Temperature forecast correction method based on dynamic weight integration

Abstract

The application provides a temperature prediction revision method based on dynamic weight integration, comprising the following steps: 1, obtaining ground live temperature data and mode prediction temperature data; 2, data preprocessing; 3, neural network revision, adopting multiple neural networks to revise the 2m temperature prediction product of CMA-GD; further, adopting BP neural network, BP neural network optimized by genetic algorithm, generalized regression neural network, wavelet neural network, long short-term memory network and the like to revise the 2m temperature prediction product of CMA-GD mode; 4, dynamic weight integration; and 5, obtaining the final revised temperature. The application is based on the 2m temperature prediction product directly output by CMA-GD mode, the mode temperature prediction is corrected by deviation through multiple neural networks, the revised result is integrated by advantage through a dynamic weight scheme, a set of localized objective temperature prediction revision product is formed, a powerful reference guidance is provided for temperature prediction, and the local temperature prediction quality is further improved.

Description

Technical Field

[0001] This invention belongs to the field of meteorological forecasting technology, and in particular relates to a temperature forecast correction method based on dynamic weight integration. Background Technology

[0002] With economic development and social progress, people's demand for weather forecasts is expanding, and their requirements for forecast accuracy and precision are also increasing. Numerical weather prediction, as an important means to improve forecast accuracy, plays a crucial role in weather forecasting operations. In recent years, numerical weather prediction technology has developed rapidly, and forecasting levels have continuously improved. However, due to uncertainties in the initial field, model errors, and atmospheric chaos, numerical model forecasts still have certain biases. How to effectively reduce model forecast errors and continuously improve forecast accuracy and precision has always been a key focus and challenge for meteorological operational departments, as well as a persistent focal point and hot topic in the field of meteorological scientific research.

[0003] To reduce the discrepancy between the direct output forecasts and actual conditions from numerical models, post-processing methods are often used to adjust the model results. Currently, the main post-processing methods include Model Output Statistics (MOS), Kalman filtering, similarity forecasting, and neural networks. Among these, neural networks, as data processing systems that can automatically realize the nonlinear mapping relationship between two sets of variables, have strong adaptive learning and collective computing capabilities and are widely used in the meteorological field. Xiong Shiwei et al. used Backpropagation (BP) neural networks to correct the fine-grid model output products provided by the European Centre for Medium Range Weather Forecasts (ECMWF), and the results showed that the nonlinear mapping relationship of neural networks gives them a significant advantage in fault tolerance. Han et al. found that Long Short Term Memory (LSTM) recurrent neural networks have good forecasting performance for hourly temperature. However, neural network methods also have certain shortcomings. Previous studies have found that due to the lack of prior knowledge of the problem, neural networks are prone to getting stuck in local extrema and overfitting during actual training. Therefore, although different neural networks can provide some effective information, the results may vary significantly. To overcome the instability of single forecast results, some scholars have introduced the concept of ensemble forecasting, which involves integrating multiple results to obtain deterministic forecasts. Numerous studies have shown that ensemble forecasts improve upon the instability of individual forecast techniques to some extent. For example, Wei Fengying designed a dynamic weighted ensemble forecasting scheme using historical forecast techniques from multiple forecasting methods as weighting coefficients. Independent sample testing over nine years of nationwide precipitation data demonstrated that the ensemble forecasting model eliminated the randomness of the original methods to a certain extent. Zhao Shengrong applied neural network methods to establish an integrated temperature forecasting system for forecast products from different operational models in three countries. The results showed that the integrated temperature forecast was significantly better than that of individual models.

[0004] Complex terrain conditions lead to significant local temperature variations, resulting in poor temperature forecasts from various numerical models and relatively low forecast quality. Therefore, research on localized correction of temperature forecasts based on direct model outputs urgently needs to be strengthened. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a temperature forecast correction method based on dynamic weight integration.

[0006] The Rapid Update Assimilation Numerical Weather Prediction System (CMA-GD) of the Guangzhou Tropical Marine Meteorological Institute, China Meteorological Administration, is a regional numerical weather prediction system independently developed in my country. It enables rapid assimilation of satellite, radar, and ground observation data within a region. This model can output hourly temperature forecast products, which to some extent meets the public's high requirements for temporal resolution in temperature forecasts and has been of significant reference value in daily operational forecasting for many years. This invention, based on the 2-meter temperature forecast product directly output by the CMA-GD model, corrects the bias in the model's temperature forecast through multiple neural networks and integrates the correction results using a dynamic weighting scheme to form a localized, objective temperature forecast correction product. This provides strong reference guidance for temperature forecasting and further improves the quality of local temperature forecasts.

[0007] The present invention adopts the following technical solution:

[0008] Temperature forecast correction methods based on dynamic weighted integration include:

[0009] Step 1. Obtain actual surface temperature data and model-predicted temperature data.

[0010] Step 2. Data Preprocessing

[0011] Before temperature correction, the data needs to be preprocessed.

[0012] Step 3. Neural Network Correction

[0013] Multiple neural networks were employed to correct the 2-meter temperature forecast products of the CMA-GD model. Furthermore, methods such as BP neural networks, BP neural networks optimized by genetic algorithms, generalized regression neural networks, wavelet neural networks, and long short-term memory networks were used to correct the 2-meter temperature forecast products of the CMA-GD model.

[0014] Step 4. Dynamic Weight Integration

[0015] Integrated forecasting is a method that uses mathematical models to synthesize forecast results from multiple forecasting tools or members for the same forecast object, thereby obtaining a unified forecast value. These forecast results can be forecasts from different models or different corrections of the same model. Furthermore, the formula is as follows, assuming there are n forecast results, denoted as F1, F2, ..., F... n The weighting coefficients for each forecast result are C1, C2, ..., C6. n Then the integrated forecast value E can be expressed as:

[0016]

[0017] In the formula, the weighting coefficients of each forecast result satisfy... Dynamic weighted ensemble forecasting is an ensemble forecasting method that dynamically adjusts weights as the forecast time progresses.

[0018] Step 5. Obtain the final corrected temperature

[0019] Furthermore, step 2 includes marking outliers as values ​​in the daily ground observation and model forecast data that deviate from three times the absolute deviation of the median. When outliers or missing values ​​appear for 3 consecutive hours, the samples for that day are removed. If the number of missing samples is less than 3 hours, linear interpolation is used to fill in the missing samples.

[0020] Furthermore, the duration of the training samples for the neural network in step 3 has a crucial impact on the training results. Therefore, it is necessary to determine the training sample parameters for different time periods. Considering the continuity of temperature and its stable trend in the short term, the hourly forecast temperature (input) and corresponding actual temperature (output) of the past N1 days' numerical model are selected as the neural network training samples. The trained neural network is then used to correct the forecast data for the current day. The training duration parameter N1 is consistent for the BP, BP_GA, GRNN, and WAVENN neural networks. The LSTM neural network differs significantly from the previous neural networks, so the training samples of the past N days are set to N2 to distinguish them. The input field of the LSTM neural network also involves hourly actual data from the past period, which is set to N3 hours.

[0021] Furthermore, for the selection of N1, N2, and N3 days in the training samples, an enumeration method was used to adjust the parameters of each neural network. Considering the prediction accuracy, mean absolute error, and computational efficiency, the values ​​of N1, N2, and N3 were set to 30 days, 10 days, and 24 hours, respectively.

[0022] Furthermore, step 3 takes into account the continuity of temperature. The selected samples should be able to accurately reflect the inherent pattern of the model-predicted temperature and the actual temperature within the current time period. If the duration of the training samples is too long and far from the forecast results of the current time period, it will have a certain interference with the training results of the current time period.

[0023] Furthermore, after eliminating pattern systematic errors through multiple neural networks in step 3, a dynamic weighting scheme is used to integrate the correction results. The hourly integrated temperature can be expressed as:

[0024]

[0025] In the formula, i represents the order of the forecast results (values ​​1, 2, ..., 6, representing model forecast, BP, BP_GA, GRNN, WAVENN, LSTM, respectively), j represents the forecast lead time (values ​​0, 1, ..., 72), and E j The weighting coefficients for each forecast result at the j-th hour forecast lead time are determined by comparing the forecast temperatures from different forecast lead times models and the historical verification and evaluation results of the neural network correction results with the corresponding actual temperatures. The calculation formula is as follows:

[0026]

[0027] In the formula, MAE ij The MAE (Modular Value Equation) is the sum of the forecasts for the j-th hour of the historical test and evaluation sample and the corresponding actual temperatures.

[0028] Furthermore, step 4 also includes verification, specifically verification using the mean absolute error.

[0029] The mean absolute error (MAE) is the average of the absolute deviations of all individual observations from the arithmetic mean. It avoids the problem of error cancellation and thus accurately reflects the magnitude of the actual prediction error. Its functional expression is:

[0030]

[0031] In the formula, F i For the predicted temperature of station i (time), O i The actual temperature at station i (time) is given.

[0032] Furthermore, step 4 also includes comparison, specifically judging by forecast accuracy. The standard for a correct forecast is an absolute forecast error of ≤2℃. Therefore, the temperature forecast accuracy is expressed as the percentage of forecasts with an absolute error ≤2℃, and the function expression is:

[0033]

[0034] In the formula, K = 2, representing |F i -O i |≤2℃, Nr K To forecast the correct number of stations (times), Nf K This represents the total number of stations (times) forecasted.

[0035] The beneficial effects of this invention are:

[0036] This invention addresses the challenges of temperature forecasting in complex terrain by integrating multiple methods within a single model for correction. The results are as follows: First, based on historical data verification and evaluation, various neural network techniques are used to correct biases in the temperature forecasts directly output by numerical models, effectively reducing systematic model errors. Second, considering the similarity of correction results from different neural networks, a dynamic weighting scheme is employed to integrate the advantages of model forecasts and corrections, significantly improving forecast quality and stability.

[0037] Taking the hourly correction results of the 72-hour forecast lead time in Guizhou Province in 2020 as an example, the average absolute error after correction was reduced by 14.93% (from 2.01℃ to 1.71℃) and the forecast accuracy was improved by 8.24% (from 60.46% to 68.70%). Attached Figure Description

[0038] Figure 1 A map showing the distribution of 365 national assessment stations in Guizhou Province;

[0039] Figure 2 A graph showing the effect of different training durations on the temperature correction of neural networks;

[0040] Figure 3 A graph showing the effect of different enumeration durations on the temperature correction of dynamic weight integration;

[0041] Figure 4 This is a flowchart of the present invention;

[0042] Figure 5 This is a graph showing the hourly temperature evolution in Guiyang within the 72-hour forecast period starting at 08:00 on July 1, 2020.

[0043] Figure 6(a) shows the deviation between the temperature forecast by the 72-hour forecast lead time model for Guiyang in July 2020, starting from 08:00 daily.

[0044] Figure 6(b) shows the deviation between the daily 72-hour forecast correction temperature and the actual temperature in Guiyang starting from 08:00 in July 2020;

[0045] Figure 7(a) shows the forecast temperature in Guiyang from 08:00 in 2020 using the forecast lead time model.

[0046] Figure 7(b) is a scatter plot of the dynamic weighted integration results and the corresponding actual temperature in Guiyang starting from 08:00 in 2020.

[0047] Figure 8 The plot shows the hourly average absolute error of model-predicted temperatures, results of different correction schemes, and actual conditions within the 72-hour forecast period for Guizhou in 2020.

[0048] Figure 9A daily average absolute error plot of the temperature forecast and dynamic weighted integrated correction results of the 72-hour forecast lead time model for Guizhou in 2020;

[0049] Figures 10(a) and 10(b) show the average absolute error of the 72-hour forecast lead time in Guizhou in 2020.

[0050] Figures 11(a) and 11(b) show the accuracy of the 72-hour forecast lead time in Guizhou in 2020.

[0051] Figures 12(a) and 12(b) show the monthly verification results of the 72-hour forecast lead time in Guizhou in 2020. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention are described clearly and completely below. Obviously, the described embodiments are only some embodiments of this invention, not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0053] Example

[0054] Source of information

[0055] The data used include CMA-GD forecast products from 2019-2020 and routine surface meteorological observation data from 365 national assessment stations in Guizhou Province during the same period. Detailed descriptions of the data are as follows:

[0056] (1) The CMA-GD forecast product is the surface 2m temperature element within the forecast period of the next 84 hours, reported from 08:00 (Beijing time, the same below) and 20:00 every day, with a time interval of 1 hour and a spatial resolution of 0.03°×0.03°.

[0057] (2) Ground temperature at 2m depth was measured at 365 national monitoring stations within Guizhou Province at 1-hour intervals. The distribution of the stations is as follows: Figure 1 As shown.

[0058] The matching principle between grid data and station data is as follows: the numerical weather prediction products are interpolated to 365 national assessment stations in Guizhou Province using the nearest neighbor interpolation method to obtain the ground 2m temperature forecast values ​​for each station.

[0059] Correction method

[0060] (1) Neural network correction

[0061] This invention employs methods such as BP neural network, BP neural network optimized by genetic algorithms (GA), generalized regression neural network (GRNN), wavelet neural network (WAVENN), and long short term memory network (LSTM) to correct the 2m temperature forecast products of the CMA-GD model.

[0062] (2) Dynamic weighted integrated forecast

[0063] Integrated forecasting is a method that uses mathematical models to synthesize forecast results from multiple forecasting tools or members for the same forecast object, thereby obtaining a unified forecast value. These forecast results can be forecasts from different models or different corrections from the same model. The specific formula is as follows, assuming there are n forecast results, denoted as F1, F2, ..., F... n The weighting coefficients for each forecast result are C1, C2, ..., C6. n Then the integrated forecast value E can be expressed as:

[0064]

[0065] In the formula, the weighting coefficients of each forecast result satisfy... Dynamic weighted ensemble forecasting is an ensemble forecasting method that dynamically adjusts weights as the forecast time progresses.

[0066] Inspection parameters

[0067] The inspection parameters and correction methods involved in this invention are as follows:

[0068] (1) Mean Absolute Error (MAE)

[0069] The mean absolute error (MAO) is the average of the absolute deviations of all individual observations from the arithmetic mean. It avoids the problem of error cancellation and thus accurately reflects the magnitude of the actual prediction error. The mean absolute error of temperature, T... MAE The function expression is:

[0070]

[0071] In the formula, N is the total number of stations (times), F i For the predicted temperature of station i (time), O i The actual temperature at station i (time) is given.

[0072] (2) Forecast Accuracy (FA)

[0073] This invention uses an absolute forecast error of ≤2℃ as the standard for accurate forecasting. Therefore, the temperature forecast accuracy rate FA is expressed as the percentage of temperature forecasts with an absolute error of ≤2℃, and its functional expression is:

[0074]

[0075] In the formula, K = 2, representing |F i -O i |≤2℃, Nr K To forecast the correct number of stations (times), Nf K This represents the total number of stations (times) forecasted.

[0076] Correction process

[0077] Since the model itself takes up to 8 hours to run and push, this invention, in accordance with actual operational needs, corrects the model's forecast products by shifting the forecast time by 12 hours. Specifically, it only corrects the 12-84 hour forecast lead time products and uses them as the corrected 0-72 hour forecast lead time products for the next reporting time. For example, if a forecast lead time product for the next 84 hours is reported starting at 08:00 on a certain day, and the local receiving time is approximately 16:00 on the same day, then the 12-84 hour forecast lead time correction product reported starting at 08:00 on that day will be used as the 0-72 hour forecast value reported starting at 20:00 on that day. Furthermore, considering the truncation issue of real-time data, and referring to the objective product transmission time requirements of the China Meteorological Administration's Smart Forecasting Technology and Methodology Exchange Competition (08:00 forecasts must be before 04:30, and 20:00 forecasts must be before 16:30), this invention removes real-time samples that are 4 hours ahead of the current reporting time to ensure smooth subsequent operational implementation.

[0078] Data preprocessing

[0079] Before temperature correction, data preprocessing is required. This invention marks outliers as values ​​in daily ground-based and model-forecasted data that deviate from three times the median absolute deviation (MAD). When outliers or missing values ​​appear for three consecutive hours, the samples for that day are removed. If the number of missing samples is less than three hours, linear interpolation is used to complete the missing samples.

[0080] Neural network correction

[0081] This invention employs multiple neural networks to correct the 2-meter temperature forecast of CMA-GD products. Specific correction schemes for each neural network are shown in Table 1. The duration of the training sample time for the neural network has a crucial impact on the training results. Therefore, it is necessary to determine the training sample parameters for different time periods. Considering the continuity of temperature and its stable trend in short periods, the hourly forecast temperatures (input) and corresponding actual temperatures (output) from the past N1 days of numerical models are selected as the neural network training samples. The trained neural network is then used to correct the forecast data for the current day. The training duration parameter N1 is consistent for the BP, BP_GA, GRNN, and WAVENN neural networks. The LSTM neural network differs significantly from the previous neural networks; therefore, the training samples from the past N days are set to N2 to distinguish them. The input field of the LSTM neural network also involves hourly actual data from a past period, which is set to N3 hours in this invention.

[0082] Table 1 Overview of Neural Networks

[0083]

[0084]

[0085] For the selection of the past N1 days of training samples, an enumeration method was used to adjust the parameters of each neural network. Taking the 72-hour forecast lead time data of representative stations in 9 cities and prefectures of Guizhou Province in 2019 as an example, different correction effects of the BP neural network were obtained by changing the value of N1, such as... Figure 2 As shown in the figure, the forecast accuracy initially decreases and then increases as N1 increases. After N1 reaches 25 days, the accuracy gradually surpasses the model itself, while the mean absolute error shows the opposite trend. The results from BP_GA, GRNN, and WAVENN neural networks are basically consistent. Considering the continuous nature of temperature, the selected samples must accurately reflect the inherent patterns between the model's forecast temperature and the actual temperature within the current period. If the training sample duration is too long and far removed from the forecast results for the current period, it may interfere with the training results for the current period. Therefore, considering the comprehensive factors of forecast accuracy, mean absolute error, and computational efficiency, the value of N1 is set to 30. Similarly, the number of days N2 for LSTM training samples and the number of past actual hours N3 are set to 10 and 24, respectively.

[0086] Dynamic weight integration

[0087] After eliminating pattern systematic errors using various neural network techniques, a dynamic weighting scheme is employed to integrate the correction results. The hourly integrated temperature can be expressed as:

[0088]

[0089] In the formula, i represents the order of the forecast results (values ​​1, 2, ..., 6, representing model forecast, BP, BP_GA, GRNN, WAVENN, LSTM, respectively), j represents the forecast lead time (values ​​0, 1, ..., 72), and E j F represents the weighting coefficient of each forecast result for the j-th hour forecast lead time. ij For the model forecast temperature and the correction results of each neural network for the j-th hour forecast lead time, C ij The weighting coefficients for each forecast result at the j-th hour forecast lead time are determined by comparing the forecast temperatures from different forecast lead times models and the historical verification and evaluation results of the neural network correction results with the corresponding actual temperatures. The calculation formula is as follows:

[0090]

[0091] In the formula, MAE ij To evaluate the MAE of each forecast result with the corresponding actual temperature at the j-th hour of the historical test sample, this invention sets the historical test sample period as the reciprocal of the average absolute error between the model forecast temperature and the corresponding actual temperature over the past N4 days, and dynamically updates the weight coefficients of the integrated forecast.

[0092] Regarding the selection of the dynamic weighted integration parameter N4, taking the 72-hour forecast lead time data of representative stations in 9 cities of Guizhou Province in 2019 as an example, an enumeration method is used for debugging. This invention corrects the 72-hour forecast lead time, so N4 should be greater than or equal to 3 days. However, due to transmission timeliness requirements, after removing the actual situation samples 4 hours before the reporting start time, when N4 is 3, only the 0-68 hour forecast lead time has corresponding actual situation data; the 69-72 hour forecast lead time has no corresponding actual situation data, making it impossible to obtain its average absolute error. Therefore, N4 is enumerated starting from 4 days, such as... Figure 3 As shown in the figure, under different enumeration durations, the dynamic weighted integrated correction temperature shows better performance than the model forecast in terms of N4. In particular, the mean absolute error reduction and forecast accuracy improvement are more significant in the 7-day and longer period, and the performance is more stable. Therefore, the N4 value is set to 7 within the allowable computing power conditions.

[0093] Correction Case

[0094] In summary, this invention first preprocesses ground-based and model-predicted data, then uses multiple neural networks to correct hourly temperatures within the 72-hour forecast lead time of the CMA-GD model. Finally, a dynamic weighting scheme is employed to integrate the model-predicted temperatures and the correction results from each neural network, and the correction results are individually verified and compared. Figure 4 As shown.

[0095] This invention takes Guiyang Station (57816, 106.73°E, 26.58°N), the capital of Guizhou Province, as an example, and presents the hourly actual temperature, model forecast, results of each neural network and dynamic weight integration, and the deviation between the dynamic weight integration results and the actual temperature within the 72-hour forecast period from 08:00 on July 1, 2020. Figure 5 As shown, the differences in correction results across different forecast lead times are significant. Specifically, after correction using BP, BP_GA, WAVENN, and GRNN neural networks, the model's low-temperature predictions were effectively improved, but the high-temperature predictions deviated from the actual data. In contrast, the LSTM neural network, by incorporating past data as input, not only changed the magnitude of the data but also altered the trend, resulting in a higher-temperature forecast that more closely matches the actual data. Due to the strength of LSTM in predicting continuous temperature trends, its correction effect is more prominent compared to other methods. In terms of ensemble performance, the dynamic weighted ensemble results for 0-54h forecast lead times show smaller deviations from the actual data, with high and low temperature forecasts closely matching actual observations, demonstrating greater stability than single models. However, after the 54-hour forecast lead time, the sharp drop in actual data leads to significant deviations in both the model itself and the correction results.

[0096] Results and Analysis

[0097] Single-station inspection

[0098] Table 2 presents the verification results of the 72-hour forecast lead time for Guiyang single-station forecasts starting from 08:00 on July 1, 2020. It can be seen that the dynamic weighted ensemble scheme demonstrates a certain positive technique in terms of mean absolute error and forecast accuracy compared to the model itself and the correction results of a single neural network. Although the mean absolute error of the dynamic weighted ensemble scheme is comparable to that of the LSTM neural network correction results (both are 1.18°C), the ensemble of multiple neural networks significantly improves the generalization ability of the learning system, resulting in a further improvement in forecast accuracy (from 79.45% to 83.56%).

[0099] Table 2. Results of the 72-hour Forecast Lead Time Verification for Guiyang at 08:00 on July 1, 2020

[0100]

[0101] Figure 6(a) compares the hourly deviations of the model-predicted temperature, the dynamic weighted integration result, and the actual temperature within the 72-hour forecast period starting at 08:00 daily in Guiyang in July 2020. The CMA-GD temperature forecast exhibits some abnormally high lead times, particularly in the 0-12h, 24-36h, and 48-60h periods, where the deviations exceed 4-5℃. This phenomenon was effectively corrected after dynamic weighted integration, with most deviations reduced to within ±1.5℃. Furthermore, the slight underestimation of temperature in other forecast lead times by CMA-GD was also improved after correction, as shown in Figure 6(b).

[0102] Further testing of the correction results over a longer period revealed that the dynamic weighted ensemble scheme effectively improves forecast reliability and exhibits a higher positive skill compared to single results. As shown in Figures 7(a) and 7(b), the average absolute error of the corrected temperature after dynamic weighted ensemble significantly decreases from 1.80℃ to 1.48℃ compared to the model's forecast, while the forecast accuracy also improves substantially from 65.96% to 74.13%. The model's forecast for low temperatures showed a significant overestimation compared to the actual conditions (Figure 7(a), with many scatter points at the bottom corner above the +2℃ line). After dynamic weighted ensemble, this overestimation was effectively corrected, with the scatter points more concentrated near the 45° diagonal (Figure 7(b)). In comparison, the correction results from 20:00 onwards were not as ideal as those from 08:00, but still demonstrated positive skill for low-temperature forecast corrections. This may be because the time interval between the 20:00 forecast and the occurrence of the high temperature (assumed to be 14:00) is longer compared to the 08:00 forecast. On the one hand, the model forecast quality gradually decreases as the forecast duration extends, and the model's forecast quality for high temperatures will decrease. On the other hand, the LSTM neural network involves input from past actual temperatures; the longer the forecast lead time, the greater the deviation from reality. Therefore, the correction effect of the 20:00 forecast is not as ideal as that of the 08:00 forecast. In addition, this invention also calculated the correction results for a 72-hour forecast lead time, which are qualitatively consistent with the conclusions of the 24-hour forecast. The forecast quality after dynamic weight integration is improved compared to the model's temperature forecast, with the forecast accuracy increasing from 63.54% to 69.75% and the MAE decreasing from 1.91℃ to 1.70℃.

[0103] Provincial inspection

[0104] Regarding temperature corrections for the entire Guizhou province, this invention provides the hourly average absolute error of model-predicted temperatures, different correction schemes, and actual temperatures over a 72-hour forecast period for 365 national assessment stations in Guizhou in 2020. For example... Figure 8As shown in the figure, the average absolute error of temperature forecasts and their corrections generally exhibits a periodic variation with the increase in forecast lead time, but the correction effects of different neural networks vary significantly across different forecast lead times. Generally, the periods when the daily maximum and minimum temperatures usually occur are afternoon and early morning, which are assumed here to correspond to 2 PM and 2 AM, respectively. Taking a forecast starting at 8 AM as an example, the average absolute error of even-numbered forecast lead times of 6 (0 / 12 / 24 / 36 / 48 / 60 / 72) is in a relatively small range, while the average absolute error of odd-numbered forecast lead times of 6 (6 / 18 / 30 / 42 / 54 / 66h) is in a relatively large range. This large range corresponds precisely to the daily maximum temperature (2 PM daily, the period when high temperatures often occur) and the minimum temperature (2 AM daily, the period when low temperatures often occur), which is precisely where the difficulty lies in forecasting extreme temperature values. Among them, the BP, BP_GA, WAVENN, and GRNN neural networks all showed significant positive skill in correcting low-temperature forecasts (daily 02:00, 18 / 42 / 66-hour forecast lead times), while the correction effect for high-temperature forecasts (daily 14:00, 6 / 30 / 54-hour forecast lead times) was poor, with some lead times showing negative skill. Although the LSTM neural network's correction effect for low temperatures was not as good as the aforementioned neural networks, it still showed positive skill in high-temperature corrections, and its overall performance was relatively stable (Table 3). From the overall forecast performance results for 0-72 hours, all neural networks can effectively reduce model systematic errors, with an average absolute error reduction of 0.02-0.17℃. The integrated temperature forecast after re-weighting showed better results than the model itself and the correction results of each neural network, with an average absolute error reduction of 0.30℃.

[0105] Table 3. Average absolute error of temperature forecasts from some forecast-leading models and actual conditions in Guizhou Province in 2020 (unit: °C)

[0106]

[0107] like Figure 9 As shown, although the mean absolute error of the model and the dynamic weighted integration fluctuates over time, the daily mean absolute error after correction is generally lower than the model forecast result, indicating that the dynamic weighted integration scheme has a significant correction effect on the model's temperature forecast and exhibits a certain degree of stability.

[0108] Furthermore, this invention examined the spatial distribution of model-forecasted temperature and dynamically weighted integrated temperature within the 72-hour forecast lead time of the 365 National Assessment Stations in Guizhou Province in 2020, such as... Figure 10(a) - Figure 1 1 (b). The results show that the dynamic weighted ensemble correction scheme can effectively solve the problem of large biases in model forecasts in most parts of Guizhou Province. The average absolute error of the 72-hour forecast lead time for the whole province decreased by 14.93% (from 2.01℃ to 1.71℃). Figures 10(a)-10(b)Meanwhile, the forecast accuracy improved by 8.24% (from 60.46% to 68.70%). Figures 11(a)-11(b) However, it is worth noting that there is some spatial unevenness in the temperature deviations of the model forecasts and the dynamic weighted integration, especially in the western part of the province, where large areas of mean absolute error still exist locally after dynamic weighted integration correction. This may be related to the topography of Guizhou, which is high in the west and low in the east. Zhu Wenda et al.'s batch experiments on the CMA-GD model using high-resolution topographic data showed that areas with higher topography often correspond to areas with large mean absolute errors in temperature. Therefore, further local corrections based on topography can be made in the future. In addition, from the perspective of model forecasts and correction effects at different forecast lead times, the correction effect is more obvious on the first day (0-24h), while the effects are basically the same on the following two days (Table 4).

[0109] Table 4. Hourly Temperature Correction Effects for the Next Three Days in Guizhou in 2020

[0110]

[0111] After monthly verification of the model forecasts and their corrections for the 72-hour forecast period across Guizhou Province in 2020, from... Figures 12(a)-12(b) It can be observed that the dynamic weighted integrated temperature correction exhibits significant correction techniques compared to the model itself in each month, reducing the mean absolute error (-0.49℃ to -0.12℃) and improving forecast accuracy (3.17% to 14.16%). Seasonally, the correction effect is more pronounced in summer compared to other seasons, with a mean absolute error reduction of 0.39℃ and a forecast accuracy improvement of 11.49%.

[0112] This invention utilizes CMA-GD model data from 2019-2020 and surface 2m temperature data from 365 national assessment stations in Guizhou Province during the same period. It eliminates model systematic errors through various neural network techniques and employs a dynamic weighting scheme to integrate the correction results, ultimately obtaining a localized temperature forecast correction product. The results show that:

[0113] (1) Based on the verification and evaluation of historical data, various neural network methods can effectively reduce model systematic errors. Neural network methods have a certain improvement effect on the model's forecast of higher than normal low temperatures in July at Guiyang station. At the same time, for the overall forecast effect of 0-72h in Guizhou Province 365 National Assessment Station, the average absolute error of temperature after correction by BP, BP_GA, WAVENN, GRNN, LSTM neural networks and other methods is reduced by 0.01-0.17℃ compared with the model.

[0114] (2) Since the correction effects of different neural networks vary at different forecast lead times, the dynamic weighting scheme can further improve the model correction effect by integrating the correction results. Whether for Guiyang station alone or for the entire Guizhou province, the mean absolute error of temperature forecasts after dynamic weighting integration is generally lower than that of the direct output of the model and the correction results of each neural network at different forecast lead times, and the forecast accuracy is also improved. Taking the hourly correction results of the 72-hour forecast lead time for Guizhou province in 2020 as an example, the mean absolute error after correction is reduced by 14.93% (from 2.01℃ to 1.71℃) compared with the model, and the forecast accuracy is improved by 8.24% (from 60.46% to 68.70%). In addition, the dynamic weighting integration correction results show a certain degree of stability in monthly and daily tests.

[0115] This invention addresses the challenges of temperature forecasting in Guizhou's complex terrain by integrating and correcting single-mode, multi-method approaches. The resulting objective forecast product can provide a reference for improving the quality of local temperature forecasts and enhancing the level of refined forecast services.

[0116] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A temperature forecast correction method based on dynamic weighted integration, characterized in that... ,include: Step 1. Obtain actual surface temperature data and model-predicted temperature data; Step 2. Data Preprocessing: Before temperature correction, the data needs to be preprocessed. Step 3. Neural Network Correction: The 2m temperature forecast products of the CMA-GD model were corrected using methods such as BP neural network, BP neural network optimized by genetic algorithm, generalized regression neural network, wavelet neural network and long short-term memory network. Step 4. Dynamic weight integration: A dynamic weighting scheme is used to integrate the correction results, and the integrated temperature is expressed hourly as follows: ; In the formula, The value 1, 2, ..., 6 represents the order of the forecast results, indicating the model forecast, BP, BP_GA, GRNN, WAVENN, and LSTM forecasts, respectively. Indicates the forecast lead time, with values ​​from 0, 1, ..., 72. For the model forecast temperature and the correction results of each neural network for the j-th hour forecast lead time, For the first The weighting coefficients of each forecast result for the hourly forecast lead time are determined by the historical verification and evaluation results of the forecast temperatures of different forecast lead time models and the correction results of each neural network with the corresponding actual temperatures. The calculation formula is as follows: ; In the formula, The first in the historical test and evaluation sample MAE of each hourly forecast lead time and corresponding actual temperature; Step 4 is followed by a comparison, specifically based on the forecast accuracy. If we take an absolute forecast error of ≤2℃ as the standard for a correct forecast, then the temperature forecast accuracy rate is expressed as the percentage of forecasts with an absolute error of ≤2℃, and the function expression is: ; In the formula, , respectively represent , To predict the correct number of stations, This represents the total number of stations predicted. For the first Station forecast temperature, For the first Station-wide temperature; Step 5. Obtain the final corrected temperature.

2. The temperature forecast correction method based on dynamic weighted integration according to claim 1, characterized in that, Step 2 specifically involves marking outliers as values ​​in the daily ground observation and model forecast data that deviate from three times the absolute deviation of the median. When outliers or missing values ​​appear for three consecutive hours, the samples for that day are removed. If the number of missing samples is less than three hours, linear interpolation is used to fill in the missing data.

3. The temperature forecast correction method based on dynamic weight integration according to claim 1, characterized in that, Step 3 also includes selecting N1 for the training time of BP, BP_GA, GRNN, and WAVENN neural networks, selecting N2 for the training samples of the past N days for the LSTM neural network, and selecting N3 hours for the hourly real-time data of the input field of the LSTM neural network over a period of time. Regarding the selection of N1, N2, and N3 for the training samples, the parameters of each neural network were adjusted using an enumeration method. Considering the prediction accuracy, mean absolute error, and computational efficiency, the values ​​of N1, N2, and N3 were set to 30 days, 10 days, and 24 hours, respectively.

4. The temperature forecast correction method based on dynamic weight integration according to claim 1, characterized in that, Step 4 is followed by a test, specifically a test using the mean absolute error. Temperature mean absolute error The function expression is: ; In the formula, For the first Station forecast temperature, For the first The actual temperature at the station, where N is the total number of stations.

Images