric motion, etc., there are always uncertainties in numerical models. In addition, models from different countries differ in model resolution, model initialization, data assimilation technology and types of assimilated data, dynamic and physical parameterization schemes, and the characterization of terrain. Different models have different forecasting capabilities for temperatures in different regions. Therefore, correcting the temperature forecasted by numerical models or integrating the temperature forecasts of different numerical models has become an important means to improve the level of temperature forecasting.

[0004] At present, linear regression, multi-model ensemble average, Bayesian model average, Kalman filter, neural network and other methods are the mainstream methods for correcting numerical model temperature forecast products or integrating multi-model products. However, no matter what interpretation method or model is used, the selection of samples and how to determine the weight of samples are crucial. From the existing research, when using the above methods to correct or integrate model products, some only use samples from a single grid point or station, some use all samples of a certain area at the same time, and some assume that all samples have the same importance.

[0005] In order to further improve the objective interpretation level of numerical model temperature forecast products, the present invention proposes a method for interpreting numerical model temperature forecast products based on a spatiotemporal local model. Summary of the invention

[0006] The present invention is designed to solve the above technical problems and is a method for interpreting and using numerical model temperature forecast products based on a spatiotemporal local model.

[0007] The technical solution adopted by the present invention to solve its technical problem is:

[0008] A method for interpreting temperature forecast products of a numerical model based on a spatiotemporal local model, the method comprising the following steps:

[0009] Step 1: Determine the area to be interpreted and the level of the numerical model temperature forecast product, and divide the area to be interpreted into several sub-areas;

[0010] Step 2: Obtain the corresponding temperature observation data and temperature forecast products of historical numerical models, and construct a sample set of temperature "observation data, forecast products" or "observation data, forecast products, error in the starting time of forecast";

[0011] Step 3: Pre-process and quality control the samples in step 2 and store them for future use;

[0012] Step 4: For each sub-region, select the corresponding interpretation model, weight function model, spatiotemporal distance model, evaluation index and sliding training period, and use the sample set processed in step 3 to optimize the parameters of the spatiotemporal local model for the interpretation of temperature forecast products, i.e., the number of neighborhood samples N, the spatiotemporal distance model parameters and the weight function bandwidth, or set the values ​​of the above parameters based on experience;

[0013] Step 5: Obtain the temperature forecast product of the numerical model during the forecast period and interpolate it to the required discrete grid;

[0014] Step 6: For each sub-region, the discrete grid numerical model temperature forecast product in step 5 is interpreted using the interpretation model, weight function model, spatiotemporal distance model and sliding training period selected in step 4 and the optimized or set spatiotemporal local model parameters for temperature forecast product interpretation;

[0015] Step 7: Establish relevant business systems and repeat steps 2-6.

[0016] The method for interpreting and using a numerical model temperature forecast product based on a spatiotemporal local model, wherein the numerical model temperature forecast product in step 1 is a deterministic forecast, an ensemble average or a combination thereof derived from one or more models.

[0017] The method for interpreting temperature forecast products of numerical models based on a spatiotemporal local model, the temperature observation data in step 2 includes but is not limited to station observation data, buoy observation data, multi-source data fusion real-time data, and reanalysis data; the forecast products and the error of the start time of the forecast in step 2 can be vectors corresponding to the forecast products and the error of the start time of the forecast from multiple models.

[0018] The method for interpreting temperature forecast products of numerical models based on spatiotemporal local models, the preprocessing and quality control of samples in step 3 include but are not limited to eliminating redundant samples and abnormal samples.

[0019] The method for interpreting temperature forecast products of numerical models based on spatiotemporal local models, the interpretation models that can be used in step 4 include but are not limited to weighted linear regression, weighted partial least squares regression, weighted conditional probability density function, weighted lasso regression and their corresponding multi-scale models.

[0020] The method for interpreting temperature forecast products of numerical models based on spatiotemporal local models, the weight function models that can be used in step 4 include but are not limited to distance threshold model ①, Gaussian model ②, exponential model ③, and double square model ④.

[0021]

[0022]

[0023]

[0024]

[0025] Among them, d i For sample x i The space-time distance to the grid point x0 to be interpreted, r is the bandwidth of the weight function, W i (x i ) is the sample x i The weight of .

[0026] The method for interpreting temperature forecast products of numerical models based on spatiotemporal local models, the evaluation indicators that can be used in step 4 include but are not limited to mean absolute error, root mean square error, CRPS score, forecast accuracy or a combination thereof.

[0027] The method for interpreting a numerical model temperature forecast product based on a spatiotemporal local model, the spatiotemporal local model for interpreting the temperature forecast product in step 4 and step 6 is specifically as follows: according to the selected sliding training period, for any discrete grid point or sample point, a relationship between the temperature observation data and the forecast product or between the temperature observation data and the forecast product and the error in the starting time is constructed using the N samples in the surrounding spatiotemporal space and the corresponding interpretation model, weight function model, and spatiotemporal distance model, and this relationship is used to interpret the numerical model temperature forecast product of the grid point or sample point.

[0028] In the method for interpreting temperature forecast products using numerical models based on a spatiotemporal local model, the spatiotemporal distance model that can be used in interpreting the temperature forecast product using the spatiotemporal local model in steps 4 and 6 includes but is not limited to equation ⑤.

[0029]

[0030] Where d is the space-time position x0 and x i The distance between (u0,v0,t0), (u i ,v i ,t i ) are the space-time positions x0 and x i The space-time coordinates of ; μ is the space-time distance model parameter, which is determined based on prior knowledge or cross-validation method.

[0031] The method for interpreting and using a numerical model temperature forecast product based on a spatiotemporal local model, the business system in step 7 includes but is not limited to a data acquisition module, a sample pair construction module, a sample quality control module, a parameter optimization and setting module, a numerical model temperature forecast product interpolation module, a numerical model temperature forecast product interpretation module or a combination thereof.

[0032] The samples described in the present invention all include corresponding unified reference time and space coordinates, which are not explicitly represented for simplicity.

[0033] The beneficial effect of the present invention is that the relationship between the numerical model temperature forecast product and the actual situation that changes with time and space is more accurately modeled by adopting a spatiotemporal local model, thereby further improving the level of interpretation of the numerical model temperature forecast product. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 A schematic diagram of space-time distance and space-time neighborhood. DETAILED DESCRIPTION

[0035] The present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0036] like Figure 1 As shown, the present invention provides a method for interpreting and using a numerical model temperature forecast product based on a spatiotemporal local model, the method comprising the following steps:

[0037] Step 1: Determine the area to be interpreted and the level of the numerical model temperature forecast product, and divide the area to be interpreted into several sub-areas;

[0038] Step 2: Obtain the corresponding temperature observation data and temperature forecast products of historical numerical models, and construct a sample set of temperature "observation data, forecast products" or "observation data, forecast products, error in the starting time of forecast";

[0039] Step 3: Pre-process and quality control the samples in step 2 and store them for future use;

[0040] Step 4: For each sub-region, select the corresponding interpretation model, weight function model, spatiotemporal distance model, evaluation index and sliding training period, and use the sample set processed in step 3 to optimize the parameters of the spatiotemporal local model for the interpretation of temperature forecast products, i.e., the number of neighborhood samples N, the spatiotemporal distance model parameters and the weight function bandwidth, or set the values ​​of the above parameters based on experience;

[0041] Step 5: Obtain the temperature forecast product of the numerical model during the forecast period and interpolate it to the required discrete grid;

[0042] Step 6: For each sub-region, the discrete grid numerical model temperature forecast product in step 5 is interpreted using the interpretation model, weight function model, spatiotemporal distance model and sliding training period selected in step 4 and the optimized or set spatiotemporal local model parameters for temperature forecast product interpretation;

[0043] Step 7: Establish relevant business systems and repeat steps 2-6.

[0044] The method for interpreting and using a numerical model temperature forecast product based on a spatiotemporal local model, wherein the numerical model temperature forecast product in step 1 is a deterministic forecast, an ensemble average or a combination thereof derived from one or more models.

[0045] The method for interpreting temperature forecast products of numerical models based on a spatiotemporal local model, the temperature observation data in step 2 includes but is not limited to station observation data, buoy observation data, multi-source data fusion real-time data, and reanalysis data; the forecast products and the error of the start time of the forecast in step 2 can be vectors corresponding to the forecast products and the error of the start time of the forecast from multiple models.

[0046] The method for interpreting temperature forecast products of numerical models based on spatiotemporal local models, the preprocessing and quality control of samples in step 3 include but are not limited to eliminating redundant samples and abnormal samples.

[0047] The method for interpreting temperature forecast products of numerical models based on spatiotemporal local models, the interpretation models that can be used in step 4 include but are not limited to weighted linear regression, weighted partial least squares regression, weighted conditional probability density function, weighted lasso regression and their corresponding multi-scale models.

[0048] The method for interpreting temperature forecast products of numerical models based on spatiotemporal local models, the weight function models that can be used in step 4 include but are not limited to one or more of the distance threshold model ①, Gaussian model ②, exponential model ③, and bisquare model ④ (for multi-scale models).

[0049]

[0050]

[0051]

[0052]

[0053] Among them, d i For sample x i The space-time distance to the grid point x0 to be interpreted, r is the bandwidth of the weight function, W i (x i ) is the sample x i The weight of .

[0054] The method for interpreting temperature forecast products of numerical models based on spatiotemporal local models, the evaluation indicators that can be used in step 4 include but are not limited to mean absolute error, root mean square error, CRPS score, forecast accuracy or a combination thereof.

[0055] The objective interpretation method of the numerical model forecast product based on the spatiotemporal local model, the parameter optimization method that can be used in step 4 includes but is not limited to the grid search parameter optimization method and the random search parameter optimization method; taking the grid search parameter optimization method as an example: the possible intervals of all parameters in step 4 are discretized and combined; for each parameter combination, the selected interpretation model, weight function model, and spatiotemporal distance model are used to construct a relationship between the temperature observation data and the forecast product or the temperature observation data and the forecast product and the error of the starting time for each sample point in the sub-area of ​​the sliding training period, and use this relationship to interpret the numerical model temperature forecast product of the sample point, and store the interpretation result for future use; according to the interpretation result, the evaluation index of each parameter combination is calculated, and the optimal parameter combination is determined according to the evaluation index.

[0056] The method for interpreting a numerical model temperature forecast product based on a spatiotemporal local model, the spatiotemporal local model for interpreting the temperature forecast product in step 4 and step 6 is specifically as follows: according to the selected sliding training period, for any discrete grid point or sample point, a relationship between the temperature observation data and the forecast product or between the temperature observation data and the forecast product and the error in the starting time is constructed using the N samples in the surrounding spatiotemporal space and the corresponding interpretation model, weight function model, and spatiotemporal distance model, and this relationship is used to interpret the numerical model temperature forecast product of the discrete grid point or sample point.

[0057] In the method for interpreting temperature forecast products using numerical models based on a spatiotemporal local model, the spatiotemporal distance model that can be used in interpreting the temperature forecast product using the spatiotemporal local model in steps 4 and 6 includes but is not limited to equation ⑤.

[0058]

[0059] Where d is the space-time position x0 and x i The distance between (u0,v0,t0), (u i ,v i ,t i ) are the space-time positions x0 and x i The space-time coordinates of ; μ is the space-time distance model parameter, which is determined based on prior knowledge or cross-validation method.

[0060] The method for interpreting and using a numerical model temperature forecast product based on a spatiotemporal local model, the business system in step 7 includes but is not limited to a data acquisition module, a sample pair construction module, a sample quality control module, a parameter optimization and setting module, a numerical model temperature forecast product interpolation module, a numerical model temperature forecast product interpretation module or a combination thereof.

[0061] Example 1

[0062] Objective interpretation is conducted on the daily minimum temperature forecast product of Hainan Island with a fine grid (0.125°×0.125°) reported by the ECMWF model at 00:00 (UTC) on the first day.

[0063] Step 1: Divide Hainan Island into 18 sub-regions according to administrative divisions;

[0064] Step 2: Obtain the Hainan Island 0.05°×0.05° China Smart Grid Live Fusion Analysis Product (CLDAS) and the Hainan Island fine grid daily minimum temperature forecast product reported by the ECMWF model at 00:00 (UTC) on the first day of the historical period; construct a sample set of daily minimum temperature (observation data, forecast product);

[0065] Step 3: Preprocess and quality control the samples in step 2;

[0066] Step 4: For each sub-region, the interpretation model uses univariate weighted linear regression, the weight function model uses the Gaussian model (Formula ②), the spatiotemporal distance model uses Formula ⑤, the evaluation index uses the mean absolute error, the sliding training period is set to 10 days, and the sample set processed in Step 3 and the random search parameter optimization method are used to optimize the spatiotemporal local model parameters of the daily minimum temperature forecast product interpretation, namely the number of neighborhood samples N, the spatiotemporal distance model parameter μ, and the weight function bandwidth r;

[0067] Step 5: Obtain the ECMWF model fine-grid daily minimum temperature forecast product within the Hainan Island during the forecast period, and use bilinear interpolation to interpolate to a 0.05°×0.05° grid consistent with the China Smart Grid Live Analysis Product (CLDAS);

[0068] Step 6: For each sub-region, the interpretation model, weight function model, spatiotemporal distance model and sliding training period selected in step 4 and the optimized spatiotemporal local model parameters for the interpretation of the daily minimum temperature forecast product are used to interpret the daily minimum temperature forecast product interpolated in step 5;

[0069] Step 7: Establish relevant business systems and repeat steps 2-6.

[0070] Example 2

[0071] Objectively interpret the Hainan Island fine grid (0.125°×0.125°) temperature forecast product for the third hour starting from 00:00 (UTC) of each day by the ECMWF model.

[0072] Step 1: Divide Hainan Island into 18 sub-regions according to administrative divisions;

[0073] Step 2: Obtain the Hainan Island 0.05°×0.05° China Smart Grid Live Analysis Product (CLDAS) and the Hainan Island fine grid temperature forecast product for the third hour starting at 00:00 (UTC) of the ECMWF model in the same period of history; and refer to the method of Zeng Xiaoqing (Zeng Xiaoqing, Xue Feng, Zhao Ruixia, Zhao Shengrong. Comparative study of several gridded temperature rolling correction forecast schemes [J]. Meteorology, 2019, 45(07): 1009-1018.) to construct a temperature (observation data, forecast product, starting time error) sample set;

[0074] Step 3: Preprocess and quality control the samples in step 2;

[0075] Step 4: For each sub-region, the interpretation model uses a binary weighted linear regression, the weight function model uses a Gaussian model (Formula ②), the spatiotemporal distance model uses Formula ⑤, the evaluation index uses the mean absolute error, the sliding training period is set to 10 days, and the sample set processed in Step 3 and the random search parameter optimization method are used to optimize the parameters of the spatiotemporal local model for the interpretation of the temperature forecast product, namely, the number of neighborhood samples N, the spatiotemporal distance model parameter μ, and the weight function bandwidth r;

[0076] Step 5: Obtain the ECMWF model fine grid 3-hour temperature forecast product within Hainan Island during the forecast period, and use the bilinear interpolation method to interpolate it to a 0.05°×0.05° grid consistent with the China Smart Grid Live Analysis Product (CLDAS);

[0077] Step 6: For each sub-region, the temperature forecast product after interpolation in step 5 is interpreted using the interpretation model, weight function model, spatiotemporal distance model, sliding training period and optimized spatiotemporal local model parameters selected in step 4;

[0078] Step 7: Establish relevant business systems and repeat steps 2-6.

[0079] Example 3

[0080] Objective interpretation is conducted on the daily minimum temperature forecast product of Hainan Island with a fine grid (0.125°×0.125°) reported by the ECMWF model at 00:00 (UTC) on the first day.

[0081] Step 1: Divide Hainan Island into 18 sub-regions according to administrative divisions;

[0082] Step 2: Obtain the Hainan Island 0.05°×0.05° China Smart Grid Live Fusion Analysis Product (CLDAS) and the Hainan Island fine grid daily minimum temperature forecast product reported by the ECMWF model at 00:00 (UTC) on the first day of the historical period; construct a sample set of daily minimum temperature (observation data, forecast product);

[0083] Step 3: Preprocess and quality control the samples in step 2;

[0084] Step 4: For each sub-region, the interpretation model uses the weighted conditional probability density function, the weight function model uses the Gaussian model (Formula ②), the spatiotemporal distance model uses Formula ⑤, the evaluation index uses the CRPS score, the sliding training period is set to 10 days, and the sample set processed in Step 3 and the random search parameter optimization method are used to optimize the spatiotemporal local model parameters of the daily minimum temperature forecast product interpretation, namely the number of neighborhood samples N, the spatiotemporal distance model parameter μ, and the weight function bandwidth r;

[0085] Step 5: Obtain the ECMWF model fine-grid daily minimum temperature forecast product within the Hainan Island during the forecast period, and use bilinear interpolation to interpolate to a 0.05°×0.05° grid consistent with the China Smart Grid Live Analysis Product (CLDAS);

[0086] Step 6: For each sub-region, the interpretation model, weight function model, spatiotemporal distance model, sliding training period and optimized spatiotemporal local model parameters of the temperature forecast product interpretation selected in step 4 are used to interpret the daily minimum temperature forecast product after interpolation in step 5;

[0087] Step 7: Establish relevant business systems and repeat steps 2-6.

Claims

1. A method for interpreting temperature forecast products of numerical models based on a spatiotemporal local model, characterized by: The method comprises the following steps: Step 1: Determine the area to be interpreted and the level of the numerical model temperature forecast product, and divide the area to be interpreted into several sub-areas; Step 2: Obtain the corresponding temperature observation data and temperature forecast products of historical numerical models, and construct a sample set of temperature "observation data, forecast products" or "observation data, forecast products, error in the starting time of forecast"; Step 3: Pre-process and quality control the samples in step 2 and store them for future use; Step 4: For each sub-region, select the corresponding interpretation model, weight function model, spatiotemporal distance model, evaluation index and sliding training period, and use the sample set processed in step 3 to optimize the parameters of the spatiotemporal local model for the interpretation of temperature forecast products, i.e., the number of neighborhood samples N, the spatiotemporal distance model parameters and the weight function bandwidth, or set the values ​​of the above parameters based on experience; The interpretation models include weighted linear regression, weighted partial least squares regression, weighted conditional probability density function, weighted lasso regression and their corresponding multi-scale models; The weight function models include distance threshold model ①, Gaussian model ②, exponential model ③, and bi-square model ④; Among them, d i For sample x i The space-time distance to the grid point x0 to be interpreted, r is the bandwidth of the weight function, W i (x i ) is the sample x i The weight of The evaluation index includes mean absolute error, root mean square error, CRPS score, prediction accuracy or a combination thereof; The temporal and spatial local model for interpreting the temperature forecast product is specifically as follows: according to the selected sliding training period, for any discrete grid point or sample point, using the surrounding N temporal and spatial samples and the corresponding interpretation model, weight function model, and temporal and spatial distance model, a relationship between the temperature observation data and the forecast product or between the temperature observation data and the forecast product and the error of the starting time is constructed, and this relationship is used to interpret the numerical model temperature forecast product of the discrete grid point or sample point; The space-time distance model includes formula ⑤; Where d is the space-time position x0 and x i The distance between (u0,v0,t0), (u i ,v i ,t i ) are the space-time positions x0 and x i The space-time coordinates of ; μ is the space-time distance model parameter, which is determined based on prior knowledge or cross-validation method; Step 5: Obtain the temperature forecast product of the numerical model during the forecast period and interpolate it to the required discrete grid; Step 6: For each sub-region, the discrete grid numerical model temperature forecast product in step 5 is interpreted using the interpretation model, weight function model, spatiotemporal distance model and sliding training period selected in step 4 and the optimized or set spatiotemporal local model parameters for temperature forecast product interpretation; Step 7: Establish relevant business systems and repeat steps 2-6; The business system in step 7 includes a data acquisition module, a sample pair construction module, a sample quality control module, a parameter optimization and setting module, a numerical model temperature forecast product interpolation module, and a numerical model temperature forecast product interpretation module.

2. The method for interpreting and using a numerical model temperature forecast product based on a spatiotemporal local model according to claim 1, characterized in that: The numerical model temperature forecast product in step 1 is a deterministic forecast and ensemble average derived from one or more models.

3. The method for interpreting and using a numerical model temperature forecast product based on a spatiotemporal local model according to claim 1, characterized in that: The temperature observation data in step 2 include station observation data, buoy observation data, multi-source data fusion real-time data, and reanalysis data; the forecast product and the error of the start time of reporting in step 2 can be a vector corresponding to the forecast products and the error of the start time of reporting of multiple models.

4. The method for interpreting and using a numerical model temperature forecast product based on a spatiotemporal local model according to claim 1, characterized in that: The sample preprocessing and quality control in step 3 includes eliminating redundant samples and abnormal samples.