Thesis (Ph. D.)--Florida State University, 2003. / Advisor: Dr. Eric A. Smith, Florida State University, College of Arts and Sciences, Dept. of Meteorology. Title and description from dissertation home page (viewed Mar. 02, 2003). Includes bibliographical references.

Thesis (M.S.)--Naval Postgraduate School, 2002. / Thesis advisor(s): Kenneth L. Davidson, John McCarthy. Includes bibliographical references (p. 61-62). Also available online.

[pt] O presente trabalho investiga o uso de métodos de Machine Learning (ML) para efetuar estimativas para o trimestre corrente (nowcasts) da taxa de crescimento do PIB Real dos EUA. Esses métodos conseguem lidar com um grande volume de dados e séries com calendários de publicação dessincronizados, e os nowcasts são atualizados cada vez que novos dados são publicados ao longo do trimestre. Um exercício pseudo-out-of-sample é proposto para avaliar a performance de previsão e analisar o padrão de seleção de variável desses modelos. O método de ML que merece o maior destaque é o Target Factor, que supera o usualmente adotado DFM para alguns vintages dentro do trimestre. Ademais, as variáveis selecionadas apresentam consistência entre os modelos e com a intuição. / [en] This paper examines the use of Machine Learning (ML) models to compute estimates of current-quarter US Real GDP growth rate (nowcasts). These methods can handle large data sets with unsynchronized release dates, and nowcasts are updated each time new data are released along the quarter. A pseudo-out-of-sample exercise is proposed to assess forecasting performance and to analyze the variable selection pattern of these models. The ML method that deserves more attention is the Target Factor, which overcomes the usually adopted dynamic factor model for some predictions vintages in the quarter. We also analyze the variables selected, which are consistent between models and intuition.

Thunderstorms are both dangerous as well as important rain-bearing structures for large parts of the world. The prediction of thunderstorm trajectories is however difficult, especially in tropical regions. This is largely due to their smaller size and shorter lifespan. To overcome this issue, this thesis investigates how well a neural network composed of long short-term memory (LSTM) units can predict the trajectories of thunderstorms, based on several years of lightning strike data. The data is first clustered, and important features are extracted from it. These are used to predict the mean position of the thunderstorms using an LSTM network. A random search is then carried out to identify optimal parameters for the LSTM model. It is shown that the trajectories predicted by the LSTM are much closer to the true trajectories than what a linear model predicts. This is especially true for predictions of more than 1 hour. Scores commonly used to measure forecast accuracy are applied to compare the LSTM and linear model. It is found that the LSTM significantly improves forecast accuracy compared to the linear model. / Åskväder är både farliga och livsviktiga bärare av vatten för stora delar av världen. Det är dock svårt att förutsäga åskcellernas banor, främst i tropiska områden. Detta beror till större delen på deras mindre storlek och kortare livslängd. Detta examensarbete undersöker hur väl ett neuralt nätverk, bestående av long short-term memory-lager (LSTM) kan förutsäga åskväders banor baserat på flera års blixtnedlslagsdata. Först klustras datan, och viktiga karaktärsdrag hämtas ut från den. Dessa används för att förutspå åskvädrens genomsnittliga position med hjälp av ett LSTMnätverk. En slumpmässig sökning genomförs sedan för att identifiera optimala parametrar för LSTM-modellen. Det fastslås att de banor som förutspås av LSTM-modellen är mycket närmare de sanna banorna, än de som förutspås av en linjär modell. Detta gäller i synnerhet för förutsägelser mer än 1 timme framåt. Värden som är vanliga för att bedöma prognosers träffsäkerhet beräknas för att jämföra LSTM-modellen och den linjära. Det visas att LSTM-modellen klart förbättrar förutsägelsernas träffsäkerhet jämfört med den linjära modellen.

Floods cannot be prevented, but their devastating effects can be minimized if advance warning of the event is available. The South African Disaster Management Act (Act 57 of 2002) advocates a paradigm shift from the current "bucket and blanket brigade" response-based mind set to one where disaster prevention or mitigation are the preferred options. It is in the context of mitigating the effects of floods that the development and implementation of a reli able flood forecasting system has major significance. In the case of flash floods, a few hours lead time can afford disaster managers the opportunity to take steps which may significantly reduce loss of life and damage to property. The engineering challenges in developing and implementing such a system are numerous. In this thesis, the design and implement at ion of a flash flood forecasting system in South Africa is critically examined. The technical aspect s relating to spatio-temporal rainfall estimation and now casting are a key area in which new contributions are made. In particular, field and optical flow advection algorithms are adapted and refined to help pred ict future path s of storms; fast and pragmatic algorithms for combining rain gauge and remote sensing (rada r and satellite) estimates are re fi ned and validated; a two-dimensional adaptation of Empirical Mode Decomposition is devised to extract the temporally persistent structure embedded in rainfall fields. A second area of significant contribution relates to real-time fore cast updates, made in response to the most recent observed information. A number of techniques embedded in the rich Kalm an and adaptive filtering literature are adopted for this purpose. The work captures the current "state of play" in the South African context and hopes to provide a blueprint for future development of an essential tool for disaster management. There are a number of natural spin-offs from this work for related field s in water resources management. / Thesis (Ph.D.Eng.)-University of KwaZulu-Natal, Durban, 2007.

Submitted by Guilherme Branco Gomes (guilherme.branco.gomes@gmail.com) on 2018-04-17T00:19:25Z No. of bitstreams: 1 dissertacao Guilherme Branco Gomes versao final.pdf: 2137139 bytes, checksum: cead1d1fa55323ea0f81e275c713796e (MD5) / Approved for entry into archive by GILSON ROCHA MIRANDA (gilson.miranda@fgv.br) on 2018-04-18T19:53:58Z (GMT) No. of bitstreams: 1 dissertacao Guilherme Branco Gomes versao final.pdf: 2137139 bytes, checksum: cead1d1fa55323ea0f81e275c713796e (MD5) / Made available in DSpace on 2018-05-08T17:43:40Z (GMT). No. of bitstreams: 1 dissertacao Guilherme Branco Gomes versao final.pdf: 2137139 bytes, checksum: cead1d1fa55323ea0f81e275c713796e (MD5) Previous issue date: 2018-03-19 / This work compares dynamic factor model’s forecasts for Brazilian GDP. Our approach takes into account mixed frequencies and can handle missing data. We implement three models: the first is based on the Principal Components Analysis methodology; the second employs a two-step estimation method with quarterly inputs; the last is similar to the former but uses monthly series. A real-time out-of-sample exercise is proposed to assess the performance of these models. A dataset is created for each day within 27 quarters - from the fourth quarter of 2010 up to the second quarter of 2017. For recent periods, the nowcasts estimated by both two-step procedures perform better than the average predictions of Focus Survey, a bulletin organized by the Brazilian Central Bank. We also show evidence that the average of GDP forecasts from this survey may be biased / Esse trabalho compara previsões para o PIB brasileiro utilizando modelos de fatores dinâmicos. Nossa abordagem leva em consideração frequências mistas e lida com dados incompletos na base (missing data). Nós implementamos três modelos: o primeiro é baseado na metodologia de componentes principais; o segundo emprega uma estimação por dois estágio com variáveis trimestrais; o último é similar ao anterior mas utiliza series mensais. Um exercício em tempo real, fora da amostra, é proposto para comparar o desempenho desses modelos. Uma base de dados é criada para cada dia dentro de 27 trimestres - do quarto trimestre de 2010 até o segundo de 2017. Para períodos recentes, os nowcasts estimados para ambos os procedimentos de dois estágios se mostram melhores do que a média de previsão da pesquisa Focus, um boletim organizado pelo Banco Central do Brasil. Nós também mostramos evidências que a média das previsões do PIB dessa pesquisa pode ser viesada

Current precipitation nowcasting systems primarily use the extrapolation of observed radar reflectivity. I used the extrapolation and studied limits of the forecast using the concept of the decorrelation time (DCT). I used data from two radars covering the territory of the Czech Republic from warm parts of four years and calculated DCT in dependence on several selected conditions describing the state of the atmosphere. I found that the mean DCT for the extrapolation is 45.4 minutes. On average the increase of the DCT in comparison when the persistence forecast is employed is 13.4 minutes. However, in dependence on current conditions the DCT may increase or decrease in more than 40 %. I also explored time evolution of the DCT during two storm events. I found that the DCT may significantly change in time, which is the consequence of changing character of the atmosphere during the storm development.

Solar photovoltaic (PV) resources are a key enabling technology in the global energy transition towards a more sustainable future. However, PV generation is highly variable due to the dynamic shading caused by clouds. To mitigate the effects of PV variability on electrical grid stability, grid operators rely on solar forecasts to proactively dispatch grid assets and balance supply and demand. To gain insights into the nature of solar variability, which is key for effective solar forecasting, this thesis presents a statistical assessment of high resolution spectral and broadband solar irradiance in Ottawa, Canada. The statistical assessment investigates the first- and second-order spectral and temporal dependencies of irradiance time series within the context of stationarity. The temporal structures indicate that solar irradiance processes are at best weakly stationary, and the implications for forecasting are discussed. The results of the statistical assessment are leveraged to develop several deterministic machine learning solar forecasting models (LSTM, XGBoost, and 1D-CNN). These models are implemented and compared in terms of computational complexity and prediction accuracy. It was found that under all sky conditions, the inclusion of spectral irradiance data improved forecasting performance compared to only using broadband irradiance. A ramp regime classification algorithm is then described, which enables the training and testing specialized ramp regime forecasting sub-models. These specialized sub-models were found to yield even greater forecasting accuracy within their respective ramp regimes, compared with the all-sky models. Further optimization and ensembling of the presented solar forecasting models is recommended for future work.

[pt] Em economia, Nowcasting é a previsão do presente, do passado recente ou mesmo a previsão do futuro muito próximo de um determinado indicador. Geralmente, um modelo nowcast é útil quando o valor de uma variável de interesse é disponibilizado com um atraso significativo em relação ao seu período de referência e/ou sua realização inicial é notavelmente revisada ao longo do tempo, se estabilizando somente após um tempo. Nesta tese, desenvolvemos e analisamos vários métodos de Nowcasting usando dados de alta dimensão (big data) em diferentes contextos: desde a previsão de séries econômicas até o nowcast de óbitos pela COVID-19. Em um de nossos estudos, comparamos o desempenho de diferentes algoritmos de Machine Learning com modelos mais naive na previsão de muitas variáveis econômicas em tempo real e mostramos que, na maioria das vezes, o Machine Learning supera os modelos de benchmark. Já no restante dos nossos exercícios, combinamos várias técnicas de nowcasting com um grande conjunto de dados (incluindo variáveis de alta frequência, como o Google Trends) para rastrear a pandemia no Brasil, mostrando que fomos capazes de antecipar os números reais de mortes e casos muito antes de estarem disponíveis oficialmente para todos. / [en] Nowcasting in economics is the prediction of the present, the recent past or even the prediction of the very near future of a certain indicator. Generally, a nowcast model is useful when the value of a target variable is released with a significant delay with respect to its reference period and/or when its value gets notably revised over time and stabilizes only after a while. In this thesis, we develop and analyze several Nowcasting methods using high-dimensional (big) data in different contexts: from the forecasting of economic series to the nowcast of COVID-19. In one of our studies, we compare the performance of different Machine Learning algorithms with more naive models in predicting many economic variables in real-time and we show that, most of the time, Machine Learning beats benchmark models. Then, in the rest of our exercises, we combine several nowcasting techniques with a big dataset (including high-frequency variables, such as Google Trends) in order to track the pandemic in Brazil, showing that we were able to nowcast the true numbers of deaths and cases way before they got available to everyone.

Resumo: Para determinar como se inicia a convecção na região sudoeste da Amazônia, foram analisados dados do TRMM/LBA (Tropical Rainfall Measuring Mission / Large-scale Biosphere Atmosphere). A base para determinar onde e quando a convecção iniciou foi o radar banda-S, com polarização dual (S-Pol), do National Center for Atmospheric Research (NCAR). Utilizaram-se, adicionalmente, dados do canal visível do satélite GOES-8 para identificar piscinas frias produzidas pela precipitação convectiva. Essas informações, em conjunto com dados topográficos de alta resolução, foram utilizadas na determinação dos mecanismos possíveis de disparos da convecção. A elevação do terreno na área de estudo varia de 100 a 600m. Este estudo apresenta os resultados de 5 de fevereiro de 1999. Um total de 315 tempestades iniciou-se dentro do raio de 130km do radar S-Pol. Nesse dia, classificado como de fraco regime de monção, a convecção desenvolveu-se em resposta ao ciclo diurno do aquecimento solar. Cúmulos rasos espalhados durante a manhã desenvolveram-se em convecção profunda no início da tarde. As tempestades tiveram início após as 11h, com um pico de iniciação entre 15 e 16h. As causas de início de tempestades foram classificadas em 4 categorias. O modo mais comum de iniciação foi o levantamento forçado por frente de rajada (36%). A categoria, que inclui forçantes topográficas (>300m), sem a influência de nenhum outro mecanismo, é responsável por 21% das iniciações e a colisão de frentes de rajada por 16%. Nos 27% restantes, não foi possível a identificação de nenhum mecanismo. O exame de todos os dias do experimento TRMM/LBA mostrou que o dia estudado em detalhe foi representativo de muitos outros dias. Um modelo conceitual para o início e a evolução de tempestades é apresentado. Esses resultados, que devem ter implicações para outros... (Resumo completo, clicar acesso eletrônico abaixo) / Abstract : Radar and satellite data from the Tropical Rainfall Measuring Mission / Large-scale Biosphere Atmosphere (TRMM/LBA) project have been examined to determine causes for convective storm initiation in the southwest Amazon region. The locations and times of storm initiation were based on the National Center for Atmospheric Research (NCAR) S-band dual-polarization Doppler radar (S-Pol). Both the radar and GOES-8 visible data were used to identify cold pools produced by convective precipitation. This data along with high-resolution topographic data were used to determine possible convective storm triggering mechanisms. The terrain elevation varied from 100 - 600 m. Tropical forests cover the area with numerous clear cut areas used for cattle grazing and farming. This study presents the results from 5 February 1999. A total of 315 storms initiated within 130 km of the S-Pol radar. This day was classified as a weak monsoon regime where convection developed in response to the diurnal cycle of solar heating. Scattered shallow cumulus during the morning developed into deep convection by early afternoon. Storm initiation began about 1100 LST and peaked around 1500-1600 LST. The causes of storm initiation were classified into 4 categories. The most common initiation mechanism was caused by forced lifting by a gust front (36%). Forcing by terrain (>300 m) without any other triggering mechanism accounted for 21% of the initiations and colliding gust fronts 16%. For the remaining 27% a triggering mechanism was not identified. Examination of all days during TRMM/LBA showed that this one detailed study day was representative of many days. A conceptual model of storm initiation and evolution is presented. The results of this study should have implications for other locations when synoptic scale forcing mechanisms are at a minimum... (Complete abstract click electronic access below) / Orientador: João Francisco Escobedo / Coorientador: Maria Assunção Faus da Silva Dias / Banca: Nelson de Jesus Ferreira / Banca: Roberto Vicente Calheiros / Banca: Oswaldo Massambani / Banca: Jonas Teixeira Nery / Doutor