Rozprawy doktorskie na temat „Hydrologic Ensemble Prediction Systems”

Nous abordons la prévision probabiliste des débits à partir de deux perspectives basées sur la complémentarité de multiples modèles hydrologiques (diversité). La première exploite une méthodologie hybride basée sur l’évaluation de plusieurs modèles hydrologiques globaux et d’outils d’apprentissage automatique pour la sélection optimale des prédicteurs, alors que la seconde fait recourt à la construction d’ensembles de réseaux de neurones en forçant la diversité. Cette thèse repose sur le concept de la diversité pour développer des méthodologies différentes autour de deux problèmes pouvant être considérés comme complémentaires. La première approche a pour objet la simplification d’un système complexe de prévisions hydrologiques d’ensemble (dont l’acronyme anglais est HEPS) qui dispose de 800 scénarios quotidiens, correspondant à la combinaison d’un modèle de 50 prédictions météorologiques probabilistes et de 16 modèles hydrologiques globaux. Pour la simplification, nous avons exploré quatre techniques: la Linear Correlation Elimination, la Mutual Information, la Backward Greedy Selection et le Nondominated Sorting Genetic Algorithm II (NSGA-II). Nous avons plus particulièrement développé la notion de participation optimale des modèles hydrologiques qui nous renseigne sur le nombre de membres météorologiques représentatifs à utiliser pour chacun des modèles hydrologiques. La seconde approche consiste principalement en la sélection stratifiée des données qui sont à la base de l’élaboration d’un ensemble de réseaux de neurones qui agissent comme autant de prédicteurs. Ainsi, chacun d’entre eux est entraîné avec des entrées tirées de l’application d’une sélection de variables pour différents échantillons stratifiés. Pour cela, nous utilisons la base de données du deuxième et troisième ateliers du projet international MOdel Parameter Estimation eXperiment (MOPEX). En résumé, nous démontrons par ces deux approches que la diversité implicite est efficace dans la configuration d’un HEPS de haute performance.
In this thesis, we tackle the problem of streamflow probabilistic forecasting from two different perspectives based on multiple hydrological models collaboration (diversity). The first one favours a hybrid approach for the evaluation of multiple global hydrological models and tools of machine learning for predictors selection, while the second one constructs Artificial Neural Network (ANN) ensembles, forcing diversity within. This thesis is based on the concept of diversity for developing different methodologies around two complementary problems. The first one focused on simplifying, via members selection, a complex Hydrological Ensemble Prediction System (HEPS) that has 800 daily forecast scenarios originating from the combination of 50 meteorological precipitation members and 16 global hydrological models. We explore in depth four techniques: Linear Correlation Elimination, Mutual Information, Backward Greedy Selection, and Nondominated Sorting Genetic Algorithm II (NSGA-II). We propose the optimal hydrological model participation concept that identifies the number of meteorological representative members to propagate into each hydrological model in the simplified HEPS scheme. The second problem consists in the stratified selection of data patterns that are used for training an ANN ensemble or stack. For instance, taken from the database of the second and third MOdel Parameter Estimation eXperiment (MOPEX) workshops, we promoted an ANN prediction stack in which each predictor is trained on input spaces defined by the Input Variable Selection application on different stratified sub-samples. In summary, we demonstrated that implicit diversity in the configuration of a HEPS is efficient in the search for a HEPS of high performance.

La prévision hydrologique consiste à évaluer quelle sera l'évolution du débit au cours des prochains pas de temps. En utilisant les systèmes actuels de prévisions hydrologiques déterministes, il est impossible d'apprécier simplement l'incertitude associée à ce type de prévision, ce que peut nuire à la prise de décisions. La prévision hydrologique d'ensemble (PHE) cherche à étayer cette incertitude en proposant, à chaque pas de temps, une distribution de probabilité, la prévision probabiliste, en place et lieu d'une estimation unique du débit, la prévision déterministe. La PHE offre de nombreux bénéfices : elle informe l'utilisateur de l'incertitude; elle permet aux autorités qui prennent des décisions de déterminer des critères d'alerte et de mettre en place des scénarios d'urgence; elle fournit les informations nécessaires à la prise de décisions tenant compte du risque. L'objectif principal de cette thèse est l'évaluation de prévisions hydrologiques d'ensemble, en mettant l'accent sur la performance et la fiabilité de celles-ci. Deux techniques pour construire des ensembles sont explorées: a) une première reposant sur des prévisions météorologiques d'ensemble (PME) et b) une seconde exploitant simultanément un ensemble de modèles hydrologiques (multimodèle). En termes généraux, les objectifs de la thèse ont été établis afin d'évaluer : a) les incertitudes associées à la structure du modèle : une étude qui repose sur des simulations journalières issues de dix-sept modèles hydrologiques globaux, pour plus de mille bassins versants français; b) les incertitudes associées à la prévision météorologique : une étude qui exploite la PME du Service Météorologique du Canada et un modèle hydrologique opérationnel semi-distribué, pour un horizon de 3 jours sur douze bassins versants québécois; c) les incertitudes associées à la fois à la structure du modèle et à la prévision météorologique : une étude qui repose à la fois sur la PME issue du ECMWF (European Centre for Medium-Range Weather Forecasts) et seize modèles hydrologiques globaux, pour un horizon de 9 jours sur 29 bassins versants français. Les résultats mets en évidence les avantages des systèmes probabilistes par rapport aux les déterministes. Les prévisions probabilistes sont toutefois souvent affectées par une sous dispersion de leur distribution prédictive. Elles exigent alors un post traitement avant d'être intégrées dans un processus de prise de décision. Plus intéressant encore, les résultats ont également montré le grand potentiel de combiner plusieurs sources d'incertitude, notamment celle associée à la prévision météorologique et celle associée à la structure des modèles hydrologiques. Il nous semble donc prioritaire de continuer à explorer davantage cette approche combinatoire.

Les simulations et prévisions hydrologiques sont sujettes à diverses sources d'incertitudes, qui sont malheureusement inévitables. La cascade d'incertitude provient de différentes composantes de la chaîne de prévision, telles que la nature chaotique de l'atmosphère, diverses conditions initiales et limites, une modélisation hydrologique conceptuelle nécessairement inexacte et des paramètres stationnaires incohérents avec un environnement en mutation. La prévision d'ensemble s'avère un outil puissant pour représenter la croissance des erreurs dans le système dynamique et pour capter les incertitudes associées aux différentes sources. Thiboult et al. (2016) ont construit un grand ensemble de 50,000 membres qui tient compte de l'incertitude des prévisions météorologiques, de celle des conditions initiales et l’incertitude structurale. Ce vaste ensemble de 50,000 membres peut également être séparé en sous-composants afin de démêler les trois principales sources d’incertitude mentionnées ci-dessus. Emixi Valdez a donc généré un autre H-EPS multimodèles et calibré pour différents bassins hydrographiques suivant un modèle similaire. Cependant, les résultats obtenus ont été simplement agrégés, en considérant les membres équiprobables. Bien que les systèmes de prévision hydrologique multimodèles puissent être considérés comme un système très complet, ils sont néanmoins exposés à d'autres incertitudes. Par exemple, les prévisions météorologiques des recherches de Thiboult et al. (2016) ont été pré-testées sur certains bassins versants. Ces tests ont montré que les performances dues à l'assimilation de données s'estompent rapidement avec l’horizon de prévision. De plus, en réalité, les utilisateurs peuvent ne pas être en mesure d’utiliser parfaitement tous les outils de prévision (c’est-à-dire les prévisions météorologiques d’ensemble, l’assimilation de données et le schéma multimodèle) conjointement. Par conséquent, il existe toujours une place pour l'amélioration permettant d'augmenter la fiabilité et la résolution des prévisions grâce à un post-traitement statistique approprié. L'objectif global de cette recherche est d'explorer l'utilisation appropriée et les compétences prévisionnelles de divers algorithmes statistiques afin de post-traiter séparément les prévisions de débit provenant d’un modèle unique ainsi que les prévisions multimodèles. Premièrement, nous avons testé l’efficacité de méthodes depost-traitement telles que le Affine Kernel Dressing (AKD) et le Non-dominated sorting genetic algorithm II (NSGA-II) en comparant les prévisions post-traitées par ces méthodes aux soties brutes de systèmes de prévision à modèle unique. Ces deux méthodes sont théoriquement / techniquement distinctes, mais partagent toutefois la même caractéristique, à savoir qu’elles ne nécessitent pas d’hypothèse paramétrique concernant la distribution des membres de la prévision d’ensemble. Elles peuvent donc être considérées comme des méthodes de post-traitement non paramétriques. Dans cette étude, l'analyse des fronts de Pareto générés avec NSGA-II a démontré la supériorité de l'ensemble post-traité en éliminant efficacement les biais des prévisions et en maintenant une bonne dispersion pour tous les horizons de prévision. Deux autres méthodes de post-traitement, à savoir le Bayesian Model Averaging (BMA) et le Copula-BMA, ont également été comparées. Ces deux méthodes ont permis d’obtenir des distributions prédictives à partir de prévisions de débit journalier émises par cinq systèmes de prévision d'ensemble hydrologiques différents. Les poids obtenus par la méthode du BMA quantifient le niveau de confiance que l'on peut avoir à l'égard de chaque modèle hydrologique candidat et conduisent à une fonction de densité prédictive (PDF) contenant des informations sur l'incertitude. Le BMA améliore la qualité globale des prévisions, principalement en maintenant la dispersion de l'ensemble avec l’horizon de prévision. Il a également la capacité d’améliorer la fiabilité des systèmes multimodèles qui n’incluent que deux sources d’incertitudes. Le BMA est donc efficace pour améliorer la fiabilité et la résolution des prévisions hydrologiques. Toutefois, le BMA souffre de limitations dues au fait que les fonctions de densité de probabilité conditionnelle (PDF) doivent suivre une distribution paramétrique connue (ex., normale, gamma). Par contre, le modèle prédictif Copula-BMA ne requiert pas une telle hypothèse et élimine aussi l'étape de transformation de puissance, qui est nécessaire pour le BMA. Dans cette étude, onze types de distributions marginales univariées et six fonctions de copule de différents niveaux de complexité ont été explorés dans un cadre Copula-BMA. Cela a permis de représenter de manière exhaustive la structure de dépendance entre des couples de débits prévus et observés. Les résultats démontrent la supériorité du Copula-BMA par rapport au BMA pour réduire le biais dans les prévisions et maintenir une dispersion appropriée pour tous les horizons de prévision.
Hydrological simulations and forecasts are subject to various sources of uncertainties. Forecast uncertainties are unfortunately inevitable when conducting the deterministic analysis of a dynamical system. The cascade of uncertainty originates from different components of the forecasting chain, such as the chaotic nature of the atmosphere, various initial conditions and boundaries, necessarily imperfect hydrologic modeling, and the inconsistent stationnarity assumption in a changing environment. Ensemble forecasting is a powerful tool to represent error growth in the dynamical system and to capture the uncertainties associated with different sources. Thiboult et al. (2016) constructed a 50,000-member great ensemble that accounts for meteorological forcing uncertainty, initial conditions uncertainty, and structural uncertainty. This large ensemble can also be separated into sub-components to untangle the three main sources of uncertainties mentioned above. In asimilar experiment, another multimodel hydrological ensemble forecasting system implemented for different catchments was produced by Emixi Valdez. However,in the latter case, model outputs were simply pooled together, considering the members equiprobable. Although multimodel hydrological ensemble forecasting systems can be considered very comprehensive, they can still underestimate the total uncertainty. For instance, the meteorological forecasts in there search of Thiboult et al. (2016) were pre-tested on some watersheds. It was found out that the forecasting performance of data assimilation fades away quickly as the lead time progresses. In addition, operational forecasts users may not able to perfectly utilize all the forecasting tools (i.e., meteorological ensemble forcing, data assimilation, and multimodel) jointly. Therefore, there is still room for improvement to enhance the forecasting skill of such systems through proper statistical post-processing.The global objective of this research is to explore the proper use and predictive skill of various statistical post-processing algorithms by testing them on single-model and multimodel ensemble stream flow forecasts. First, we tested the post-processing skills of Affine kernel dressing (AKD) and Non-dominated sorting genetic algorithm II (NSGA-II) over single-model H-EPSs. Those two methods are theoretically/technically distinct yet are both non-parametric. They do not require the raw ensemble members to follow a specific parametric distribution.AKD-transformed ensembles and the Pareto fronts generated with NSGA-II demonstrated the superiority of post-processed ensembles compared to raw ensembles. Both methods where efficient at eliminating biases and maintaining a proper dispersion for all forecasting horizons. For multimodel ensembles, two post-processors, namely Bayesian model averaging (BMA) and the integrated copula-BMA, are compared for deriving a pertinent joint predictive distribution of daily streamflow forecasts issued by five different single-model hydrological ensemble prediction systems (H-EPSs). BMA assign weights to different models. Forecasts from all models are then combined to generate more skillful and reliable probabilistic forecasts. BMA weights quantify the level of confidence one can have regarding each candidate hydrological model and lead to a predictive probabilistic density function (PDF) containing information about uncertainty. BMA improves the overall quality of forecasts mainly by maintaining the ensemble dispersion with the lead time. It also improves the reliability and skill of multimodel systems that only include two sources of uncertainties compared to the 50,000-member great ensemble from Thiboult et al (2016). Furthermore, Thiboult et al. (2016) showed that the meteorological forecasts they used were biased and unreliable on some catchments. BMA improves the accuracy and reliability of the hydrological forecasts in that case as well.However, BMA suffers from limitations pertaining to its conditional probability density functions (PDFs), which must follow a known parametric distribution form (e.g., normal, gamma). On the contrary, Copula-BMA predictive model fully relaxes this constraint and also eliminates the power transformation step. In this study, eleven univariate marginal distributions and six copula functions are explored in a Copula-BMA framework for comprehensively reflecting the dependence structure between pairs of forecasted and observed streamflow. Results demonstrate the superiority of the Copula-BMAcompared to BMA in eliminating biases and maintaining an appropriate ensemble dispersion for all lead-times.

Artificial Neural Networks (ANNs) have been comprehensively researched, both from a computer scientific perspective and with regard to their use for predictive modelling in a wide variety of applications including hydrology and the environment. Yet their adoption for live, real-time systems remains on the whole sporadic and experimental. A plausible hypothesis is that this may be at least in part due to their treatment heretofore as “black boxes” that implicitly contain something that is unknown, or even unknowable. It is understandable that many of those responsible for delivering Early Warning Systems (EWS) might not wish to take the risk of implementing solutions perceived as containing unknown elements, despite the computational advantages that ANNs offer. This thesis therefore builds on existing efforts to open the box and develop tools and techniques that visualise, analyse and use ANN weights and biases especially from the viewpoint of neural pathways from inputs to outputs of feedforward networks. In so doing, it aims to demonstrate novel approaches to self-improving predictive model construction for both regression and classification problems. This includes Neural Pathway Strength Feature Selection (NPSFS), which uses ensembles of ANNs trained on differing subsets of data and analysis of the learnt weights to infer degrees of relevance of the input features and so build simplified models with reduced input feature sets. Case studies are carried out for prediction of flooding at multiple nodes in urban drainage networks located in three urban catchments in the UK, which demonstrate rapid, accurate prediction of flooding both for regression and classification. Predictive skill is shown to reduce beyond the time of concentration of each sewer node, when actual rainfall is used as input to the models. Further case studies model and predict statutory bacteria count exceedances for bathing water quality compliance at 5 beaches in Southwest England. An illustrative case study using a forest fires dataset from the UCI machine learning repository is also included. Results from these model ensembles generally exhibit improved performance, when compared with single ANN models. Also ensembles with reduced input feature sets, using NPSFS, demonstrate as good or improved performance when compared with the full feature set models. Conclusions are drawn about a new set of tools and techniques, including NPSFS and visualisation techniques for inspection of ANN weights, the adoption of which it is hoped may lead to improved confidence in the use of ANN for live real-time EWS applications.

Thesis (M.S. in Meteorology)--Naval Postgraduate School, March 2006.
Thesis Advisor(s): Carlyle H. Wash, Patrick A. Harr. "March 2006." Includes bibliographical references (p. 129). Also available online.

Approved for public release; distribution is unlimited.
The skill of a combined grand ensemble (GE), which is constructed from three operational global ensemble prediction systems (EPS), is evaluated with respect to the probability forecast of a tropical cyclone (TC) being within a specified area. Anisotropic probability ellipses are defined from the GE to contain 68% of the ensemble members. Forecast reliability is based on whether the forecast verifying position is within the ellipse. A sharpness parameter is based on the size of the GE-based probability ellipse relative to other operational forecast probability ellipses. For the 2010 Atlantic TC season, results indicate that the GE ellipses exhibit a high degree of reliability whereas the operational probability circle tends to be over-dispersive. Additionally, the GE ellipse tends to be sharper than the operational product for forecast intervals beyond 48 hours. The size and shape of the GE ellipses varied with TC track types, which suggests that information about the physics of the flow-dependent system is retained whereas isotropic probability ellipses may not reflect variability associated with track type. It is concluded that the GE probability ellipse demonstrates utility for combined EPS to enhance probabilistic forecasts for use as TC-related decision aids, as there is a potential for reducing the sizes of warning areas.

Classification is a data mining problem that arises in many real-world applications. A popular approach to tackle these classification problems is using an ensemble of classifiers that combines the collective knowledge of several classifiers. Most popular methods create a static ensemble, in which a single ensemble is constructed or chosen from a pool of classifiers and used for all new data instances. Two factors that have been frequently used to construct a static ensemble are the accuracy of and diversity among the individual classifiers. There have been many studies investigating how these factors should be combined and how much diversity is required to increase the ensemble's performance. These results have concluded that it is not trivial to build a static ensemble that generalizes well. Recently, a different approach has been undertaken: dynamic ensemble construction. Using a different set of classifiers for each new data instance rather than a single static ensemble of classifiers may increase performance since the dynamic ensemble is not required to generalize across the feature space. Most studies on dynamic ensembles focus on classifiers' competency in the local region in which a new data instance resides or agreement among the classifiers. In this thesis, we propose several other approaches for dynamic class prediction. Existing methods focus on assigned labels or their correctness. We hypothesize that using the class probability estimates returned by the classifiers can enhance our estimate of the competency of classifiers on the prediction. We focus on how to use class prediction probabilities (confidence) along with accuracy and diversity to create dynamic ensembles and analyze the contribution of confidence to the system. Our results show that confidence is a significant factor in the dynamic setting. However, it is still unclear how accurate, diverse, and confident ensemble can best be formed to increase the prediction capability of the system. Second, we propose a system for dynamic ensemble classification based on a new distance measure to evaluate the distance between data instances. We first map data instances into a space defined by the class probability estimates from a pool of two-class classifiers. We dynamically select classifiers (features) and the k-nearest neighbors of a new instance by minimizing the distance between the neighbors and the new instance in a two-step framework. Results of our experiments show that our measure is effective for finding similar instances and our framework helps making more accurate predictions. Classifiers' agreement in the region where a new data instance resides has been considered a major factor in dynamic ensembles. We postulate that the classifiers chosen for a dynamic ensemble should behave similarly in the region in which the new instance resides, but differently outside of this area. In other words, we hypothesize that high local accuracy, combined with high diversity in other regions, is desirable. To verify the validity of this hypothesis we propose two approaches. The first approach focuses on finding the k-nearest data instances to the new instance, which then defines a neighborhood, and maximizes simultaneously local accuracy and distant diversity, based on data instances outside of the neighborhood. The second method considers all data instances to be in the neighborhood, and assigns them weights depending on the distance to the new instance. We demonstrate through several experiments that weighted distant diversity and weighted local accuracy outperform all benchmark methods.

The aim of the thesis is the assessment of precipitation input uncertainty into hydrological response in hydrometeorological ensemble systems for flood prediction. The study has been preliminary focused on the development of a hydrometeorological system that couples a statistical precipitation downscaling model, known as STRAIN, with a fully-distributed hydrological model, known as tRIBS. In a subsequent part of the research, a rigorous method has been designed to test the consistency hypothesis (i.e. ensemble and observations are drawn from the same distribution) of the ensemble precipitation fields generated by downscaling models. The verification procedure has been tested by means of numerical experiments. Results permit us to conclude that: (i) ensemble members generated using model parameters estimated on the observed event are overdispersed; (ii) the adoption of a single calibration relation linking model parameters and coarse meteorological observable can lead to the generation of consistent ensemble members; (iii) when a single calibration relation is not able to explain observed events variability, storm-specific calibration relation should be adopted to return consistent forecasts. Finally, in the last part of the work, a rigorous method has been developed to assess consistency of ensemble streamflows produced by hydrometeorological systems. The method has been tested with numerical experiments using the prediction system designed in the preliminary phase of the study with the purpose of evaluating the propagation of uncertainty of downscaled precipitation input into hydrological response. The innovative aspects of the thesis rely on (i) the development of rigorous verification methods for ensemble outputs of hydrometeorological systems; and (ii) the application of these procedure on a great number of events in order to draw statistically significant conclusions.

The aim of this work is to assess the added value of the enhanced horizontal resolution in the probabilistic prediction of upper-level and surface fields. In particular, the performance of three different ensemble prediction systems were compared: ECMWF-ENS (51 members, 18 km horizontal resolution), COSMO-LEPS (16 members, 7 km horizontal resolution) and COSMO-2I-EPS (10 members, 2.2 km horizontal resolution). While the first 2 ensemble systems are operational, COSMO-2I-EPS is still in a development phase. Therefore, the intercomparison window covers a limited period, which ranges from 20 to 27 June 2016. In this work, both upper-level and surface variables are analyzed. As for upper-level, both temperature and the geopotential height at three different pressure levels are considered; the ensemble spread and the root mean square error are computed using the available Italian radiosounding data every 12/24 hours for verification. As for the surface, 2-metre temperature and precipitation cumulated over six hours are verified against the non-conventional station network provided by the National Civil Protection Department. The ensemble spread and the root mean square error of 2-metre temperature are computed, while a number of probabilistic scores (Brier Skill Score, Ranked Probability Score, Roc-Area, Outliers Percentage and others) are considered for precipitation. For both upper-level and surface verification, it turns out that the best scores are mainly obtained by the COSMO-based ensemble systems with higher horizontal resolution and lower ensemble size. The added value of high resolution in mesoscale ensemble systems seems to play a crucial role in the probabilistic prediction of atmospheric fields at all levels. In particular, the more detailed description of mesoscale and orographic-related processes in COSMO-ensembles provides an added value for the prediction of localised High-Impact Weather events.

L'objectiu principal d'aquesta tesi és millorar l'actual capacitat de predicció de fenòmens meteorològics de pluja intensa potencialment perillosos a la Mediterrània occidental. Es desenvolupen i verifiquen tres sistemes de predicció per conjunts (SPC) que tenen en compte incerteses presents en els models numèrics i en les condicions inicials. Per generar els SPC s'utilitza la connexió entre les estructures de vorticitat potencial (VP) i els ciclons, a més de diferents esquemes de parametrització física. Es mostra que els SPC proporcionen una predicció més hàbil que la determinista. Els SPC generats pertorbant les condicions inicials han obtingut millor puntuació en verificacions estadístiques. Els resultats d'aquesta tesi mostren la utilitat i la idoneïtat dels mètodes de predicció basats en la pertorbació d'estructures de VP de nivells alts, precursors de les situacions ciclòniques. Els resultats i estratègies presentats pretenen ser un punt de partida per a futurs estudis que facin ús d'aquests mètodes.
The main goal of this thesis is to improve the current prediction skill of potentially hazardous heavy precipitation weather events in the western Mediterranean region. We develop and test three different ensemble prediction systems (EPSs) that account for uncertainties present in both the numerical models and the initial conditions. To generate the EPSs we take advantage of the connection between potential vorticity (PV) structures and cyclones, and use different physical parameterization schemes. We obtain an improvement in forecast skill when using an EPS compared to a determinist forecast. The EPSs generated perturbing the initial conditions perform better in the statistical verification scores. The results of this Thesis show the utility and suitability of forecasting methods based on perturbing the upper-level precursor PV structures present in cyclonic situations. The results and strategies here discussed aim to be a basis for future studies making use of these methods.

This thesis advances the understanding of the application of artificial neural networks ensemble to clinical data by addressing the following fundamental question: What is the potentiality of an ensemble of neural networks models as a filter and classifier in a complex clinical situation? A novel neural networks ensemble classification model called Rules and Information Driven by Consistency in Artificial Neural Networks Ensemble (RIDCANNE) is developed for the purpose of prediction of medical outcomes or events, such as kidney transplants. The proposed classification model is based on combination of initial data preparations, preliminary classification by ensembles of Neural Networks, and generation of new training data based on criteria of highly accuracy and model agreement. Furthermore, it can also generate decision tree classification models to provide classification of data and the prediction results. The case studies described in this thesis are from a kidney transplant database and two well-known collections of benchmark data known as the Pima Indian Diabetes and Wisconsin Cancer datasets. An implication of this study is that further attention needs to be given to both data collection and preparation stages. This study revealed that even neural network ensemble models that are known for their strong generalization ability might not be able to provide a high level of accuracy for complex, noisy and incomplete clinical data. However, by using a selective subset of data points, it is possible to improve the overall accuracy. In summary, the research conducted for this thesis advances the current clinical data preparation and classification techniques in which the task is to extract patterns that contain higher information content from a sea of noisy and incomplete clinical data, and build accurate and transparent classifiers. The RIDC-ANNE approach improves an analyst�s ability to better understand the data. Furthermore, it shows great promise for use in clinical decision making systems. It can provide us with a valuable data mining tool with great research and commercial potential.

In this study, the prediction model based on the Stacking principle is called the Stacking fusion model. Little evidence demonstrates that the Stacking fusion model possesses better prediction performance in the field of heart disease diagnosis than other classification models. Since this model belongs to the family of ensemble learning models, which has a bad interpretability, it should be used with caution in medical diagnoses. The purpose of this study is to verify whether the Stacking fusion model has better prediction performance than stand-alone machine learning models and other ensemble classifiers in the field of heart disease diagnosis, and to find ways to explain this model. This study uses experiment and quantitative analysis to evaluate the prediction performance of eight models in terms of prediction ability, algorithmic stability, false negative rate and run-time. It is proved that the Stacking fusion model with Naive Bayes classifier, XGBoost and Random forest as the first-level learners is superior to other classifiers in prediction ability. The false negative rate of this model is also outstanding. Furthermore, the Stacking fusion model is explained from the working principle of the model and the SHAP framework. The SHAP framework explains this model’s judgement of the important factors that influence heart disease and the relationship between the value of these factors and the probability of disease. Overall, two research problems in this study help reveal the prediction performance and reliability of the cardiac disease prediction model based on the Stacking principle. This study provides practical and theoretical support for hospitals to use the Stacking principle in the diagnosis of heart disease.

The thesis has for objective to study new statistical methods to correct temperature predictionsthat may be implemented on the ensemble prediction system (EPS) of Meteo France so toimprove its use for the electric system management, at EDF France. The EPS of Meteo Francewe are working on contains 51 members (forecasts by time-step) and gives the temperaturepredictions for 14 days. The thesis contains three parts: in the first one we present the EPSand we implement two statistical methods improving the accuracy or the spread of the EPS andwe introduce criteria for comparing results. In the second part we introduce the extreme valuetheory and the mixture models we use to combine the model we build in the first part withmodels for fitting the distributions tails. In the third part we introduce the quantile regressionas another way of studying the tails of the distribution.

River streamflow forecasting is traditionally based on real-time measurements of rainfall over catchments and discharge at the outlet and upstream. These data are processed in mathematical models of varying complexity and allow to obtain accurate predictions for short times. In order to extend the forecast horizon to a few days - to be able to issue early warning - it is necessary to take into account the weather forecasts. However, the latter display the property of sensitivity to initial conditions, and for appropriate risk management, forecasts should therefore be considered in probabilistic terms. Currently, ensemble predictions are made using a numerical weather prediction model with perturbed initial conditions and allow to assess uncertainty.

The research began by analyzing the meteorological predictions at the medium-range (up to 10-15 days) and their use in hydrological forecasting. Precipitation from the ensemble prediction system of the European Centre for Medium-Range Weather Forecasts (ECMWF) were used. A semi-distributed hydrological model was used to transform these precipitation forecasts into ensemble streamflow predictions. The performance of these forecasts was analyzed in probabilistic terms. A simple decision model also allowed to compare the relative economic value of hydrological ensemble predictions and some deterministic alternatives.

Numerical weather prediction models are imperfect. The ensemble forecasts are therefore affected by errors implying the presence of biases and the unreliability of probabilities derived from the ensembles. By comparing the results of these predictions to the corresponding observed data, a statistical model for the correction of forecasts, known as post-processing, has been adapted and shown to improve the performance of probabilistic forecasts of precipitation. This approach is based on retrospective forecasts made by the ECMWF for the past twenty years, providing a sufficient statistical sample.

Besides the errors related to meteorological forcing, hydrological forecasts also display errors related to initial conditions and to modeling errors (errors in the structure of the hydrological model and in the parameter values). The last stage of the research was therefore to investigate, using simple models, the impact of these different sources of error on the quality of hydrological predictions and to explore the possibility of using hydrological reforecasts for post-processing, themselves based on retrospective precipitation forecasts.

/

La prévision des débits des rivières se fait traditionnellement sur la base de mesures en temps réel des précipitations sur les bassins-versant et des débits à l'exutoire et en amont. Ces données sont traitées dans des modèles mathématiques de complexité variée et permettent d'obtenir des prévisions précises pour des temps courts. Pour prolonger l'horizon de prévision à quelques jours – afin d'être en mesure d'émettre des alertes précoces – il est nécessaire de prendre en compte les prévisions météorologiques. Cependant celles-ci présentent par nature une dynamique sensible aux erreurs sur les conditions initiales et, par conséquent, pour une gestion appropriée des risques, il faut considérer les prévisions en termes probabilistes. Actuellement, les prévisions d'ensemble sont effectuées à l'aide d'un modèle numérique de prévision du temps avec des conditions initiales perturbées et permettent d'évaluer l'incertitude.

La recherche a commencé par l'analyse des prévisions météorologiques à moyen-terme (10-15 jours) et leur utilisation pour des prévisions hydrologiques. Les précipitations issues du système de prévisions d'ensemble du Centre Européen pour les Prévisions Météorologiques à Moyen-Terme ont été utilisées. Un modèle hydrologique semi-distribué a permis de traduire ces prévisions de précipitations en prévisions d'ensemble de débits. Les performances de ces prévisions ont été analysées en termes probabilistes. Un modèle de décision simple a également permis de comparer la valeur économique relative des prévisions hydrologiques d'ensemble et d'alternatives déterministes.

Les modèles numériques de prévision du temps sont imparfaits. Les prévisions d'ensemble sont donc entachées d'erreurs impliquant la présence de biais et un manque de fiabilité des probabilités déduites des ensembles. En comparant les résultats de ces prévisions aux données observées correspondantes, un modèle statistique pour la correction des prévisions, connue sous le nom de post-processing, a été adapté et a permis d'améliorer les performances des prévisions probabilistes des précipitations. Cette approche se base sur des prévisions rétrospectives effectuées par le Centre Européen sur les vingt dernières années, fournissant un échantillon statistique suffisant.

A côté des erreurs liées au forçage météorologique, les prévisions hydrologiques sont également entachées d'erreurs liées aux conditions initiales et aux erreurs de modélisation (structure du modèle hydrologique et valeur des paramètres). La dernière étape de la recherche a donc consisté à étudier, à l'aide de modèles simples, l'impact de ces différentes sources d'erreur sur la qualité des prévisions hydrologiques et à explorer la possibilité d'utiliser des prévisions hydrologiques rétrospectives pour le post-processing, elles-même basées sur les prévisions rétrospectives des précipitations.
Doctorat en Sciences
info:eu-repo/semantics/nonPublished

The thermally-driven rotating annulus is a laboratory experiment used to study the dynamics of planetary atmospheres under controlled and reproducible conditions. The predictability of this experiment is studied by applying the same principles used to predict the atmosphere. A forecasting system for the annulus is built using the analysis correction method for data assimilation and the breeding method for ensemble generation. The results show that a range of flow regimes with varying complexity can be accurately assimilated, predicted, and studied in this experiment. This framework is also intended to demonstrate a proof-of-concept: that the annulus could be used as a testbed for meteorological techniques under laboratory conditions. First, a regime diagram is created using numerical simulations in order to select points in parameter space to forecast, and a new chaotic flow regime is discovered within it. The two components of the framework are then used as standalone algorithms to measure predictability in the perfect model scenario and to demonstrate data assimilation. With a perfect model, regular flow regimes are found to be predictable until the end of the forecasts, and chaotic regimes are predictable over hundreds of seconds. There is a difference in the way predictability is lost between low-order chaotic regimes and high-order chaos. Analysis correction is shown to be accurate in both regular and chaotic regimes, with residual velocity errors about 3-8 times the observational error. Specific assimilation scenarios studied include information propagation from data-rich to data-poor areas, assimilation of vortex shedding observations, and assimilation over regime and rotation rate transitions. The full framework is used to predict regular and chaotic flow, verifying the forecasts against laboratory data. The steady wave forecasts perform well, and are predictable until the end of the available data. The amplitude and structural vacillation forecasts lose quality and skill by a combination of wave drift and wavenumber transition. Amplitude vacillation is predictable up to several hundred seconds ahead, and structural vacillation is predictable for a few hundred seconds.

Streamflow forecasts are very useful for a variety of applications such as flood warning, reservoir operation and water resources planning and management, especially in countries like India where streamflow can be highly variable. Methods available for streamflow forecasting can be broadly classified as process-driven and data-driven methods. Forecasts always have uncertainty associated with them due to limitations in modelling complex processes in the hydrologic system, and factors such as scarcity of data and measurement errors. It is important to quantify the forecast uncertainty for making informed decisions. Hydrologic Ensemble Prediction Systems (HEPS), which use ensembles in process-driven approach for generating probabilistic forecasts to quantify uncertainty are gaining popularity in the world. However, there is dearth of studies on application of HEPS for forecasting streamflows in Indian rivers. Recently, United States (US) National Weather Service developed a HEPS called Hydrologic Ensemble Forecast Service (HEFS) to generate seamless probabilistic hydrologic forecasts from short to long lead times. The first objective of this thesis is to investigate the potential of HEFS in generating skilful streamflow forecasts for an Indian river, as there is no prior application of HEFS outside US. Tel river, which is one of the tributaries of Mahanadi river (which is frequently prone to floods) was chosen for case study. Forecasts of meteorological variables (precipitation and temperature) required as input to HEFS were obtained from Global Ensemble Forecast System (GEFS). The HEFS consists of three main components - (i) Meteorological Ensemble Forecast Processor (MEFP), (ii) Hydrologic Processor and (iii) Hydrologic Ensemble Postprocessor (EnsPost). MEFP accounts for meteorological uncertainty by generating bias corrected ensemble meteorological forecast which is subsequently propagated through the Hydrologic Processor initialised with basin conditions. The resulting hydrologic ensemble forecast is input to EnsPost to generate postprocessed hydrologic ensemble forecast which reflects the total forecast uncertainty accounting for both meteorological and hydrologic uncertainties. A lumped rainfall-runoff model called GR4J was used as the Hydrologic Processor. Verification of retrospective daily streamflow forecasts generated using HEFS for Tel river against corresponding observations indicated that the forecasts have fairly good skill at short lead times (1 to 3 days). The forecasts were found to have higher skill compared to climatological forecasts and forecasts generated by an ARIMA model. Statistical methods are widely used operationally for forecasting streamflow at coarser time scales such as seasonal. For some applications (e.g., coordinated operation of a system of reservoirs), contemporaneous streamflow forecasts may be required at many sites in a basin. Forecasts generated using separate statistical models for each site may not preserve spatial correlation structure between flows at different sites. The second objective of this thesis is to explore the potential of regularised Multivariate Multiple Linear Regression (MMLR) models in generating skilful multisite streamflow forecasts. Three regularisation methods namely ridge regression, lasso and MRCE (Multivariate Regression with Covariance Estimation) were considered. The potential of the regularised MMLR models was examined through a case study on seasonal streamflow forecasting in upper Colorado river basin of US. Performance of the models was compared with that of four other multisite forecasting methods based on (i) Schaake Shuffle, (ii) Principal Component Analysis, (iii) disaggregation and (iv) k-nearest neighbour resampling, which were available in literature. Considering both forecast skill and ability to preserve inter-site correlations, the method based on MMLR and ridge regression was found to perform better than the other methods considered.

博士
國立成功大學
水利及海洋工程學系
87
The fractal geometry and the chaotic dynamics are two important components in nonlinear sciences. They are powerful tools to analyze the complex hydrologic systems currently. In this study, we focus the attention on hydrosciences. Eight topics are concerned and described as follows: (1) estimation of point-fractal dimension and its variability for hydrologic process, (2) analysis and clustering for areal fractal characteristics, (3) identification of time-scale law and fractal approach of frequency analysis for hydrologic variable, (4) existence identification of attractor for daily streamflow data, (5) complexity comparison of a system with different transformation, (6) mechanism description of subset regression model by phase-space concept, (7) nonlinear modeling of streamflow process and its one-lead time forecasting and (8) nonlinear model for multi-lead time forecasting in areal rainfall of watershed. Firstly, the fractal geometry is used to analyze the tempo-spatial variation for hydrologic variables and three parts are included. On the estimation of point-fractal dimension and its variability for hydrologic process, the capacity dimension, information dimension and correlation dimension are calculated. The results show that the rainfall and runoff processes are both multi-fractals. The former is far more complex than the latter. When the observation time-scale increases, the lacunarity decreases. The decreasing patterns are different in degree and have a strong non-uniform tendency on the temporal distribution. On the analysis and clustering for areal fractal characteristics, the spatial variation and explanation with the temporal dimension of rainfall are the same of our understanding from experience. As for the relation between the temporal dimension of watershed runoff process and the spatial dimension of river basin network, they are correlative in degree. Not only the lacunarity measure can be served as the verification tool, once the systems have the same dimensions, but also is a powerful tool to detect or test whether the external impact or not. On the identification of time-scale law and fractal approach of frequency analysis for hydrologic variable, the rainfall is evident that scale invariance exists in time and clustering will decrease in accord with the increase of threshold. Under the variation of threshold, it can be verified that the maximum values of the homogenous scale-invariant interval are just the same. In addition, taking the probability-scale law on different levels of threshold, the relation can be established between the saturation scale (return period) and the threshold (design hydrologic variable). The methodology is different from the traditional one, i.e. frequency analysis method. When the time-scale we concerned is increasing up to one-year, the fractal scaling law of storm can be established readily. Secondly, the chaotic dynamics is used to analyze the temporal variation of hydrologic variables and three parts are also included. On the existence identification of attractor for the daily streamflow data, the correlation dimension and maximum Lyapunov exponent are calculated. The results show that the chaotic attractor exists and suggests that it may be described by 3 or 4 variables. This information can be used to reconstruct the motion trajectory of a system. On the complexity comparison of a system with different transformation, four types of data are selected and the complexity is white noise, daily rainfall, daily streamflow and Lorenz attractor in descending order. Only the white noise is apparently in disorder, the others have some kinds of pattern or structure. This measure shows the power on the verification. Once the data are transformed by the integration and moving average, the complexity will be decreased. It implies that the model can be established with a simpler way and have a longer lead-time predictability. On the mechanism description of subset regression model by phase-space concept, the subset-order model has the dual properties, the framework of mathematical theory and the mechanism of the physical explanation, and this result is proved by the case studies. Moreover, the multi-dimensional nonlinear subset-order model is extended to the tempo-spatial related model. Finally, the nonlinear model is used for the time series simulation and forecasting and two aspects are concerned: (1)On the nonlinear modeling of streamflow process and its one-lead time forecasting, many models are compared. The results shows that the combined phase-space theory with the artificial neural network framework is excellent and it is powerful on the learning of nonlinear behavior for the artificial neural network. (2)On the nonlinear model for multi-lead time forecasting in areal rainfall of watershed, it is verified that the nonlinear model based on the chaotic dynamics is superior to the linear one based on the time-lag correlation. When the lead-time for forecasting is increased, the former is apparently better than the latter on the overall performance and it is still robust. If a time series is proved a chaotic process, it belongs a nonlinear problem. It must come from the nonlinear point of view and not treated as a linear one. Many modern concepts and tools are used to analyze the complex hydrologic systems in this study, but much of them are also introduced the results from the traditional one. The purpose is to provide a reference and a comparison with different methodologies. Viewing the differences, we overcome the shortcomings, instead of the strong points of the other. The description of complex system will be improved in advance.