Temperature correction method and device based on full-connection neural network
The EMFCNN temperature model, constructed using a fully connected neural network, solves the problems of high computational complexity and error accumulation in traditional temperature forecast models. It achieves multi-factor fusion and feature selection, improving the accuracy and timeliness of temperature forecasts, and showing significant advantages, especially in temperature forecasts in Guangdong Province.
Patent Information
- Application Number
- CN202510674853.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-23
- Publication Date
- 2026-01-30
AI Technical Summary
Existing technologies, especially traditional numerical weather prediction models, suffer from high computational complexity and accumulated model errors, leading to insufficient timeliness and accuracy in short-term and nowcasting. Statistical learning-based methods have limitations in handling multivariate nonlinear relationships, making it difficult to construct reasonable input feature sets to characterize multi-scale atmospheric processes and to integrate model forecast products with real-world observation data to improve model generalization capabilities.
A fully connected neural network is used to construct an EMFCNN temperature model. The model is constructed by concatenating station number embedding encoding with model forecast elements and combining it with the local temperature variation equation to screen high-impact elements. The model is trained using the Adam optimizer and the Inverse Time Decay learning rate decay strategy to achieve deep fusion and feature selection of multimodal meteorological data and construct a multi-element fusion temperature forecast model.
It significantly reduces the mean absolute error between forecast values and actual conditions, improves the accuracy and timeliness of temperature forecasts, and shows higher accuracy and stability, especially in the 24-hour maximum temperature, 2-hour minimum temperature and hourly temperature forecasts in Guangdong Province, overcoming the systematic bias of traditional models.
Smart Images
Figure CN121434601A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of weather forecasting, and in particular to a temperature correction method and device based on a fully connected neural network. BACKGROUND
[0002] With the intensification of global climate change, extreme weather events occur frequently, and the accuracy and refinement of weather forecasting are increasingly urgent. Temperature prediction, as one of the key elements of meteorological services, directly affects disaster warning, agricultural production, energy scheduling, and public life in many fields. Although traditional numerical weather prediction (NWP) has significant advantages in physical mechanism modeling, its high computational complexity and accumulated model errors limit the timeliness and accuracy of short-term and near-term forecasts. At the same time, although post-processing methods based on statistical learning can partially correct model biases, they still have limitations in dealing with multivariate nonlinear relationships.
[0003] In recent years, machine learning technology, with its powerful nonlinear modeling capability, has provided a new technical path for the refinement of meteorological elements. Fully connected neural network (FCNN), as a classic deep learning model, can effectively integrate multi-source heterogeneous data and mine complex relationships between high-dimensional features, and has shown potential in meteorological element downscaling and model error correction. However, for regional temperature prediction tasks, how to reasonably construct input feature sets to represent atmospheric multi-scale processes and how to integrate model prediction products and real-time observation data to improve model generalization ability are still technical difficulties to be solved. SUMMARY
[0004] The purpose of the present application is to at least solve one of the deficiencies of the prior art, and to provide a temperature correction method based on a fully connected neural network.
[0005] To achieve the above-mentioned purpose, the present application adopts the following technical solutions: Specifically, a temperature correction method based on a fully connected neural network is proposed, which includes the following: Constructing an EMFCNN temperature model based on a fully connected neural network; Training the EMFCNN temperature model, and obtaining an EMFCNN_24h maximum temperature model, an EMFCNN_24h minimum temperature model, and an EMFCNN_hourly temperature model after completing the training; Performing temperature correction based on the EMFCNN_24h maximum temperature model, the EMFCNN_24h minimum temperature model, and the EMFCNN_hourly temperature model; Specifically, the formation process of the training set, the validation set, and the test set for training the EMFCNN temperature model is as follows, obtain CLDAS and ECMWF THIN data of a target position, and perform data cleaning on the CLDAS and ECMWF THIN data to obtain cleaned CLDAS and ECMWF THIN data; select high-impact elements from the cleaned CLDAS and ECMWF THIN data according to local changes in air temperature to obtain CLDAS high-impact element data and ECMWF THIN high-impact element data; unify the ECMWF THIN high-impact element data to the CLDAS grid system based on a Cressman interpolation algorithm; use the CLDAS high-impact element data and the ECMWF THIN high-impact element data of the CLDAS grid system as a data set, and divide the data set based on a time sequence to obtain a training set, a validation set, and a test set for training; Specifically, the station number is subjected to Embedding coding, and then is spliced with the mode prediction elements before being input into a full-connection neural network to output 24-hour maximum temperature, 24-hour minimum temperature, and 1-hourly point temperature prediction results. The station number includes CLDAS grid points and G1099 stations.
[0006] Further, specifically, the EMFCNN temperature model is trained using an Adam optimizer and an Inverse Time Decay learning rate decay strategy, and the specific settings are as follows, The initial learning rate is 0.01, the learning rate is decayed once every epoch, the decay index is set to 0.1, the early stopping rule based on the validation set performance is used, that is, if the validation loss does not improve within 10 consecutive epochs, the training process is stopped, and the MAE is used as the validation loss function.
[0007] Further, specifically, The grid resolution of the obtained CLDAS data is 0.0625°x0.0625°, the time resolution is hourly, and the variables include 2-meter temperature, The grid resolution of the obtained ECMWF THIN data is 0.125°x0.125°, the time resolution is hourly, and the variables include 2-meter air temperature, 10-meter east-west wind, 10-meter north-south wind, air pressure, 925-hPa relative humidity, 925-hPa specific humidity, 925-hPa north-south wind, 925-hPa east-west wind, 925-hPa temperature, and 850-hPa relative humidity.
[0008] Further, specifically, the CLDAS and ECMWFTHIN data are cleaned by a systematic detection and correction mechanism to process abnormal values, error records and missing fields.
[0009] Further, specifically, The CLDAS data includes maximum temperature, minimum temperature, and hourly temperature observation data at whole points; The ECMWFTHIN high-impact factor data includes forecast day key meteorological factors, including 2-meter air temperature, 10-meter / 925-hPa wind field components, sea level pressure, 925-hPa temperature and humidity parameters, and 850-hPa relative humidity. Among them, the high-impact factor for the maximum temperature is the forecast day 14:00 key meteorological factor, the high-impact factor for the minimum temperature is the forecast day 04:00 key meteorological factor, and the high-impact factor for the hourly temperature at whole points is the forecast day key meteorological factor at each time period.
[0010] Further, specifically, the training set, validation set and test set for training are obtained based on time series division of the data set, including, The data from April to September in 2019-2022 in the data set is set as the training set, the data in the same period in 2023 is set as the validation set, and the data in the same period in 2024 is set as the test set, forming a complete meteorological modeling data architecture.
[0011] Further, specifically, the high-impact factors are selected from the cleaned CLDAS and ECMWFTHIN data according to the local variation of air temperature, including, A simplified equation for the local variation of air temperature is constructed, , Among them, is the local variation of temperature, i.e., the temperature change per unit time, is the horizontal temperature advection term, i.e., the temperature change caused by the horizontal transport of heat by the wind field, is the adiabatic change caused by vertical motion, is the non-adiabatic heating term, i.e., the influence of latent heat release; Therefore, firstly, the ground pressure, air temperature, U wind, V wind, total cloud amount, low cloud amount, precipitation, relative humidity and specific humidity of the ground, 925 hPa and 850 hPa are selected as the initial screening factors; then, the correlation coefficients of the factors and the highest temperature and the lowest temperature are calculated, and the first 10 factors with the highest correlation coefficients and passing the 90% reliability test are selected as the high-impact factors; the factors are 2-meter air temperature, 10-meter east-west wind, 10-meter south-north wind, ground pressure, 925 hPa relative humidity, 925 hPa specific humidity, 925 hPa south-north wind, 925 hPa east-west wind, 925 hPa temperature and 850 hPa relative humidity.
[0012] The application further provides a temperature correction device based on a full-connection neural network, which comprises the following: a model construction module, configured to construct an EMFCNN temperature model based on a full-connection neural network; a model training module, configured to train the EMFCNN temperature model, and obtain an EMFCNN_24h highest temperature model, an EMFCNN_24h lowest temperature model and an EMFCNN_per-hour temperature model after the training is completed; a temperature correction module, configured to perform temperature correction based on the EMFCNN_24h highest temperature model, the EMFCNN_24h lowest temperature model and the EMFCNN_per-hour temperature model; Specifically, the forming process of the training set, the verification set and the test set for training the EMFCNN temperature model is as follows, obtain CLDAS and ECMWFTHIN data of a target position, and perform data cleaning on the CLDAS and ECMWFTHIN data to obtain cleaned CLDAS and ECMWFTHIN data; select high-impact factors from the cleaned CLDAS and ECMWFTHIN data according to local temperature change, to obtain CLDAS high-impact factor data and ECMWFTHIN high-impact factor data; unify the ECMWFTHIN high-impact factor data to the CLDAS grid system based on a Cressman interpolation algorithm; use the CLDAS high-impact factor data and the ECMWFTHIN high-impact factor data of the CLDAS grid system as a data set, and divide the data set based on time sequence to obtain a training set, a verification set and a test set for training; Specifically, the station number is subjected to Embedding coding, and then is spliced with the mode prediction element, and then is accessed into the full connection neural network to output the 24h maximum temperature, 24h minimum temperature and 1h temperature prediction result. The station number includes the CLDAS grid points and the G1099 station.
[0013] The present application has the following beneficial effects: The present application provides a temperature correction method based on a full connection neural network, On the one hand, a multi-modal meteorological data fusion model (EMFCNN) based on a full connection neural network is provided, the spatial features and the meteorological elements are deeply fused through the splicing operation of the station number Embedding coding and the mode prediction element, the limitation of the single data source of the traditional numerical mode is broken through, and the grid station features and the numerical prediction data are jointly modeled; On the other hand, a prediction element screening strategy based on the local change equation of air temperature is provided: the key meteorological element combination of the prediction day is selected through the physical equation constraint, including the temperature and humidity parameters of the 925hPa layer and the wind field of multiple height layers. The scientific nature of feature selection is improved, and the blindness of the traditional machine learning method is avoided. BRIEF DESCRIPTION OF DRAWINGS
[0014] The above and other features of the present disclosure will become more apparent from the following detailed description taken in conjunction with the accompanying drawings, in which like reference numerals denote like elements or features in the accompanying drawings. As apparent from the following description, the drawings are only some embodiments of the present disclosure, and other drawings can be obtained from these drawings by those skilled in the art without creative effort. Figure 1 The flowchart of the temperature correction method based on the full connection neural network is shown. Figures 2(a), (b) and (c) respectively show the highest temperature of the national station in Guangdong Province on September 20, 2023 (20:00-20:00), the highest temperature of the national station in Guangdong Province on September 21, 2023 (20:00-20:00), and the highest temperature variation distribution of the national station in Guangdong Province on September 21, 2023. Figure 3 The weight distribution of each feature of the highest temperature model in the verification set when the EMFCNN temperature model is verified by the SHAP method is shown. Figure 4 The model network structure diagram of the full connection neural network involved in the present application is shown. DETAILED DESCRIPTION
[0015] The concept, specific structure and generated technical effects of the present application will be described clearly and completely in combination with embodiments and drawings to fully understand the purposes, schemes and effects of the present application. It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict. The same reference signs used in the drawings indicate the same or similar parts.
[0016] Considering that the current various numerical modes still have a large deviation from the actual temperature for the target position, especially the 24h maximum temperature, 2h minimum temperature and 1h temperature prediction in Guangdong Province in spring and summer, especially in the process of temperature rise, the strength and range of high temperature prediction are still not satisfactory, Embodiment 1, refer to Figure 1 The present application proposes a temperature correction method based on a full connection neural network, proposes a multi-element fusion temperature prediction model EMFCNN temperature model based on a full connection neural network, reduces the average absolute error between the predicted value and the actual value, and realizes more accurate temperature prediction service for the target position, especially in Guangdong, specifically, the method comprises the following: Constructing an EMFCNN temperature model based on a full connection neural network; Training the EMFCNN temperature model, and obtaining a 24h maximum temperature, 24h minimum temperature and 1h temperature neural network model after completing the training, which are named as EMFCNN_24h maximum temperature model, EMFCNN_24h minimum temperature model and EMFCNN_1h temperature model respectively; Based on the EMFCNN_24h maximum temperature model, the EMFCNN_24h minimum temperature model and the EMFCNN_1h temperature model, the temperature is corrected; wherein the temperature model is a post-processing model of the mode, and the prediction time limit is 12-36h in the future, Specifically, the formation process of the training set, the verification set and the test set for training the EMFCNN temperature model is as follows, Obtain the CLDAS and ECMWFTHIN data of the target position, and clean the CLDAS and ECMWFTHIN data to obtain cleaned CLDAS and ECMWFTHIN data; According to the local change of temperature, select high-impact elements from the cleaned CLDAS and ECMWFTHIN data to obtain CLDAS high-impact element data and ECMWFTHIN high-impact element data; Based on the Cressman interpolation algorithm, the ECMWFTHIN high-impact element data is unified to the CLDAS grid system; The CLDAS high-impact element data and the ECMWFTHIN high-impact element data of the CLDAS grid system are taken as a data set, and the data set is divided into a training set, a verification set and a test set based on a time sequence for training; Specifically, the station number is encoded by Embedding, and then spliced with the mode prediction elements, and then input into the full connection neural network to output the 24h maximum temperature, 24h minimum temperature and 1h temperature prediction results (corresponding to the EMFCNN_24h maximum temperature model, the EMFCNN_24h minimum temperature model and the EMFCNN_1h temperature model respectively). The station number includes the CLDAS grid points and the G1099 station.
[0017] The model network structure of the full connection neural network is as follows Figure 4 : The Embedding encoding information of the station number is in the left black box, and the mode prediction element information is in the left pink long box, and the two are spliced and then input into the full connection neural network.
[0018] In addition, the EMFCNN temperature model is verified by SHAP, and SHAP (SHapley Additive exPlanations) is a method for explaining the output of a machine learning model, which quantifies the contribution of each feature to the prediction result based on the Shapley value in cooperative game theory. The weight distribution of each feature of the maximum temperature model is calculated by using the SHAP method Figure 3 The weight of each prediction element in the model falls within the expected reasonable range, which shows that the performance of the model meets our expected target. Specifically, the influence degree of each prediction element on the final prediction result neither overestimates nor underestimates, which shows that the model has good stability and reliability in processing these elements. This consistency not only verifies the effectiveness of the model design, but also enhances our confidence in the output of the model, because the importance score of each element reflects its role in the actual meteorological process, which is consistent with the domain knowledge.
[0019] The comprehensive test and evaluation results of the EMFCNN temperature model are shown in Table 2 (MAE score results and improvement rate), and the EMFCNN temperature model shows superior performance in both grid point prediction and station prediction.
[0020] Table 2
[0021] In terms of grid point prediction, on the 2023 validation set, the MAE of the highest temperature, the lowest temperature, and the hourly temperature prediction of the EMFCNN temperature model was reduced by 36.2%, 23.0%, and 40.3% respectively compared with the traditional objective method (mos36h), especially in the hourly temperature prediction. On the 2024 test set, although the MAE of some tasks increased slightly (such as the grid MAE of the highest temperature from 0.7872℃ to 0.834℃), the improvement rate remained at a high level (such as the grid improvement rate of the lowest temperature of 20.6%), indicating that the model has strong adaptability and robustness to changes in data distribution. In addition, compared with the previous traditional objective method, the performance of EMFCNN in the test set has a smaller decay, further highlighting its technical advantages.
[0022] In terms of site prediction, the EMFCNN temperature model also performed well. On the 2023 validation set, the improvement rates of the highest temperature, the lowest temperature, and the hourly temperature prediction were 9.2%, 13.3%, and 22.5% respectively, especially in the hourly temperature prediction, the improvement rate was more than 20%. On the 2024 test set, the site prediction improvement rate of EMFCNN did not decrease significantly, and even some tasks improved slightly (such as the site improvement rate of the highest temperature from 9.2% to 9.8%), indicating that the model has strong stability and generalization ability in complex site tasks. Compared with the traditional objective method, the performance advantage of EMFCNN in site prediction tasks is also significant, further verifying its potential in practical applications.
[0023] As a preferred embodiment of the present application, specifically, the Adam optimizer is used, and the Inverse Time Decay learning rate decay strategy is used to train the EMFCNN temperature model, and the specific settings are as follows, The initial learning rate is 0.01, the learning rate decay is performed once every epoch, the decay exponent is set to 0.1, the early stopping rule based on the validation set performance is used, that is, if the validation loss does not improve for 10 consecutive epochs, the training process is stopped, and the MAE is used as the validation loss function.
[0024] In this preferred embodiment, the specific super parameter settings are shown in Table 1: Table 1
[0025] As a preferred embodiment of the present application, specifically, The grid resolution of the obtained CLDAS data is 0.0625°x0.0625°, the time resolution is hourly, and the variables include 2-meter temperature, The grid resolution of the acquired ECMWF THIN data is 0.125°*0.125°, the time resolution is hourly, and the variables include 2-meter air temperature, 10-meter east-west wind, 10-meter north-south wind, air pressure, 925-hPa relative humidity, 925-hPa specific humidity, 925-hPa north-south wind, 925-hPa east-west wind, 925-hPa temperature, and 850-hPa relative humidity.
[0026] Specifically, high-impact elements are selected from the cleaned CLDAS and ECMWF THIN data according to local changes in air temperature, including, a simplified equation of local change in air temperature is constructed, , wherein, is the local change in temperature, i.e., the change in temperature per unit time, is the horizontal temperature advection term, i.e., the change in temperature caused by horizontal heat transport of the wind field, is the adiabatic change caused by vertical motion, is the non-adiabatic heating term, i.e., the influence of latent heat release; As can be seen, the local change in air temperature is related to the temperature change caused by the advection of the atmosphere, the adiabatic change caused by vertical motion, and non-adiabatic heating. Therefore, first, the ground, 925-hPa, and 850-hPa air pressure, air temperature, U-wind (east-west wind), V-wind (north-south wind), total cloud cover, low cloud cover, precipitation, relative humidity, and specific humidity are selected as preliminary screening factors; then, the correlation coefficients of these factors with the maximum and minimum temperatures are calculated, and the top 10 factors with the highest correlation coefficients that pass the 90% reliability test are selected as high-impact elements; these elements are 2-meter air temperature, 10-meter east-west wind, 10-meter north-south wind, ground air pressure, 925-hPa relative humidity, 925-hPa specific humidity, 925-hPa north-south wind, 925-hPa east-west wind, 925-hPa temperature, and 850-hPa relative humidity.
[0027] As a preferred embodiment of the present application, the CLDAS and ECMWF THIN data are cleaned by a systematic detection and correction mechanism to process abnormal values, error records, and missing fields therein.
[0028] As a preferred embodiment of the present application, specifically, The CLDAS data include maximum temperature, minimum temperature, and hourly temperature observation data at the whole point; The ECMWF THIN high-impact element data include key meteorological elements on the forecast day, and the key meteorological elements on the forecast day include 2-meter air temperature, 10-meter / 925-hPa wind field components, sea level air pressure, 925-hPa temperature and humidity parameters, and 850-hPa relative humidity.
[0029] As a preferred embodiment of the present application, specifically, the training set, the validation set and the test set for training are obtained based on time series division of the data set, including, The data from April to September in 2019-2022 in the data set is set as the training set, the data in the same period in 2023 is set as the validation set, and the data in the same period in 2024 is set as the test set, thereby forming a complete meteorological modeling data architecture.
[0030] The present application also provides a temperature correction device based on a full connection neural network, comprising the following: A model construction module is configured to construct an EMFCNN temperature model based on a full connection neural network; A model training module is configured to train the EMFCNN temperature model, and obtain an EMFCNN_24h highest temperature model, an EMFCNN_24h lowest temperature model and an EMFCNN_1h hourly temperature model after the training is completed; A temperature correction module is configured to perform temperature correction based on the EMFCNN_24h highest temperature model, the EMFCNN_24h lowest temperature model and the EMFCNN_1h hourly temperature model; Specifically, the training set, the validation set and the test set for training the EMFCNN temperature model are formed as follows, The CLDAS and ECMWFTHIN data of the target position are obtained, and the CLDAS and ECMWFTHIN data are cleaned to obtain cleaned CLDAS and ECMWFTHIN data; According to the local change of air temperature, the high-impact elements are selected from the cleaned CLDAS and ECMWFTHIN data to obtain CLDAS high-impact element data and ECMWFTHIN high-impact element data; The ECMWFTHIN high-impact element data is unified to the CLDAS grid system based on the Cressman interpolation algorithm; The CLDAS high-impact element data and the ECMWFTHIN high-impact element data of the CLDAS grid system are taken as a data set, and the training set, the validation set and the test set for training are obtained based on time series division of the data set; Specifically, the station number is encoded by Embedding, and then spliced with the mode prediction elements, and then input into the full connection neural network to output the 24h highest temperature, the 24h lowest temperature and the 1h hourly temperature prediction result, Wherein the station number includes the CLDAS grid points and the G1099 station.
[0031] In addition, a specific case of the temperature correction method and device based on the full connection neural network is as follows, Referring to (a), (b), (c) in Figure 2, on September 21, 2023, Guangdong Province experienced a significant large-scale high-temperature weather process. The daily maximum temperature in most parts of the province generally rose by 1-3℃ compared with the previous day (20th, see (a) of Figure 2) (Figure 2 (c)), among which the Pearl River Delta and northern Guangdong were particularly obvious, and the high-temperature area above 35℃ significantly expanded compared with the previous day (Figure 2 (b)).
[0032] In the forecast of this high-temperature process, each traditional numerical model showed obvious differences.
[0033] In contrast, the EMFCNN_24h maximum temperature model showed significant advantages in the forecast of this high-temperature process. The 24-hour maximum temperature forecast MAE of the model from 20:00 on the 20th to 20:00 on the 21st was only 0.62℃, which was reduced by 77% compared with the EC model. In terms of high-temperature threshold and spatial distribution characteristics, the forecast results of the EMFCNN_24h model were highly consistent with the actual situation, showing more accurate and detailed forecasting ability than traditional models. Especially in the spatial distribution and intensity prediction of the high-temperature area above 35℃, the model successfully overcame the systematic bias problem commonly existing in traditional models.
[0034] In summary, through the temperature correction method and device based on a full connection neural network proposed in the present application, a multi-element fusion temperature prediction model based on a full connection neural network, EMFCNN temperature model, is constructed for the 24-hour maximum temperature, minimum temperature and 1-hour temperature prediction demand of Guangdong Province. The prediction results are significantly better than those of traditional models in both grid prediction and station prediction scenarios, providing an expandable technical framework for fine temperature prediction in Guangdong.
[0035] Although the description of the present application has been quite detailed and particularly described with respect to several described embodiments, it is not intended to be limited to any of these details or embodiments or any special embodiment, but should be considered to provide a broad interpretation of the prior art to the appended claims by considering the prior art, thereby effectively encompassing the intended scope of the present application. In addition, the present application is described above in embodiments that the inventors can foresee, and the purpose is to provide a useful description, and non-essential modifications to the present application that have not yet been foreseen can still represent equivalent modifications to the present application.
[0036] The above is only a preferred embodiment of the present application, and the present application is not limited to the above embodiments, as long as the same means achieve the technical effect of the present application, it should belong to the protection scope of the present application. The technical solutions and / or embodiments within the protection scope of the present application can have various modifications and changes.
Claims
1. A temperature correction method based on a fully connected neural network, characterized in that, The method comprises the following steps: constructing an EMFCNN temperature model based on a full connection neural network; training the EMFCNN temperature model, and obtaining an EMFCNN_24h highest temperature model, an EMFCNN_24h lowest temperature model and an EMFCNN_hourly air temperature model after the training is completed; performing temperature correction based on the EMFCNN_24h highest temperature model, the EMFCNN_24h lowest temperature model and the EMFCNN_hourly air temperature model; Specifically, the formation process of a training set, a validation set and a test set for training the EMFCNN temperature model is as follows, obtaining CLDAS and ECMWFTHIN data of a target position, and performing data cleaning on the CLDAS and ECMWFTHIN data to obtain cleaned CLDAS and ECMWFTHIN data; selecting high-influence elements from the cleaned CLDAS and ECMWFTHIN data according to local changes in air temperature to obtain CLDAS high-influence element data and ECMWFTHIN high-influence element data; unifying the ECMWFTHIN high-influence element data to a CLDAS grid system based on a Cressman interpolation algorithm; taking the CLDAS high-influence element data and the ECMWFTHIN high-influence element data of the CLDAS grid system as a data set, and dividing the data set based on a time sequence to obtain the training set, the validation set and the test set for training; Specifically, the station number is subjected to Embedding coding, and then is spliced with the model prediction elements before being input into the full connection neural network to output 24h highest temperature, 24h lowest temperature and 1h hourly air temperature prediction results. The station number comprises CLDAS grid points and G1099 stations.
2. The temperature correction method based on a fully connected neural network according to claim 1, characterized in that, Specifically, the EMFCNN temperature model is trained by using an Adam optimizer and combining an Inverse Time Decay learning rate decay strategy, and the specific settings are as follows, the initial learning rate is 0.01, the learning rate is decayed once every epoch, the decay index is set to 0.1, the early stopping rule based on the validation set performance is used, that is, if the validation loss does not improve for 10 consecutive epochs, the training process is stopped, and the MAE is used as the validation loss function.
3. The temperature correction method based on a fully connected neural network according to claim 1, characterized in that, Specifically, The grid resolution of the obtained CLDAS data is 0.0625°×0.0625°, the time resolution is hourly, and the variables include 2m temperature. The grid resolution of the obtained ECMWFTHIN data is 0.125°×0.125°, the time resolution is hourly, and the variables include 2m air temperature, 10m east-west wind, 10m north-south wind, air pressure, 925hPa relative humidity, 925hPa specific humidity, 925hPa north-south wind, 925hPa east-west wind, 925hPa temperature and 850hPa relative humidity.
4. The temperature correction method based on a fully connected neural network according to claim 1, characterized in that, Specifically, the CLDAS and ECMWFTHIN data are subjected to data cleaning through a systematic detection and correction mechanism to process abnormal values, error records and missing fields.
5. The temperature correction method based on a fully connected neural network according to claim 3, characterized in that, Specifically, The CLDAS data includes maximum temperature, minimum temperature, and hourly temperature observation data at integral points; The ECMWFTHIN high-impact factor data includes key meteorological factors on the forecast day, including 2-meter air temperature, 10-meter / 925-hPa wind field components, sea level pressure, 925-hPa temperature and humidity parameters, and 850-hPa relative humidity; wherein the high-impact factor for the maximum temperature is the key meteorological factor at 14:00 on the forecast day, the high-impact factor for the minimum temperature is the key meteorological factor at 04:00 on the forecast day, and the high-impact factor for the hourly temperature at integral points is the key meteorological factor at each time period throughout the day on the forecast day.
6. The temperature correction method based on a fully connected neural network according to claim 1, characterized in that, Specifically, the training set, the validation set, and the test set obtained based on the time series division of the data set include, The data from April to September in 2019-2022 in the data set is set as the training set, the data in the same period in 2023 is set as the validation set, and the data in the same period in 2024 is set as the test set, forming a complete meteorological modeling data architecture.
7. The temperature correction method based on a fully connected neural network according to claim 1, characterized in that, Specifically, the high-impact factors are selected from the cleaned CLDAS and ECMWFTHIN data according to the local variation of air temperature, including, A simplified equation for the local variation of air temperature is constructed, where is the local change of temperature, i.e. temperature change per time, is the horizontal temperature advection, i.e. temperature change due to horizontal transport of heat by the wind field, ω(Γ-Γ d is the diabatic change due to vertical motion, is the non-adiabatic heating term, i.e. influence of latent heat release; As can be seen, the local variation of air temperature is related to the temperature change caused by the advection of the atmosphere, the adiabatic change caused by the vertical motion, and the non-adiabatic heating. Therefore, first, the ground pressure, air temperature, U wind (east-west wind), V wind (north-south wind), total cloud cover, low cloud cover, precipitation, relative humidity, and specific humidity at 925 hPa and 850 hPa are selected as the initial screening factors; then, the correlation coefficients of these factors with the maximum temperature and the minimum temperature are calculated, and the top 10 factors that pass the 90% reliability test and have the highest correlation coefficients are selected as the high-impact factors; these factors are 2-meter air temperature, 10-meter east-west wind, 10-meter north-south wind, ground pressure, 925-hPa relative humidity, 925-hPa specific humidity, 925-hPa north-south wind, 925-hPa east-west wind, 925-hPa temperature, and 850-hPa relative humidity.
8. A temperature correction device based on a fully connected neural network, characterized by It includes the following: A model construction module for constructing an EMFCNN temperature model based on a fully connected neural network; A model training module for training the EMFCNN temperature model, and obtaining an EMFCNN_24h maximum temperature model, an EMFCNN_24h minimum temperature model, and an EMFCNN_hourly temperature model after the training is completed; A temperature correction module for performing temperature correction based on the EMFCNN_24h maximum temperature model, the EMFCNN_24h minimum temperature model, and the EMFCNN_hourly temperature model; Specifically, the formation process of the training set, the validation set, and the test set for training the EMFCNN temperature model is as follows, Get the CLDAS and ECMWFTHIN data of the target location, and clean the CLDAS and ECMWFTHIN data to obtain cleaned CLDAS and ECMWFTHIN data; According to the local variation of air temperature, high-impact elements are selected from the CLDAS and ECMWFTHIN data after cleaning, to obtain CLDAS high-impact element data and ECMWFTHIN high-impact element data; Based on the Cressman interpolation algorithm, the ECMWFTHIN high-impact element data is unified to the CLDAS grid system; The CLDAS high-impact element data and the ECMWFTHIN high-impact element data of the CLDAS grid system are taken as a data set, and the data set is divided based on the time sequence to obtain a training set, a validation set and a test set for training; Specifically, the station number is subjected to Embedding coding, then splicing operation is performed with the mode prediction elements, and then the station number and the mode prediction elements are jointly input into a full-connection neural network to output 24h maximum temperature, 24h minimum temperature and 1h temperature prediction results, Wherein the station number includes CLDAS grid points and G1099 stations.