Improving the Short-Range Precipitation Forecast of Numerical Weather Prediction through a Deep Learning-Based Mask Approach

Due to various technical issues,existing numerical weather prediction(NWP)models often perform poorly at forecasting rainfall in the first several hours.To correct the bias of an NWP model and improve the accuracy of short-range precipitation forecasting,we propose a deep learning-based approach called UNet Mask,which combines NWP forecasts with the output of a convolutional neural network called UNet.The UNet Mask involves training the UNet on historical data from the NWP model and gridded rainfall observations for 6-hour precipitation forecasting.The overlap of the UNet output and the NWP forecasts at the same rainfall threshold yields a mask.The UNet Mask blends the UNet output and the NWP forecasts by taking the maximum between them and passing through the mask,which provides the corrected 6-hour rainfall forecasts.We evaluated UNet Mask on a test set and in real-time verification.The results showed that UNet Mask outperforms the NWP model in 6-hour precipitation prediction by reducing the FAR and improving CSI scores.Sensitivity tests also showed that different small rainfall thresholds applied to the UNet and the NWP model have different effects on UNet Mask's forecast performance.This study shows that UNet Mask is a promising approach for improving rainfall forecasting of NWP models.

Current Status and Future Challenges of Weather Radar Polarimetry: Bridging the Gap between Radar Meteorology/Hydrology/Engineering and Numerical Weather Prediction

After decades of research and development, the WSR-88 D(NEXRAD) network in the United States was upgraded with dual-polarization capability, providing polarimetric radar data(PRD) that have the potential to improve weather observations,quantification, forecasting, and warnings. The weather radar networks in China and other countries are also being upgraded with dual-polarization capability. Now, with radar polarimetry technology having matured, and PRD available both nationally and globally, it is important to understand the current status and future challenges and opportunities. The potential impact of PRD has been limited by their oftentimes subjective and empirical use. More importantly, the community has not begun to regularly derive from PRD the state parameters, such as water mixing ratios and number concentrations, used in numerical weather prediction(NWP) models.In this review, we summarize the current status of weather radar polarimetry, discuss the issues and limitations of PRD usage, and explore potential approaches to more efficiently use PRD for quantitative precipitation estimation and forecasting based on statistical retrieval with physical constraints where prior information is used and observation error is included. This approach aligns the observation-based retrievals favored by the radar meteorology community with the model-based analysis of the NWP community. We also examine the challenges and opportunities of polarimetric phased array radar research and development for future weather observation.

The Predictability Problems in Numerical Weather and Climate Prediction

The uncertainties caused by the errors of the initial states and the parameters in the numerical model are investigated. Three problems of predictability in numerical weather and climate prediction are proposed, which are related to the maximum predictable time, the maximum prediction error, and the maximum admissible errors of the initial values and the parameters in the model respectively. The three problems are then formulated into nonlinear optimization problems. Effective approaches to deal with these nonlinear optimization problems are provided. The Lorenz’ model is employed to demonstrate how to use these ideas in dealing with these three problems.

Progresses of Researches on Numerical Weather Prediction in China: 1999-2002

The recent progresses in the research and development of (NWP) in China are reviewed in this paper. The most impressive achievements are the development of direct assimilation of satellite irradiances with a 3DVAR (three-dimentional variational) data assimilation system and a non-hydrostatic modei with a semi-Lagrangian semi-implicit scheme. Progresses have also been made in modei physics and modei application to precipitation and environmental forecasts. Some scientific issues of great importance for further development are discussed.

Numerical Weather Prediction in China in the New Century——Progress,Problems and Prospects

This paper summarizes the recent progress of numerical weather prediction （NWP） research since the last review was published. The new generation NWP system named GRAPES （the Global and Regional Assimilation and Prediction System）, which consists of variational or sequential data assimilation and nonhydrostatic prediction model with options of configuration for either global or regional domains, is briefly introduced, with stress on their scientific design and preliminary results during pre-operational implementation. In addition to the development of GRAPES, the achievements in new methodologies of data assimilation, new improvements of model physics such as parameterization of clouds and planetary boundary layer, mesoscale ensemble prediction system and numerical prediction of air quality are presented. The scientific issues which should be emphasized for the future are discussed finally.

Analogue correction method of errors and its application to numerical weather prediction

In this paper, an analogue correction method of errors （ACE） based on a complicated atmospheric model is further developed and applied to numerical weather prediction （NWP）. The analysis shows that the ACE can effectively reduce model errors by combining the statistical analogue method with the dynamical model together in order that the information of plenty of historical data is utilized in the current complicated NWP model, Furthermore, in the ACE, the differences of the similarities between different historical analogues and the current initial state are considered as the weights for estimating model errors. The results of daily, decad and monthly prediction experiments on a complicated T63 atmospheric model show that the performance of the ACE by correcting model errors based on the estimation of the errors of 4 historical analogue predictions is not only better than that of the scheme of only introducing the correction of the errors of every single analogue prediction, but is also better than that of the T63 model.

An approach to estimating and extrapolating model error based on inverse problem methods:towards accurate numerical weather prediction

Model error is one of the key factors restricting the accuracy of numerical weather prediction （NWP）. Considering the continuous evolution of the atmosphere, the observed data （ignoring the measurement error） can be viewed as a series of solutions of an accurate model governing the actual atmosphere. Model error is represented as an unknown term in the accurate model, thus NWP can be considered as an inverse problem to uncover the unknown error term. The inverse problem models can absorb long periods of observed data to generate model error correction procedures. They thus resolve the deficiency and faultiness of the NWP schemes employing only the initial-time data. In this study we construct two inverse problem models to estimate and extrapolate the time-varying and spatial-varying model errors in both the historical and forecast periods by using recent observations and analogue phenomena of the atmosphere. Numerical experiment on Burgers＇ equation has illustrated the substantial forecast improvement using inverse problem algorithms. The proposed inverse problem methods of suppressing NWP errors will be useful in future high accuracy applications of NWP.

A Forecast Error Correction Method in Numerical Weather Prediction by Using Recent Multiple-time Evolution Data

The initial value error and the imperfect numerical model are usually considered as error sources of numerical weather prediction （NWP）. By using past multi-time observations and model output, this study proposes a method to estimate imperfect numerical model error. This method can be inversely estimated through expressing the model error as a Lagrange interpolation polynomial, while the coefficients of polyno- mial are determined by past model performance. However, for practical application in the full NWP model, it is necessary to determine the following criteria： （1） the length of past data sufficient for estimation of the model errors, （2） a proper method of estimating the term ＂model integration with the exact solution＂ when solving the inverse problem, and （3） the extent to which this scheme is sensitive to the observational errors. In this study, such issues are resolved using a simple linear model, and an advection diffusion model is applied to discuss the sensitivity of the method to an artificial error source. The results indicate that the forecast errors can be largely reduced using the proposed method if the proper length of past data is chosen. To address the three problems, it is determined that （1） a few data limited by the order of the corrector can be used, （2） trapezoidal approximation can be employed to estimate the ＂term＂ in this study; however, a more accurate method should be explored for an operational NWP model, and （3） the correction is sensitive to observational error.

System of Multigrid Nonlinear Least-squares Four-dimensional Variational Data Assimilation for Numerical Weather Prediction(SNAP):System Formulation and Preliminary Evaluation

A new forecasting system-the System of Multigrid Nonlinear Least-squares Four-dimensional Variational(NLS-4DVar)Data Assimilation for Numerical Weather Prediction(SNAP)-was established by building upon the multigrid NLS-4DVar data assimilation scheme,the operational Gridpoint Statistical Interpolation(GSI)−based data-processing and observation operators,and the widely used Weather Research and Forecasting numerical model.Drawing upon lessons learned from the superiority of the operational GSI analysis system,for its various observation operators and the ability to assimilate multiple-source observations,SNAP adopts GSI-based data-processing and observation operator modules to compute the observation innovations.The multigrid NLS-4DVar assimilation framework is used for the analysis,which can adequately correct errors from large to small scales and accelerate iteration solutions.The analysis variables are model state variables,rather than the control variables adopted in the conventional 4DVar system.Currently,we have achieved the assimilation of conventional observations,and we will continue to improve the assimilation of radar and satellite observations in the future.SNAP was evaluated by case evaluation experiments and one-week cycling assimilation experiments.In the case evaluation experiments,two six-hour time windows were established for assimilation experiments and precipitation forecasts were verified against hourly precipitation observations from more than 2400 national observation sites.This showed that SNAP can absorb observations and improve the initial field,thereby improving the precipitation forecast.In the one-week cycling assimilation experiments,six-hourly assimilation cycles were run in one week.SNAP produced slightly lower forecast RMSEs than the GSI 4DEnVar(Four-dimensional Ensemble Variational)as a whole and the threat scores of precipitation forecasts initialized from the analysis of SNAP were higher than those obtained from the analysis of GSI 4DEnVar.

EXPERIMENTAL STUDY OF THE ROLE OF INITIAL AND BOUNDARY CONDITIONS IN MESOSCALE NUMERICAL WEATHER PREDICTION

Based on the real case of a frontal precipitation process affecting South China, 27 controlled numerical experiments was made for the effects of hydrostatic and non-hydrostatic effects, different driving models, combinations of initial/boundary conditions, updates of lateral values and initial time levels of forecast, on model predictions. Features about the impact of initial/boundary conditions on mesoscale numerical weather prediction (NWP) model are analyzed and discussed in detail. Some theoretically and practically valuable conclusions are drawn. It is found that the overall tendency of mesoscale NWP models is governed by its driving model, with the initial conditions showing remarkable impacts on mesoscale models for the first I0 hours of the predictions while leaving lateral boundary conditions to take care the period beyond; the latter affect the inner area of mesoscale predictions mainly through the propagation and movement of weather signals (waves) of different time scales; initial values of external model parameters such as soil moisture content may affect predictions of more longer time validity, while fast signals may be filtered away and only information with time scale 4 times as large as or more than the updated period of boundary values may be introduced, through lateral boundary, to mesoseale models, etc. Some results may be taken as important guidance on mesoseale model and its data a.ssimilation developments of the future.

Statistical downscaling of numerical weather prediction based on convolutional neural networks

Numerical Weather Prediction(NWP)is a necessary input for short-term wind power forecasting.Existing NWP models are all based on purely physical models.This requires mainframe computers to perform large-scale numerical calculations and the technical threshold of the assimilation process is high.There is a need to further improve the timeliness and accuracy of the assimilation process.In order to solve the above problems,NWP method based on artificial intelligence is proposed in this paper.It uses a convolutional neural network algorithm and a downscaling model from the global background field to establish a given wind turbine hub height position.We considered the actual data of a wind farm in north China as an example to analyze the calculation example.The results show that the prediction accuracy of the proposed method is equivalent to that of the traditional purely physical model.The prediction accuracy in some months is better than that of the purely physical model,and the calculation efficiency is considerably improved.The validity and advantages of the proposed method are verified from the results,and the traditional NWP method is replaced to a certain extent.

Parameterized Forward Operators for Simulation and Assimilation of Polarimetric Radar Data with Numerical Weather Predictions

Many weather radar networks in the world have now provided polarimetric radar data(PRD)that have the potential to improve our understanding of cloud and precipitation microphysics,and numerical weather prediction(NWP).To realize this potential,an accurate and efficient set of polarimetric observation operators are needed to simulate and assimilate the PRD with an NWP model for an accurate analysis of the model state variables.For this purpose,a set of parameterized observation operators are developed to simulate and assimilate polarimetric radar data from NWP model-predicted hydrometeor mixing ratios and number concentrations of rain,snow,hail,and graupel.The polarimetric radar variables are calculated based on the T-matrix calculation of wave scattering and integrations of the scattering weighted by the particle size distribution.The calculated polarimetric variables are then fitted to simple functions of water content and volumeweighted mean diameter of the hydrometeor particle size distribution.The parameterized PRD operators are applied to an ideal case and a real case predicted by the Weather Research and Forecasting(WRF)model to have simulated PRD,which are compared with existing operators and real observations to show their validity and applicability.The new PRD operators use less than one percent of the computing time of the old operators to complete the same simulations,making it efficient in PRD simulation and assimilation usage.

Chen-Chao Koo and the Early Numerical Weather Prediction Experiments in China

Although the first successful numerical weather prediction(NWP)project led by Charney and von Neumann is widely known,little is known by the international community about the development of NWP during the 1950s in China.Here,a detailed historical perspective on the early NWP experiments in China is provided.The leadership in NWP of the late Professor Chen-Chao Koo,a protége of C.G.Rossby at the University of Stockholm during the late 1940s and a key leader of modern meteorology(particularly of atmospheric dynamics and physics)in China during the 1950s−70s,is highlighted.The unique contributions to NWP by Koo and his students,such as the ideas of formulating NWP as an“evolution”problem,in which the past data over multiple time steps are utilized,rather than an initial-value problem,and on the cybernetic aspects of atmospheric processes,i.e.,regarding the motion of the atmosphere at various time scales as an optimal control system,are also emphasized.

Hydrological Evaluation with SWAT Model and Numerical Weather Prediction for Flash Flood Warning System in Thailand

Flash floods are a natural disaster that occurs annually, especially in the mountainous terrain and steep slopes of northern Thailand. The current flood forecasting systems and tools are available but have low accuracy and efficiency. The numbers of rainfall and runoff stations are less, because the access to the station area is difficult. Additionally, the operation and maintenance costs are high. Hydrological modeling of a SWAT （Soil and Water Assessment Tool） was used in this study with the application of three days weather forecast from the NWP （numerical weather prediction）, which provided temperature, relative humidity, rainfall, sunshine and wind speed. The data from NWP and SWAT were used to simulate the runoff from the Nan River in the last 10 years （2000-2010）. It was found that the simulated flow rate for the main streams using data from NWP were higher than the observations. At the N64 and Nl stations, the ratios of the maximum simulated flow rate to the observations were equal to 108% and 118%, respectively. However, for the tributaries, it was found that the simulated flow rate using NWP data was lower than the observations, but, it was still within the acceptable range of not greater than 20%,6. At N65, D090201 and D090203 stations, the ratio of the maximum simulated flow rate were 90.0%, 83.0% and 86.0%, respectively. This was due to the rainfall from the NWP model being greater than the measured rainfall. The NWP rainfall was distributed all over the area while the rainfall data from the measurements were obtained from specific points. Therefore, the rain from the NWP model is very useful especially for the watershed areas without rain gauge stations. In summary, the data from the NWP can be used with the SWAT model and provides relatively sound results despite the value for the main river being slightly higher than the observed data. Consequently, the output can be used to create a flood map for flash flood warning in the area.

Development and assessment of artificial neural network models for direct normal solar irradiance forecasting using operational numerical weather prediction data

Accurate operational solar irradiance forecasts are crucial for better decision making by solar energy system operators due to the variability of resource and energy demand.Although numerical weather prediction(NWP)models can forecast solar radiation variables,they often have significant errors,particularly in the direct normal irradiance(DNI),which is especially affected by the type and concentration of aerosols and clouds.This paper presents a method based on artificial neural networks(ANN)for generating operational DNI forecasts using weather and aerosol forecasts from the European Center for Medium-range Weather Forecasts(ECMWF)and the Copernicus Atmospheric Monitoring Service(CAMS),respectively.Two ANN models were designed:one uses as input the predicted weather and aerosol variables for a given instant,while the other uses a period of the improved DNI forecasts before the forecasted instant.The models were developed using observations for the location of´Evora,Portugal,resulting in 10 min DNI forecasts that for day 1 of forecast horizon showed an improvement over the downscaled original forecasts regarding R2,MAE and RMSE of 0.0646,21.1 W/m^(2)and 27.9 W/m^(2),respectively.The model was also evaluated for different timesteps and locations in southern Portugal,providing good agreement with experimental data.

基于机器学习的NWP ZTD长短期预测模型

对流层延迟是影响全球卫星导航系统(GNSS)定位精度的主要误差源之一,利用数值天气预报(NWP)模型估计天顶对流层延迟(ZTD)是常用的方法之一,但NWP模型预报资料估计的ZTD精度有限;NWP模型再分析资料估计的ZTD不能用于GNSS实时定位,且目前大多数文献未能对ZTD长短期预测分别进行研究。因此,利用欧洲中期天气预报中心(ECMWF)的第五代全球气候再分析资料数据集(ERA5)和国际GNSS服务组织(IGS)的高精度ZTD数据,研究基于反向传播(BP)神经网络、支持向量机和长短期记忆网络3种机器学习算法构建以年为时间窗口的ZTD长期预测模型和以24h为时间窗口的ZTD短期预测模型的可行性。实验结果表明:构建的ZTD长期预测模型和短期预测模型可以有效提高预测ZTD的精度。

计及多误差场景集划分的超短期NWP风速修正方法

超短期风电功率预测对于机组运行控制和能源调度有着指导性的作用。为了削弱数值天气预报(numerical weather prediction,NWP)风速对超短期预测精度的影响,提出了一种计及多误差场景集划分的超短期NWP风速修正方法。采用双向长短期记忆网络(bidirectional long-short term memory,BILSTM)对NWP风速未来4 h的预报误差进行预测,对风速误差预测值进行误差场景集划分,根据误差场景集训练不同的BILSTM网络进行误差匹配和风速预报误差预测对风速进行修正,再根据修正结果采用多模型进行风电功率超短期预测。将所提方法应用于中国内蒙古某风电场进行算例验证。结果表明,该方法有效降低了NWP风速预报误差,在原有数据基础上,相较于未修正NWP的风速,RMSE值降低了1.859,MAE值降低了1.464,MAPE值降低了26.01%。其中,BP神经网络超短期功率预测精度提高了7.5%,GRU深度网络提高了8.7%,多元线性回归模型提高了9.6%,证明了该方法的有效性。

基于NWP-LSTM的短期供热负荷预测方法

为提高短期供热负荷预测精度,提出了一种基于数值天气预报(NWP)和长短期记忆神经网络(LSTM)的短期供热负荷预测方法。该方法首先对NWP数据和历史供热负荷数据进行Pearson相关性分析,得出对供热负荷影响较大的天气因素,与历史供热负荷数据一起组成神经网络的输入,并通过反复实验设计出最优结构的NWP-LSTM神经网络模型。通过与其他常见供热负荷预测方法比较,提出的NWP-LSTM模型可以获得更精确的预测结果,适合实际工程应用。

数值预报中气象卫星资料同化前处理技术进展

在数值天气预报变分同化中,利用同化前处理将卫星资料完成有效信息优选、资料拼接和稀疏化、初级通道选择、下边界参数耦合等处理,实现卫星资料同化对数值天气预报业务的正贡献,是决定海量卫星资料同化效率、质量和效果的重要环节。针对多种格式的卫星资料,中国气象局研发标准格式的高时效卫星资料拼接等技术,有效减小整轨卫星资料时间滞后对数值天气预报业务的负面影响。对于风云气象卫星资料,将云和降水检测、资料质量分析等处理置于同化前处理中,实现多光谱资料融合的同化预质量控制,保证了风云卫星微波温度探测资料和红外高光谱资料的同化正贡献。利用统一资料格式对预处理卫星资料进行再处理,拓展针对卫星成像和主动探测资料的处理,将卫星资料同化的部分质量控制功能置于卫星资料同化前处理中,是风云卫星资料同化前处理技术发展的重要趋势。

多源极轨卫星微波温度计资料实时区域同化系统

基于中尺度数值预报模式WRF和WRFDA同化系统,实现多源极轨卫星微波温度计资料实时区域同化,并对同化产品进行评估和应用。2018年同化试验结果表明:通过质量控制和偏差订正,AMSU-A资料第5~9通道亮温观测增量O-B(观测值O和背景场的正演辐射模拟值B的差值)的标准差有效降低,同化后各通道亮温分析残差O-A(观测值O和分析场的正演辐射模拟值A的差值)的标准差有效降低。同化预报产品被应用在暴雨强对流个例和台风个例中,取得良好效果。