Aristotle University Of Thessaloniki (AUTH)
Προσομοίωση της σχέσης βροχόπτωσης - απορροής με μεθόδους μηχανικής μάθησης
Abstract
dc:descriptionThe conversion of precipitation to runoff is considered one of the most important research areas in the science of hydrology. The rainfall-runoff modelling provides considerable information about the hydrologic behavior of a watershed and given the probability of appearance of flood events which usually caused from extreme weather phenomena. The watershed management and the planning of hydraulic technical infrastructure depends on water influx, losses and final discharges. The main water influx in the watershed comes from precipitation (rainfall and snowfall) and the losses include infiltration, evapotranspiration and interception. Finally the remaining water from the difference between influx and losses is the water discharge. The complexity of the rainfall-runoff relationship and the numerous parameters it involves make it difficult to determine a mathematical function that will describe precisely the final runoff. Artificial intelligence and machine learning methods have the important advantage that they can approach the hydrological phenomena without requiring the exact knowledge of the physical parameters that affect the rainfall-runoff relationship. The fact that these methods have the ability to determine a relationship between inputs and target data makes them powerful computational tools for the simulation of complex water processes such as the transformation of precipitation to runoff. In the present dissertation four models were developed for the rainfall-runoff simulation, each one using a different machine learning approach. More specifically the author has developed models of artificial neural networks, Bayesian neural networks, support vector regression models and ensemble methods (Boosting and Bagging). We make a special mention to the Boosting learning algorithm, which was developed in order to be adaptable to the special characteristics of the meteorological and hydrological data, providing accurate predictions both of normal and extreme runoff values. This particular model named HydroBAL which is the initialism for “Hydrologic Boosting ALgorithm”. The models were applied to three watersheds for which data of precipitation and runoff were available. The application of the models showed that all of them simulate efficiently the rainfall-runoff relationship and their differences are minor. An important aspect of the models is their considerable accuracy in forecasting the extreme values. These values rarely appear in the time series data (inputs-targets) and diverge from the normal data range, so their prediction is very difficult. However these values are important because they are directly connected to flood events. Finally the effectiveness of these models at all levels of runoff simulation and forecasting create positive expectations for their generalization applicability.
Degree
thesis:*- Grantor dc:publisher
- Aristotle University Of Thessaloniki (AUTH)
- Year dc:date
- 2016
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Botsis, Dimitris
- Μπότσης, Δημήτριος
Subjects
dc:subject × 16Rights
- Language dc:language
- gre
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/37295