Bayesian Networks (BNs), as an explicit representation of the joint probability distribution characterising a problem domain and providing a topological description of the causal relationship among the variables, are rapidly becoming the tool of choice for uncertain reasoning in artificial intelligence (AI). In contrast to other fields of AI, BNs can be learned by combining domain knowledge with statistical data due to explicit representation of the expert or domain knowledge. Consequently, one can more easily interpret and understand the knowledge encoded in the representation. In addition, BNs can readily handle incomplete data sets and even offers an efficient and principled approach for avoiding the over fitting of data. The thesis investigates the applicability of BNs in the two different fields of hydroinformatics, namely hydraulics and hydrology. In the first application, static BNs, in which the variables and the topology of the variables do not change in the time, are used to assess the risk of structural collapse of the sewer (ROSCOS). Two years records of actual structural collapse data of the sewer in the part of city of London obtained from the Thames Water are analysed and used to build the BNs with combination of domain knowledge. Then the constructed BNs are validated using unseen actual collapse cases of the sewer from the same locality. The validation of BNs shows quite satisfactory results in assessment of the ROSCOS. In the second application, Dynamic Bayesian Networks (DBNs) are used for rainfall-runoff modelling of Xixian catchment which is the sub catchment of Huai River basin of P.R. China. As the runoff generation process is highly non-linear, time varying and spatially distributed, the underlying dynamics of the system is investigated using deterministic Chaos theory. Firstly, Local Linear Models (LLMs) are used to forecast runoff on the basis of observed runoff only. Secondly, rainfall is also used as input to the LLMs and the results show considerable increase in the predictive performance. Finally, Hidden Markov Mixture of Experts (HMME), which is one of the types of DBNs, is coupled with LLMs. The results shows good matching between observed and predicted runoff. The investigation of these techniques show promising results in the above two fields of Hydroinformatics in solving water-related problems. In general, the objective of this study has been achieved, but more investigation should be done for application of HMME.
Back to the list of MSc abstracts