Systematic Review on Ground-Based Cloud Tracking Methods for Photovoltaics Nowcasting

Renewable energies are highly dependent on local weather conditions, with photovoltaic energy being particularly affected by intermittent clouds. Anticipating the impact of cloud shadows on power plants is crucial, as clouds can cause partial shading, excessive irradiation, and operational issues. This study focuses on analyzing cloud tracking methods for short-term forecasts, aiming to mitigate such impacts. We conducted a systematic literature review, highlighting the most significant articles on cloud tracking from ground-based observations. We explore both traditional image processing techniques and advances in deep learning models. Additionally, we discuss current challenges and future research directions in this rapidly evolving field, aiming to provide a comprehensive overview of the state of the art and identify opportunities for significant advancements in the next generation of cloud tracking systems based on computer vision and deep learning.

Advances in Deep-Learning-based Precipitation Nowcasting Techniques

Precipitation nowcasting,as a crucial component of weather forecasting,focuses on predicting very short-range precipitation,typically within six hours.This approach relies heavily on real-time observations rather than numerical weather models.The core concept involves the spatio-temporal extrapolation of current precipitation fields derived from ground radar echoes and/or satellite images,which was generally actualized by employing computer image or vision techniques.Recently,with stirring breakthroughs in artificial intelligence(AI)techniques,deep learning(DL)methods have been used as the basis for developing novel approaches to precipitation nowcasting.Notable progress has been obtained in recent years,manifesting the strong potential of DL-based nowcasting models for their advantages in both prediction accuracy and computational cost.This paper provides an overview of these precipitation nowcasting approaches,from which two stages along the advancing in this field emerge.Classic models that were established on an elementary neural network dominated in the first stage,while large meteorological models that were based on complex network architectures prevailed in the second.In particular,the nowcasting accuracy of such data-driven models has been greatly increased by imposing suitable physical constraints.The integration of AI models and physical models seems to be a promising way to improve precipitation nowcasting techniques further.

Evaluation of the Added Value of Probabilistic Nowcasting Ensemble Forecasts on Regional Ensemble Forecasts

Ensemble forecasting systems have become an important tool for estimating the uncertainties in initial conditions and model formulations and they are receiving increased attention from various applications.The Regional Ensemble Prediction System(REPS),which has operated at the Beijing Meteorological Service(BMS)since 2017,allows for probabilistic forecasts.However,it still suffers from systematic deficiencies during the first couple of forecast hours.This paper presents an integrated probabilistic nowcasting ensemble prediction system(NEPS)that is constructed by applying a mixed dynamicintegrated method.It essentially combines the uncertainty information(i.e.,ensemble variance)provided by the REPS with the nowcasting method provided by the rapid-refresh deterministic nowcasting prediction system(NPS)that has operated at the Beijing Meteorological Service(BMS)since 2019.The NEPS provides hourly updated analyses and probabilistic forecasts in the nowcasting and short range(0-6 h)with a spatial grid spacing of 500 m.It covers the three meteorological parameters:temperature,wind,and precipitation.The outcome of an evaluation experiment over the deterministic and probabilistic forecasts indicates that the NEPS outperforms the REPS and NPS in terms of surface weather variables.Analysis of two cases demonstrates the superior reliability of the NEPS and suggests that the NEPS gives more details about the spatial intensity and distribution of the meteorological parameters.

Convective Storm VIL and Lightning Nowcasting Using Satellite and Weather Radar Measurements Based on Multi-Task Learning Models

Convective storms and lightning are among the most important weather phenomena that are challenging to forecast.In this study,a novel multi-task learning(MTL)encoder-decoder U-net neural network was developed to forecast convective storms and lightning with lead times for up to 90 min,using GOES-16 geostationary satellite infrared brightness temperatures(IRBTs),lightning flashes from Geostationary Lightning Mapper(GLM),and vertically integrated liquid(VIL)from Next Generation Weather Radar(NEXRAD).To cope with the heavily skewed distribution of lightning data,a spatiotemporal exponent-weighted loss function and log-transformed lightning normalization approach were developed.The effects of MTL,single-task learning(STL),and IRBTs as auxiliary input features on convection and lightning nowcasting were investigated.The results showed that normalizing the heavily skew-distributed lightning data along with a log-transformation dramatically outperforms the min-max normalization method for nowcasting an intense lightning event.The MTL model significantly outperformed the STL model for both lightning nowcasting and VIL nowcasting,particularly for intense lightning events.The MTL also helped delay the lightning forecast performance decay with the lead times.Furthermore,incorporating satellite IRBTs as auxiliary input features substantially improved lightning nowcasting,but produced little difference in VIL forecasting.Finally,the MTL model performed better for forecasting both lightning and the VIL of organized convective storms than for isolated cells.

Segmentation and Classification of Individual Clouds in Images Captured with Horizon-Aimed Cameras for Nowcasting of Solar Irradiance Absorption

One important aspect of solar energy generation especially in inter-tropical sites is the local variability of clouds. Satellite images do not have temporal resolution enough to nowcast its impacts on solar plants, this monitoring is made by local cameras. However, cloud detection and monitoring are not trivial due to cloud shape dynamics, the camera is a linear and self-adjusting device, with fish-eye lenses generating a flat image that distorts images near the horizon. The present work focuses on cloud identification to predict its effects on solar plants that are distinct for every site’s climatology and geography. We used RASPBERY-PI-based cameras pointed at the horizon to allow observation of clouds’ vertical distribution, not possible with a unique fish-eye lens. A large number of cloud image identification analyses led the researchers to use deep learning methods such as U-net, HRnet, and Detectron. We use transfer learning with weights trained over the “2012 ILSVRC ImageNet” data set and architecture configurations like Resnet, Efficient, and Detectron2. While cloud identification proved a difficult task, we achieved the best results by using Jaccard Coefficient as a validation metric, with the best model being a U-net with Resnet18 using 486 × 648 resolution. This model had an average IoU of 0.6, indicating a satisfactory performance in cloud segmentation. We also observed that the data imbalance affected the overall performance of all models, with the tree class creating a favorable bias. The HRNet model, which works with different resolutions, showed promising results with a more refined segmentation at the pixel level, but it was not necessary to detect the most predominant clouds in the sky. We are currently working on balancing the dataset and mapping out data augmentation transformations for our next experiments. Our ultimate goal is to use such models to predict cloud motion and forecast the impact it will have on solar power generation. The present work has contributed to a better understanding of what techniques work best for cloud identification and paves the way for future studies on the development of a better overall cloud classification model.

Application of Multi-Scale Tracking Radar Echoes Scheme in Quantitative Precipitation Nowcasting

A new radar echo tracking algorithm known as multi-scale tracking radar echoes by cross-correlation （MTREC） was developed in this study to analyze movements of radar echoes at different spatial scales. Movement of radar echoes, particularly associated with convective storms, exhibits different characteristics at various spatial scales as a result of complex interactions among meteorological systems leading to the formation of convective storms. For the null echo region, the usual correlation technique produces zero or a very small magnitude of motion vectors. To mitigate these constraints, MTREC uses the tracking radar echoes by correlation （TREC） technique with a large ＂box＂ to determine the systematic movement driven by steering wind, and MTREC applies the TREC technique with a small ＂box＂ to estimate small-scale internal motion vectors. Eventually, the MTREC vectors are obtained by synthesizing the systematic motion and the small-scale internal motion. Performance of the MTREC technique was compared with TREC technique using case studies： the Khanun typhoon on 11 September 2005 observed by Wenzhou radar and a squall-line system on 23 June 2011 detected by Beijing radar. The results demonstrate that more spatially smoothed and continuous vector fields can be generated by the MTREC technique, which leads to improvements in tracking the entire radar reflectivity pattern. The new multi-scMe tracking scheme was applied to study its impact on the performance of quantitative precipitation nowcasting. The location and intensity of heavy precipitation at a 1-h lead time was more consistent with quantitative precipitation estimates using radar and rain gauges.

Methods of Lightning Nowcasting Based on Radar Echo Extrapolation Technology

An improved echo extrapolation technology( MOD-COTREC) was introduced firstly,and then two plans for lightning nowcasting based on MOD-COTREC and both isothermal radar reflectivity and MOD-COTREC were proposed based on the technology. Afterwards,the two plans for lightning nowcasting were tested by a case respectively. It is concluded that during the process of lightning nowcasting singly based on MOD-COTREC,the appearance and disappearance of lightning are not considered,and only lightning position is predicted when lightning density is constant,so the prediction error is big. The plan for lightning nowcasting based on both isothermal radar reflectivity and MOD-COTREC is still at an experimental stage,and the nowcasting products of cloud-to-ground lightning based on the plan are very different from the actual density and position of cloud-to-ground lightning,so it needs to be improved further.

Development of typhoon driven wave nowcasting model in Southeast China Sea

Using optimal interpolation data assimilation of observed wave spectrum around Northeast coast of Taiwan Island, the typhoon driven wave nowcasting model in Southeast China Sea is setup. The SWAN (simulating waves nearshore) model is used to calculate wave field and the input wind field is the QSCAT/NCEP (Quick Scatterometer/National Centers for Environmental Prediction) data. The two-dimensional wavelet transform is applied to analyze the X-band radar image of nearshore wave field and it reveals that the observed wave spectrum has shoaling characteristics in frequency domain. The reverse calculation approach of wave spectrum in deep water is proposed and validated with experimental tests. The two-dimensional digital low-pass filter is used to obtain the initialization wave field. Wave data during Typhoon Sinlaku is used to calibrate the data assimilation parameters and test the reverse calculation approach. Data assimilation corrects the significant wave height and the low frequency spectra energy evidently at Beishuang Station along Fujian Province coast, where the entire assimilation indexes are positive in verification moments. The nowcasting wave field shows that the present model can obtain more accurate wave predictions for coastal and ocean engineering in Southeast China Sea.

Summary on Applications of Stratiform Clear Air Echo in Nowcasting

It was difficult to probe the clear air echo by the general traditional radar for echo's weak intensity.Therefore,its investigation was less because of the restrictions of probe technique and data.In recent years,with the probe tools improving,more clear air echoes were probed,and the relative investigations were more and more.However,most investigations stayed in the theory at present,and the relative literatures about its application in the practical forecast work were few.For a new generation of Doppler radars' powers and sensitivities were all high,they were put into service successively in China.People could observe more and more the clear air atmospheric echoes in the daily business.Its Doppler radar velocity provided the important basis for daily short-term predication and had very important indication meaning for the nowcasting of seasons which were spring,summer and fall.It was important to forecast the precipitation,especially the abrupt rainstorm by using the symptom of clear air echo which was probed by the new generation of Doppler radar products.Therefore,the advances on clear air echo research at home and abroad were summarized simply.

Lightning Nowcasting with an Algorithm of Thunderstorm Tracking Based on Lightning Location Data over the Beijing Area

A thunderstorm tracking algorithm is proposed to nowcast the possibility of lightning activity over an area of concern by using the total lightning data and neighborhood technique.The lightning radiation sources observed from the Beijing Lightning Network(BLNET)were used to obtain information about the thunderstorm cells,which are significantly valuable in real-time.The boundaries of thunderstorm cells were obtained through the neighborhood technique.After smoothing,these boundaries were used to track the movement of thunderstorms and then extrapolated to nowcast the lightning approaching in an area of concern.The algorithm can deliver creditable results prior to a thunderstorm arriving at the area of concern,with accuracies of 63%,80%,and 91%for lead times of 30,15,and 5 minutes,respectively.The real-time observations of total lightning appear to be significant for thunderstorm tracking and lightning nowcasting,as total lightning tracking could help to fill the observational gaps in radar reflectivity due to the attenuation by hills or other obstacles.The lightning data used in the algorithm performs well in tracking the active thunderstorm cells associated with lightning activities.

THE INFLUENCE OF CLOUD PARAMETERIZATION ADJUSTMENT USING REFLECTIVITY OF DOPPLER ON NOWCASTING WITH GRAPES MODEL

In this study, we attempted to improve the nowcasting of GRAPES model by adjusting the model initial field through modifying the cloud water, rain water and vapor as well as revising vapor-following rain water. The results show that the model nowcasting is improved when only the cloud water and rain water are adjusted or all of the cloud water, rain water and vapor are adjusted in the initial field. The forecasting of the former(latter) approach during 0-3(0-6) hours is significantly improved. Furthermore, for the forecast for 0-3 hours, the latter approach is better than the former. Compared with the forecasting results for which the vapor of the model initial field is adjusted by the background vapor with those by the revised vapor, the nowcasting of the revised vapor is much better than that of background vapor. Analysis of the reasons indicated that when the vapor is adjusted in the model initial field, especially when the saturated vapor is considered, the forecasting of the vapor field is significantly affected. The changed vapor field influences the circulation, which in turn improves the model forecasting of radar reflectivity and rainfall.

Modelling the ZR Relationship of Precipitation Nowcasting Based on Deep Learning

Sudden precipitations may bring troubles or even huge harm to people’s daily lives.Hence a timely and accurate precipitation nowcasting is expected to be an indispensable part of our modern life.Traditionally,the rainfall intensity estimation from weather radar is based on the relationship between radar reflectivity factor(Z)and rainfall rate(R),which is typically estimated by location-dependent experiential formula and arguably uncertain.Therefore,in this paper,we propose a deep learning-based method to model the ZR relation.To evaluate,we conducted our experiment with the Shenzhen precipitation dataset.We proposed a combined method of deep learning and the ZR relationship,and compared it with a traditional ZR equation,a ZR equation with its parameters estimated by the least square method,and a pure deep learning model.The experimental results show that our combined model performsmuch better than the equation-based ZRformula and has the similar performance with a pure deep learning nowcasting model,both for all level precipitation and heavy ones only.

A Novel Method for Precipitation Nowcasting Based on ST-LSTM

Precipitation nowcasting is of great significance for severe convective weather warnings.Radar echo extrapolation is a commonly used precipitation nowcasting method.However,the traditional radar echo extrapolation methods are encountered with the dilemma of low prediction accuracy and extrapolation ambiguity.The reason is that those methods cannot retain important long-term information and fail to capture short-term motion information from the long-range data stream.In order to solve the above problems,we select the spatiotemporal long short-term memory(ST-LSTM)as the recurrent unit of the model and integrate the 3D convolution operation in it to strengthen the model’s ability to capture short-term motion information which plays a vital role in the prediction of radar echo motion trends.For the purpose of enhancing the model’s ability to retain long-term important information,we also introduce the channel attention mechanism to achieve this goal.In the experiment,the training and testing datasets are constructed using radar data of Shanghai,we compare our model with three benchmark models under the reflectance thresholds of 15 and 25.Experimental results demonstrate that the proposed model outperforms the three benchmark models in radar echo extrapolation task,which obtains a higher accuracy rate and improves the clarity of the extrapolated image.

3DuA-Net:融合3D卷积和注意力的雷达回波外推预报

针对短临降雨预测模型对历史雷达数据的建模结果存在视觉性能模糊和低估高回波值的问题,提出一种融合3D卷积和双端注意力机制的短临降雨雷达回波外推模型3DuA-Net。以ST-LSTM时空长短期记忆网络为循环单元,将普通卷积替换为3D卷积,增强模型从全局视角强化对短期运动特征信息的捕获能力。并提出DuAtt高效双端注意力机制,提高模型对长期雷达图像序列局部和全局重要特征信息的保存及结合能力。采用深圳气象局公开的多普勒雷达数据集进行实验,结果表明:在10、20、40 dBz阈值下,该模型相比Conv-LSTM基线模型的CSI指标平均提升7.74%,HSS指标平均提升5.54%,MAE指标下降3.8%,SSIM指标提升8.86%。

基于时空图卷积的强对流降水临近预报研究

降水临近预报对于强对流天气的预报具有重要的支撑作用。气象业务中主要采用雷达回波外推方法解决此问题。然而,现有方法通常缺乏从序列雷达数据中有效学习的能力,导致预测精度不佳。为了解决这一问题,本文提出了一种改进的时空图卷积模型ASTGCN(A Spatio-Temporal Graph Convolution Neural Network)用于强对流降水的临近预报。利用时空图卷积网络,有效地捕获相邻雷达帧之间的时间依赖性。此外,利用注意力机制和自动编码器来增强模型捕获时空相关性的能力。结果表明,该模型可以从数据中发现隐藏的图结构,从而捕获隐藏的空间关系。与现有模型(Transformer)相比,该模型的临界成功指数(CSI)提高了28%,表明其在强对流降水临近预报方面具有优越的性能。

基于PhyDNet-ATT的能见度预报方法

本文采用PhyDNet-ATT深度学习方法建立江苏省能见度预报模型PhyDNet-ATT-VIS,该模型融合了高时空分辨率地面观测数据和模式产品,实现了空间分辨率为3 km、时间分辨率为1 h、预报时效为6~18 h的能见度预报,并且对模型结果进行了检验评估。与ECMWF能见度产品相比,PhyDNet-ATT-VIS预报的均方根误差和平均绝对偏差分别降低201%和310%;对于不同能见度等级,命中率显著提高,空报率显著降低,TS评分显示预报技巧优势明显,但15~18 h低能见度预报仍存在很大提升空间。PhyDNet-ATT-VIS在观测站点密集区域的预报误差显著低于观测站点稀疏区域。在区域性雾过程和局地雾过程预报方面,PhyDNet-ATT-VIS均能较准确地预报雾的落区、强度、生消等关键特征参数。研究为能见度短时临近预报技巧的提升提供了新思路。

Quantitative Applications of Weather Satellite Data for Nowcasting:Progress and Challenges

Monitoring and predicting highly localized weather events over a very short-term period,typically ranging from minutes to a few hours,are very important for decision makers and public action.Nowcasting these events usually relies on radar observations through monitoring and extrapolation.With advanced high-resolution imaging and sounding observations from weather satellites,nowcasting can be enhanced by combining radar,satellite,and other data,while quantitative applications of those data for nowcasting are advanced through using machine learning techniques.Those applications include monitoring the location,impact area,intensity,water vapor,atmospheric instability,precipitation,physical properties,and optical properties of the severe storm at different stages(pre-convection,initiation,development,and decaying),identification of storm types(wind,snow,hail,etc.),and predicting the occurrence and evolution of the storm.Satellite observations can provide information on the environmental characteristics in the preconvection stage and are very useful for situational awareness and storm warning.This paper provides an overview of recent progress on quantitative applications of satellite data in nowcasting and its challenges,and future perspectives are also addressed and discussed.

光流法雷达外推产品在突发强降水预报中的应用

地形条件复杂的山丘区中小河流洪水突发性强、汇流时间短,高准确率和长时效性降水短时临近预报产品对提高突发洪水预报精度尤为关键。以2021年9月河南省鸭河口水库出现的千年一遇特大洪水为例,利用国家气象信息中心提供的三源融合格点实况降水资料,检验基于改进光流法的雷达外推降水预报产品在本次洪水过程中0~1 h和0~3 h降水预报的TS评分和预报偏差。结果表明:(1)改进光流法在0~1 h的逐小时降水预报上较为精准,累计雨量在50 mm以下时,TS评分在0.45~0.85之间;雨量在50~70 mm之间时,TS评分在0.35~0.70之间;雨量在70 mm以上时,TS评分在0.25~0.35之间。50 mm以上雨量有较高TS评分表现出改进光流法在极端强降水预报中的优势性。(2)改进光流法在0~3 h的降水预报上,累计雨量在50 mm以下时,TS评分在0.55~0.85之间;在50 mm以上时,TS评分在0.35~0.75之间。该降水预报产品不仅对极端性降水预报效果较好,且预报时效长达3 h,可为防洪调度提供更长的决策时间。(3)改进光流法在0~3 h的降水预报产品与融合实况格点降水相比,雨量在20 mm以下的预报结果比较接近,平均绝对误差在10 mm以内;雨量在20 mm以上时,随雨量增大,平均误差、平均绝对误差、均方根误差均逐渐增大。(4)改进光流法在0~3 h的降水预报产品对影响范围小、降水强度大、维持时间长、累计雨量大的极端强降水有较好的预报表现。研究成果可为洪水预报模型提供一种较为可靠的降水输入预报。

基于分层生成对抗网络的短临降水预报方法研究

本文使用深度学习方法中的生成对抗网络(GAN)来提升短临降水预报的准确率,提出了一个基于历史雷达回波图序列预测未来雷达回波图序列的分层生成对抗网络(HGAN)方法。HGAN方法由全局生成器和局部鉴别器两部分组成,全局生成器以多子网的层次结构构建,采用上采样过程训练模型,捕捉雷达回波的演变趋势,有利于生成清晰的未来雷达回波图。局部鉴别器根据局部区域将预测的雷达回波图与观测的雷达回波图区分开,并引入缓冲区机制,保存历史预测序列,使最终预测的结果更加符合时序性。两者以对抗的方式加以训练,得到的模型能够生成足够清晰且接近真实的未来雷达回波序列,对于回波强度极值和范围的刻画更为准确。对HGAN和GAN进行测试集检验及个例分析,分析结果验证了HGAN对雷达回波预测的有效性。同时在检验反射率阈值相同的情况下,HGAN的临界成功指数命中率高于GAN,而虚警率低于GAN,且在相同预测时长下,HGAN结构相似性指数(SSIM)优于GAN。

基于双向Transformer的降水临近预报模型

精准的降水临近预报对日常生活至关重要,但现行预报模型的准确度有待进一步提升。为此,提出一种新的预报模型BTPN。该模型引入双向Transformer,从时空序列的正逆方向提取特征,捕获关键信息,减少时空特征丢失;使用卷积Transformer模块结合卷积的局部编码和Transformer的全局编码特性,强化时空信息提取和关联性,缓解时空长时序信息丢失问题;结合细节提取模块,有助于减少局部细节的损失,并缓解高值区域消散问题。在HKO-7数据集上的评估显示,BTPN模型在MAE、SSIM及CSI指标上超越了其他先进模型,并在大面积降水和台风极端天气情境中显示出优异的预测能力。实验表明BTPN模型具备更高的预报精确性,具备较好的应用前景。