Marine low-altitude atmospheric waveguide forecasting method based on artificial intelligence set error correction

Abstract

The invention belongs to the technical field of marine meteorology, and relates to a marine low-altitude atmospheric waveguide forecasting method based on artificial intelligence set error correction. According to the method, by establishing a forecast error correction algorithm set and dynamically adjusting weights of set members, error set correction of forecast results is achieved, and the problems that a single correction algorithm is high in space-time uncertainty and unstable in correction precision are solved. Compared with a pure data-driven low-altitude atmospheric waveguide forecasting method, the low-altitude atmospheric waveguide forecasting method has the advantages that the result reasonability and stability are higher, and the effective forecasting time efficiency is longer.

Description

Technical Field

[0001] The present invention belongs to the technical field of marine meteorology and relates to a method for predicting low-altitude atmospheric ducts based on artificial intelligence ensemble error correction. Background Art

[0002] Atmospheric ducts are special weather phenomena that severely affect the propagation process of electromagnetic waves in the atmosphere due to drastic changes in the vertical gradient of atmospheric refractive index. According to the differences in duct height and profile shape, atmospheric ducts are generally classified into evaporation ducts, surface ducts, and elevated ducts. Among them, surface ducts and elevated ducts with relatively high duct heights are generally collectively referred to as low-altitude atmospheric ducts. The calculation of low-altitude atmospheric ducts usually involves calculating the atmospheric modified refractive index based on elements such as atmospheric temperature and humidity, and then diagnosing indicators such as the height, thickness, and intensity of low-altitude atmospheric ducts based on the vertical gradient changes of the atmospheric modified refractive index.

[0003] Low-altitude atmospheric duct forecasting usually adopts numerical forecasting methods, that is, obtaining the three-dimensional spatial distribution changes of elements such as atmospheric temperature, humidity, and pressure within the next few days based on numerical weather forecast products made by numerical weather models, and diagnosing based on these elements. Finally, the future changes of elements such as the type, height, thickness, intensity, and radio wave band affected by the atmospheric duct in the target area are made. Generally speaking, there are three methods to improve the forecasting accuracy of low-altitude atmospheric ducts: First, expand the observation data sources and develop more advanced data assimilation systems to provide more accurate physical constraints for numerical simulations; Second, improve the numerical model difference method or grid division method, improve the physical process of the numerical model, and achieve the purpose of improving local forecast simulations by coupling more comprehensive physical processes or optimizing key parameters; Third, based on regional historical observation data and hindcast results, establish a mapping relationship between observations and forecast results through various mathematical methods, and use post-processing methods to eliminate the systematic errors of forecast results.

[0004] However, under the condition of sparse observations at sea, it has been very difficult to continue to improve the forecasting accuracy simply by relying on traditional methods such as data assimilation and model optimization. In recent years, the forecasting post-processing method based on artificial intelligence algorithms has provided new ideas for improving low-altitude atmospheric duct forecasting. However, as a data-driven empirical method, the technical approach of using artificial intelligence to correct the forecasting errors of low-altitude atmospheric ducts is difficult to ensure high-precision error correction under all spatio-temporal conditions. Its actual application effects vary greatly in different sea areas, seasons, and weather environments. Currently, no artificial intelligence algorithm can guarantee that the duct correction effect is better than other algorithms under all spatio-temporal conditions. Summary of the Invention

[0005] Based on the problems existing in the current prior art, the present invention provides a method for forecasting low-altitude atmospheric ducts at sea based on artificial intelligence ensemble error correction. This method establishes an ensemble of forecast error correction algorithms through multiple methods such as polynomial fitting, machine learning, and deep learning, and realizes the error ensemble correction of the forecast results by dynamically adjusting the weights of the ensemble members, overcoming the problems of high spatio-temporal uncertainty and unstable correction accuracy of a single correction algorithm.

[0006] The method for forecasting low-altitude atmospheric ducts at sea based on artificial intelligence ensemble error correction provided by the present invention includes the following steps: (1) Construct an atmospheric modified refractive index forecast framework based on the COAWST model for calculating the atmospheric modified refractive index; (2) Conduct local optimization experiments for the target area, with the atmospheric temperature and humidity below 5 km altitude as the main evaluation variables, screen the physical parameterization schemes in the coupled model, evaluate the impacts of changes in cloud microphysical schemes, radiation schemes, cumulus convection schemes, surface layer schemes, planetary boundary layer schemes, wave bottom dissipation schemes of ocean waves, and ocean vertical turbulent mixing schemes on the simulation of atmospheric temperature and humidity, and select the 4-5 most sensitive schemes to further conduct multiple groups of sensitivity experiments; (3) Combine historical observations and reanalysis data, and use the simulation observation correlation coefficient and root mean square error as evaluation indicators to evaluate the simulation capabilities of various option combinations of the schemes; (4) Generate a parameterization scheme configuration for the coupled model to run based on the sensitivity evaluation results, and calculate the atmospheric modified refractive index; (6) Construct an ensemble of error correction algorithm models; and conduct an atmospheric refractive index forecast correction evaluation experiment, and determine the final ensemble members according to the evaluation results; (7) Adopt the Bayesian Model Averaging (BMA) algorithm, perform ensemble averaging on all ensemble members of the error correction algorithm based on the performance of each algorithm during the historical period, assign initial weights, and then dynamically adjust the weights according to the subsequent error correction performance; distribute the weights of each member according to the posterior probabilities of all ensemble members after normalization, and then perform weighted summation to obtain the final ensemble forecast result of the atmospheric modified refractive index; (8) Conduct low-altitude atmospheric duct diagnosis based on the ensemble forecast result of the atmospheric modified refractive index to obtain the waveguide characteristic diagnosis result.

[0007] Preferably, the error correction algorithm is selected from any combination of two or more of polynomial fitting, machine learning algorithms, and deep learning algorithms.

[0008] Preferably, step (7) specifically includes: taking days as the unit, using the observation data of the past 120 hours, using the Expectation-Maximization (EM) algorithm to solve the maximum likelihood estimation, and converting the observation data into a normal distribution; based on the forecast results corrected by errors in the same period, using the Markov Chain Monte Carlo (MCMC) method, calculating the marginal likelihood function by constructing the ideas of Markov chain and Monte Carlo integration, and statistically calculating the posterior probability of each member of the error correction algorithm; according to the posterior probabilities of all members, distributing the weights of each member after normalization, and then performing weighted summation to obtain the final ensemble forecast result of the atmospheric modified refractive index.

[0009] The method of the present invention combines a numerical weather prediction method with empirical statistical post-processing technology. Compared with a purely data-driven low-altitude atmospheric duct prediction method, its results are more reasonable and stable, and the effective prediction time is longer. Compared with the traditional low-altitude atmospheric duct numerical prediction method, its prediction results have higher accuracy after error correction and stronger adaptability to different environmental changes. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 is a schematic flow chart of a low-altitude atmospheric duct prediction method based on artificial intelligence ensemble error correction provided in an embodiment of the invention; Figure 2 is an error correction sequence diagram of the atmospheric modified refractive index at a height of 869 m calculated by different correction algorithms and the BMA method; Figure 3 is an error correction sequence diagram of the atmospheric modified refractive index at a height of 1046 m calculated by different correction algorithms and the BMA method; Figure 4 is an error correction sequence diagram of the atmospheric modified refractive index at a height of 1342 m calculated by different correction algorithms and the BMA method; Figure 5 is the root mean square error of the atmospheric modified refractive index calculated by different correction algorithms and the BMA method. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0011] For the convenience of understanding the present invention, the present invention will be described in more detail below with reference to the drawings and specific embodiments. The preferred embodiments of the present invention are shown in the drawings. However, the present invention can be implemented in many different forms and is not limited to the embodiments described in this specification. On the contrary, the purpose of providing these embodiments is to make the understanding of the disclosed content of the present invention more thorough and comprehensive.

[0012] The low-altitude atmospheric duct prediction method based on artificial intelligence ensemble error correction provided by the present invention is implemented as Figure 1 shown, and specifically includes the following steps: First, based on the COAWST air-sea coupled model, a forecasting operational framework for the atmospheric modified refractive index within the forecasting region is established. According to the regularly released data of the GFS global atmospheric forecast field, GOFS global ocean forecast field, and WaveWatch III global ocean wave operational forecast field, unified processing is carried out in terms of data format conversion, regional cutting, and projection transformation interpolation to provide the initial field and boundary field driving for each component model of COAWST. In the COAWST model, the three component models of the atmosphere, ocean, and ocean waves perform numerical simulations independently, and at each exchange time step, meteorological and hydrological variables calculated by each component model are provided to the model coupler, and variables calculated by other component models are read as inputs. Based on the simulation results, the distributions of elements such as air pressure, humidity, and air temperature within the region are extracted to calculate the atmospheric modified refractive index.

[0013] The atmospheric refractive index can be calculated by the following formula: ; where N represents the refractive index (unit: N-units), T represents the temperature (unit: K), P represents the air pressure (unit: hPa), and e represents the water vapor pressure, which can be calculated by the following formula: ; where q represents the specific humidity (unit: g / kg), and ε is a constant (0.622); ; where M represents the modified atmospheric refractive index (unit: M-units), R represents the radius of the Earth (6371 km), and h represents the height (unit: m).

[0014] Secondly, local optimization experiments are carried out for the target region to obtain the parameterization scheme configuration with the optimal simulation accuracy. Focusing on the atmospheric temperature and humidity below 5 km height as the main evaluation variables, the physical parameterization schemes in the coupled model are screened, and the impacts of changes in the cloud microphysics scheme, radiation scheme, cumulus convection scheme, surface layer scheme, planetary boundary layer scheme, ocean wave bottom dissipation scheme, and ocean vertical turbulent mixing scheme on the simulation of atmospheric temperature and humidity are evaluated. Four to five schemes with the strongest sensitivity are selected to further carry out multiple groups of sensitivity experiments. Subsequently, combined with historical observations and reanalysis data, taking the simulation-observation correlation coefficient and root mean square error as evaluation indicators, the simulation capabilities of various scheme option combinations are evaluated. According to the sensitivity evaluation results, a parameterization scheme configuration is generated for the operation of the coupled model, and forecasting information such as the atmospheric refractive index is calculated. With the optimized parameterization scheme configuration, historical long-term hindcast simulation experiments are carried out to produce a training data set for error correction modeling.

[0015] Subsequently, based on the historical long-term return simulation data of the atmospheric modified refractive index elements as the forecast training data, and the atmospheric modified refractive index calculated from the sounding, analysis field, and reanalysis data in the same period within the region as the observation reference data, an error correction algorithm set is constructed. In terms of the selection of error correction algorithms, algorithms with a large difference in principles are selected to construct the error correction model. The algorithms are divided into three categories. The first category is the traditional arithmetic method, which uses mathematical methods such as polynomial fitting (e.g., the Model Output Statistics MOS method) for error correction; the second category is the classical machine learning algorithms, such as the BP neural network, random forest, and support vector machine methods; the third category is the typical deep learning algorithms, such as the Convolutional Neural Network CNN, Long Short-Term Memory LSTM, and Gated Recurrent Unit GRU algorithms.

[0016] Based on the above algorithms, an atmospheric refractive index forecast correction evaluation experiment is carried out. To ensure the diversity of the ensemble members, algorithms with extremely similar correction results are screened out and replaced with other algorithms (e.g., Extreme Gradient Boosting XGBoost, Generative Adversarial Network GAN). After evaluation, the error correction ensemble members are finally determined to carry out ensemble calculation.

[0017] During the ensemble calculation, the Bayesian Model Averaging BMA algorithm is adopted. According to the performance of each algorithm in the historical period, the ensemble average is performed on all error correction algorithm ensemble members, initial weights are assigned, and dynamic weight adjustment is carried out according to the subsequent error correction performance. The ensemble average framework realizes the inclination of the ensemble result to the excellent members by assigning different conditional probability weights to each member.

[0018] The specific implementation steps are as follows: Taking days as the unit, using the observation data of the past 120 hours, the Expectation-Maximization EM algorithm is used to solve the maximum likelihood estimation, and the observation data is converted into a normal distribution; subsequently, based on the forecast results corrected by errors in the same period, the Markov Chain Monte Carlo MCMC method is adopted, and the marginal likelihood function is calculated through the idea of constructing a Markov chain and Monte Carlo integration, and the posterior probability of each error correction algorithm member is statistically calculated respectively; finally, according to the posterior probabilities of all members, weights are assigned to each member after normalization, and then weighted summation is carried out to obtain the final ensemble forecast result of the atmospheric modified refractive index: ; where yBMA is the ensemble forecast result; t is a variable, here it is the atmospheric modified refractive index; Mk is the model member numbered k; Pr(Mk|D) is the posterior probability of Mk under the condition of the given data sample D.

[0019] Finally, based on the atmospheric modified refractive index forecast result, the low-altitude atmospheric duct diagnosis is carried out to produce the duct characteristic diagnosis product. The specific steps are as follows: For the atmospheric modified refractive index profile of each grid point, starting from the sea surface h0 (height is 0), the diagnosis is carried out layer by layer upwards. The specific process includes: (1) Search upward to correct the appearance of the first maximum peak of the refractive index and record the current height h1 and the current corrected refractive index M1. If there is no h1, it is determined that no waveguide occurs and the diagnosis is completed; (2) If h1 exists, continue to search upward to determine the first minimum value of the corrected refractive index. Then record the current height h2 and the corrected refractive index M2; (3) If there is no increasing interval of the corrected refractive index within the height range from h0 to h1, it is determined that the current waveguide type is a surface waveguide without a base layer. At this time, the bottom height of the waveguide is 0 m, the top height is h2, the waveguide strength is M1–M2, and the waveguide thickness is h2–h0; (4) If there is an increasing interval of the corrected refractive index at heights below h1, search downward from h1 to find the first point where the corrected refractive index is less than M2 and record the height h3 of this point; (5) If no point less than M2 is found and h0 is still the initial height of 0 m, the current waveguide is identified as a surface waveguide with a base layer. At this time, the bottom height of the waveguide is 0 m, the top height is h2, the waveguide strength is M1–M2, and the waveguide thickness is h2–h0; (6) If a point less than M2 is found, record the height h3 of this point and determine the current waveguide type as a suspended waveguide. The bottom height of the waveguide is h3, the top height is h2, the waveguide strength is M1–M2, and the waveguide thickness is h2–h3; (7) After completing the waveguide diagnosis process for this layer, replace the initial height h0 with the height h2. Then return to step (2) and continue to search upward to diagnose the waveguide of the next layer until the maximum height is reached.

[0020] After the diagnosis is completed, output parameters such as the waveguide type, height, thickness, and strength of each grid point at each moment, and finally form the forecast data of the low-altitude atmospheric waveguide in the area.

[0021] To evaluate the performance of the method of the present invention, numerical simulations are carried out on the meteorological data for the whole year of 2021 based on the COAWST air-sea coupled model. According to the simulation results, elements such as air pressure, humidity, and temperature in the area are extracted, the atmospheric corrected refractive index at different heights is calculated, and the atmospheric corrected refractive index at different heights in the synchronous sounding observation data in the area is calculated as the reference value. Five error correction algorithms (LSTM, GRU, CNN, BP, MOS) are used to correct the error of the atmospheric corrected refractive index at different heights, and the Bayesian Model Averaging (BMA) method is innovatively applied to perform ensemble averaging on the correction results of multiple models. The results show that compared with a single error correction algorithm, the BMA ensemble method shows better correction effects at different height layers, significantly improving the accuracy of the atmospheric corrected refractive index, as shown in Figure 2 -4.

[0022] As Figure 5 shown, the root mean square error (RMSE) between the correction results of five error correction algorithms (LSTM, GRU, CNN, BP, MOS) and the Bayesian model averaging (BMA) method and the sounding observation data of the same period was calculated respectively, and it was found that the BMA method can significantly improve the calculation accuracy of the atmospheric modified refractive index by effectively integrating the advantages of each single model.

Claims

1. A method for predicting low-altitude atmospheric ducts over the sea based on artificial intelligence ensemble error correction, characterized in that The following steps are involved: (1) Construct an atmospheric corrected refractivity forecast framework based on the COAWST model to calculate the atmospheric corrected refractivity; (2) Carry out local optimization experiments in the target area, take the atmospheric temperature and humidity below 5 km as the main evaluation variables, screen the physical parameterization schemes in the coupled model, evaluate the impact of changes in cloud microphysics scheme, radiation scheme, cumulus convection scheme, surface layer scheme, planetary boundary layer scheme, wave bottom dissipation scheme, and ocean vertical turbulence mixing scheme on the simulation of atmospheric temperature and humidity, and select the 4 to 5 schemes with the strongest sensitivity to further carry out multiple groups of sensitivity experiments; (3) Combine historical observations and reanalysis data, use the simulated observation correlation coefficient and root mean square error as evaluation indicators to evaluate the simulation capability of each option combination; (4) Based on the sensitivity assessment results, generate parameterized scheme configurations for coupled model operation and calculate the atmospheric corrected refractivity; (6) Construct an ensemble of error correction algorithm models; conduct an atmospheric refractivity forecast correction evaluation test, and determine the final ensemble members based on the evaluation results; (7) Using the Bayesian model average (BMA) algorithm, all error correction algorithm ensemble members are averaged according to the performance of each algorithm in the historical period, and initial weights are assigned. The weights are then dynamically adjusted according to the subsequent error correction performance. The weights of each member are assigned according to the posterior probabilities of all ensemble members after normalization, and then the weighted sum is taken to obtain the final atmospheric corrected refractivity ensemble forecast result. (8) Based on the atmospheric corrected refractive index ensemble forecast results, low-altitude atmospheric waveguide diagnosis is performed to obtain waveguide characteristic diagnosis results.

2. The method for predicting the offshore low-altitude atmospheric duct based on artificial intelligence set error correction according to claim 1, wherein The error correction algorithm is selected from a combination of any two or more of a polynomial fitting algorithm, a machine learning algorithm, and a deep learning algorithm.

3. The method for predicting low-altitude atmospheric ducts over the sea based on artificial intelligence set error correction according to claim 1, wherein Step (7) specifically includes: taking the observation data of the past 120 hours as a unit, using the maximum expectation EM algorithm to solve the maximum likelihood estimation, and converting the observation data into a normal distribution; based on the forecast results after error correction in the same period, using the Markov Monte Carlo MCMC method, by constructing the Markov chain and Monte Carlo integration to calculate the marginal likelihood function, and statistically calculating the posterior probability of each error correction algorithm member; according to the posterior probabilities of all members, assigning weights to each member after normalization, and then weighted summing them to obtain the final atmospheric corrected refractive index ensemble forecast result.

Images