Precipitation is the driving force affecting the temporal and spatial variability of many hydrosystem responses, e.g., drainage basin, dam and sewer networks. When the reliability or sensitivity of such systems is concerned, simulation study is usually utilized to generate synthesized rainfall sequence for model input. Since long-term physical models of rainfall are not easily accessible, most of the current adopted models to synthesize rainfall sequence are essentially empirical and descriptive, and they fall into either parametric or nonparametric category, or mixture of the two. No consensus, however, is achieved regarding their appropriateness in application....[ Read more ]

Precipitation is the driving force affecting the temporal and spatial variability of many hydrosystem responses, e.g., drainage basin, dam and sewer networks. When the reliability or sensitivity of such systems is concerned, simulation study is usually utilized to generate synthesized rainfall sequence for model input. Since long-term physical models of rainfall are not easily accessible, most of the current adopted models to synthesize rainfall sequence are essentially empirical and descriptive, and they fall into either parametric or nonparametric category, or mixture of the two. No consensus, however, is achieved regarding their appropriateness in application.

Motivated by the above problem, this study aims to provide an insightful investigation into these two classes of methods and especially detail the advantages of nonparametric methods in the context of hydrological engineering. We propose that a reasonable rainfall sequence generation mechanism should be based on past observations, faithful in the probabilistic sense and infinitely producible. Based on these realizations, argument of this thesis is developed.

First, it is pointed out the parametric models, imposing structural information, may inappropriately represent the data, termed ‘mis-specification’. On the contrary, the nonparametric method is in general asymptotically consistent. Chapters 1, 2 and 3 are related to this topic. Moreover, we investigate two nonparametric methods in detail, especially their finite sample performance. This is mainly through simulation study. For one of them, termed ‘Taylor-Thompson method’, we provide its global property in close form and propose a data-based smoothing parameter selection method. We also address that in multi-dimension, nonparametric-based generation of random vector is more natural. We find no method universally outperforms the other one in preserving all the characteristics of sample, so selection of two methods depends on issue of interest. These topics are included in Chapter 3. At last, an example is given in Chapter 4 for illustrative purpose. Artificial rainfall sequences are generated by the two nonparametric methods under investigation. The sequence is combined in a rainfall-runoff-routing model, with downstream flood the ultimate focus. The difference of model outputs is examined in the engineering point of view. Comments in Chapter 5 conclude the thesis.