A high-resolution temperature prediction method

Abstract

The present invention provides a high-resolution temperature prediction method, including data preparation and high-precision, high-spatial-resolution temperature prediction, wherein the data preparation includes collecting the temperature, air pressure, relative humidity, specific humidity and dew point temperature products for the next 3-24 hours released by GFS each time, with a time resolution of 3 hours and a spatial resolution of 25km respectively. The GFS forecast products are interpolated into grid data with a spatial resolution of 250m using an inverse distance weighted interpolation algorithm. The present invention implements a method for refined temperature forecasting, and the method can be used to obtain a high-precision, high-spatial-resolution temperature grid forecast.

Description

Technical Field

[0001] The invention belongs to the technical field of meteorological forecasting, and in particular relates to a high-resolution temperature forecasting method. Background Art

[0002] In recent years, numerical forecasting technology has developed rapidly, and the accuracy of temperature forecasts has been further improved. However, there is still a certain gap compared to the current society's demand for forecast refinement and accuracy.

[0003] The method of obtaining refined forecast temperature based on large-scale numerical forecast products is mainly to downscale the forecast products. The current method mainly downscales the forecast products at a single moment. Generally speaking, temperature has the property of time continuity, and there is a certain correlation between temperatures close in time. The same GFS forecast temperature also has a certain correlation between temperatures at similar moments. The current method does not take into account the temporal correlation of forecast products released by GFS at different forecast times, resulting in the inability to give full play to the value of the forecast products. Summary of the invention

[0004] The purpose of the present invention is to solve the defects of the above-mentioned prior art and provide a high-resolution temperature prediction method to realize a method for refined temperature forecasting. By using this method, high-precision and high-spatial-resolution temperature grid prediction can be obtained.

[0005] The present invention combines the advantages of different neural networks to deeply mine the information in numerical forecast products from the perspective of space and time and integrate satellite data to make high-precision and high-spatial-resolution temperature forecasts, providing a new method for refined temperature forecasts based on numerical forecast products.

[0006] The present invention adopts the following technical solution:

[0007] A high-resolution temperature prediction method comprises the following steps:

[0008] 1. Data preparation

[0009] Collect the temperature (T) for the next 3-24 hours published by GFS each time gfs), air pressure (P), relative humidity (RH), specific humidity (SH) and dew point temperature (DPT) products, with temporal and spatial resolutions of 3 hours and 25km respectively. The GFS forecast products are interpolated into grid data with a spatial resolution of 250m using the inverse distance weighted interpolation algorithm. The national vegetation cover index (NDVI) products with a spatial resolution of 250m and a temporal resolution of 16 days provided by Modis are collected and spliced ​​to obtain complete data for a certain area. The digital elevation (DEM) with a spatial resolution of 90m provided by the Shuttle Radar Topography Mission (SRTM) is collected and the spatial resolution of the digital elevation (DEM) is processed to 250m by simple average resampling. The measured temperature (T sta ), the station temperature is used as the real temperature for the target variable of the temperature forecast model. Finally, the data is cleaned and outliers are removed.

[0010] 2. Research on high-precision and high-spatial-resolution temperature prediction

[0011] (2.1) Using the temperature forecast for 3-24 hours published by GFS (T gfs ), air pressure (P), relative humidity (RH), dew point temperature (DPT) and specific humidity (SH) products, temperature observed by meteorological stations, vegetation cover index (NDVI) and digital elevation (DEM), and matched the data set required for the forecast model. Then, the advantages of feedforward neural network (FNN) and long short-term memory neural network (LSTM) were combined to develop a refined temperature forecast model based on GFS forecast data; the temperature T released by GFS with a forecast period of 3-24 hours was converted into gfs , air pressure P, relative humidity RH, dew point temperature DPT and specific humidity SH products, measured temperature T sta , vegetation cover index NVDI, digital elevation DEM and longitude and latitude as the input of the model, temperature T gfs As the output of the model, when the temperature forecast model is used to forecast the temperature, its mathematical expression is as follows:

[0012] T pre =f(T gfs ,P,RH,SH,DPT,NDVI,DEM,LON,LAT) (1)

[0013] (2.2) Use the Pearson correlation coefficient to analyze the correlation between various parameters and measured temperature to understand the impact of each variable on the model itself;

[0014] (2.3) Establish a refined temperature forecast model using feedforward neural network (FNN) and long short-term memory neural network (LSTM)

[0015] The model can be used to forecast high spatial resolution and high precision temperatures for the next 3, 6, 9, 12, 15, 18, 21 and 24 hours based on the time when the GFS data is released, forming a 250m high-resolution temperature grid forecast. The feedforward neural network (FNN) has a strong performance in enhancing the correlation between a single input feature and the output. The long short-term memory neural network (LSTM) is good at processing time series problems and can better extract time correlation information that FNN cannot learn. Therefore, the input data of the model is divided into two parts: the long short-term memory neural network (LSTM) part and the feedforward neural network (FNN) part.

[0016] (a) The 3-24 hour data released by GFS is regarded as a time series with a time interval of 3 hours and put into the LSTM neural network to let the network learn the information hidden in the time series.

[0017] (b) Considering that the GFS forecast data itself has a large error, the feedforward neural network (FNN) part not only needs to process the vegetation cover index (NDVI), digital elevation (DEM) and LAT, LON, but also needs to hand over the data with time series characteristics to the feedforward neural network (FNN) for synchronous processing.

[0018] (c) Finally, a feedforward neural network (FNN) is used to combine the input part of the feedforward neural network (FNN) and the long short-term memory neural network (LSTM) to integrate the effective information extracted from these two parts and output the predicted temperature.

[0019] (2.4) Use the measured temperature data from meteorological stations to evaluate the accuracy of temperature forecasts.

[0020] Beneficial effects of the present invention:

[0021] The present invention improves the accuracy of numerical temperature forecast products.

[0022] The present invention improves the spatial resolution of numerical temperature forecast products.

[0023] The present invention realizes high-precision, high-spatial-resolution (250m) gridded temperature forecasting. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 A roadmap for temperature forecast models;

[0025] Figure 2(a)-Figure 2(h) To compare the temperature predicted by the model with the measured data from the meteorological stations;

[0026] Table 1 shows the parameter information used in the given forecast model. DETAILED DESCRIPTION

[0027] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention is described clearly and completely below. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0028] Example

[0029] The measured data of near-surface atmospheric temperature were collected from 97 meteorological stations in Hunan Province in the summer of 2021 (June-August), and the temperature, air pressure, relative humidity, specific humidity and dew point temperature products with a forecast period of 3-24 hours released by GFS, Modis NDVI products and DEM were collected simultaneously. Table 1 gives the main information of the data used in the present invention, including the temporal resolution, spatial resolution and source of the data. Among them, the temporal resolution of the GFS data product is 3 hours, the spatial resolution is 0.25°, the temporal resolution of NDVI is 16 days, the spatial resolution is 250m, and the spatial resolution of DEM is 90m. During model training, the datasets of 9, 10, 19, 20 and the last two days of each month in Hunan Province from June to August 2021 were used for model testing, and the datasets at other times were used for model training.

[0030] The collected data is preprocessed, including spatial interpolation, data cleaning and data standardization. Then, the pixels closest to the station are selected according to the longitude and latitude of the station, and they are regarded as the same location. Then, the corresponding output data in the pixel is matched with the station temperature to form the data set required for modeling. Then, the data set is divided into two parts and processed by LSTM and FNN respectively. Finally, FNN is used to output the forecast result. The whole temperature forecast is as follows: Figure 1 As shown. For the constructed temperature forecast model, the data sets of 9th, 10th, 19th, 20th and the last two days of each month in Hunan Province from June to August 2021 were selected for verification. The verification results show that the forecast model constructed by the present invention can use the time when the GFS data is released as the reference moment to predict the gridded temperature with high spatial resolution and high precision for the next 3, 6, 9, 12, 15, 18, 21 and 24 hours, as shown in the figure. Figure 2(a)-Figure 2(h) As shown in Figure 2, where ah is the two-dimensional histogram of the eight forecast times.

[0031] Table 1

[0032]

[0033] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A high-resolution temperature prediction method, characterized in that: The following steps are involved: Step 1. Collect the T of the next 3-24 hours published by GFS each time gfs , P, RH, SH and DPT products, with time and spatial resolutions of 3 hours and 25 km respectively. The GFS forecast products are interpolated into grid data with a spatial resolution of 250 m using the inverse distance weighted interpolation algorithm. The NDVI products with a spatial resolution of 250 m and a temporal resolution of 16 days provided by Modis are collected and spliced ​​to obtain complete data for a certain area. The DEM with a spatial resolution of 90 m provided by the Space Shuttle Radar Topography Mission is collected and resampled to a spatial resolution of 250 m by simple average. The T sta , the station temperature is used as the real temperature for the target variable of the temperature forecast model, and finally the data is cleaned to remove outliers; Where T gfs is air temperature, P is air pressure, RH is relative humidity, SH is specific humidity, DPT is dew point temperature, NDVI is vegetation cover index, DEM is digital elevation, T sta is the measured temperature; Step 2. High-precision and high-spatial-resolution temperature forecast (2.1) Using the forecast time of 3-24 hours issued by GFS gfs , P, RH, DPT and SH products, T sta , NVDI and DEM, matching the data sets required for the forecast model, combining the feedforward neural network and the long short-term memory neural network to establish a refined temperature forecast model based on GFS forecast data, converting the T forecast time of 3-24 hours issued by GFS into gfs , P, RH, DPT and SH products, T sta , NVDI, DEM and longitude and latitude are used as the input of the model. sta As the output of the model, when the temperature forecast model is used for temperature forecasting, its mathematical expression is as follows: （1）； (2.2) Use the Pearson correlation coefficient to analyze the correlation between various parameters and the measured temperature, and understand the influence of each variable on the model itself; (2.3) Using FNN and LSTM to establish a refined temperature forecast model A refined temperature forecast model is established through FNN and LSTM to forecast high spatial resolution and high precision temperature for the next 3, 6, 9, 12, 15, 18, 21 and 24 hours based on the time when GFS data is released, forming a 250m high-resolution temperature grid forecast. The input data of the model is divided into two parts, the LSTM part and the FNN part. (a) The 3-24 hour data released by GFS is treated as a time series with a time interval of 3 hours and put into the LSTM neural network to let the network learn the information hidden in the time series; (b) NVDI, DEM, LAT, LON, and data with time series characteristics are sent to FNN for synchronous processing; (c) Finally, FNN is used to combine the input FNN and LSTM to integrate the effective information extracted from the two parts and output the predicted temperature; (2.4) Use the measured temperature data from meteorological stations to evaluate the accuracy of temperature forecasts.

Images