Dam risk analysis and dynamic decision making

Dam failures are among the most catastrophic natural or engineering disasters. A single dam failure could claim the lives of more than 26,000 people as in the case of the 1975 Banqiao dam failure in China. In addition to man-made dams, numerous landslide dams could form during strong earthquakes and heavy storms. For instance more than 250 landslide dams of significant risks were created by the 2008 Wenchuan earthquake. The need for mitigation of the risks due to dam breaks raises questions about the proper risk analysis and efficient emergency management. The main objectives of this thesis are to develop a landslide dam database and a set of empirical models for the prediction of breaching parameters based on the database, establish a new model for the analysis of human risks from dam break floods, and develop a framework of risk-based dynamic decision making for dam break emergency management. Landslide dams differ from embankment dams in geometry, soil compositions, and flood control facilities. The differences largely influence the failure modes of these two types of dams. A large database of landslide dams with 1267 cases from all over the world has been compiled in this study. Based on the database, a set of empirical models are developed for estimating five important breaching parameters (peak flow rate, breaching time, breach depth, breach top width, and breach bottom width) of landslide dams. The breaching parameters of landslide dams and man-made earth and rockfill dams are compared. Direct application of the empirical models for man-made earth and rockfill dams to landslide dams would, on average, overestimate the breach size by more than 60% and the peak outflow rate by approximately 200%, and underestimate the breaching duration by approximately 50%. Analysis of human risks due to dam-break floods is very complex because both objective uncertainties and subjective uncertainties such as human thinking, decision, and behaviour are involved. A HUman Risk Analysis Model for dam-break floods (HURAM) is established in this thesis using Bayesian networks. The model is able to take into account a large number of important parameters (fourteen) and their inter-relationships in a systematic structure; include the uncertainties of these parameters and their inter-relationships; incorporate more available information of physical mechanisms and historical data; and update the predictions using Bayes’ theorem based on available information in specific cases. HURAM allows not only cause-to-result inference, but also result-to-cause inference by updating the Bayesian network with specific information from the study case. A change in any parameter in the model may affect other parameters and the loss of life. The uncertainties of the parameters and their relationships are studied both at the global level using multiple sources of information and at the local level by updating the prior probabilities. Evacuation, sheltering, and building damage are also simulated in HURAM, which is required in estimating evacuation costs and flood damage. Emergent evacuation of the population at risk (PAR) before a dam-break is an efficient way to save human life and properties. A late decision may lead to loss of lives and properties, but a very early evacuation will incur unnecessary expenses. A risk-based framework of dynamic decision making for dam-break emergency management (DYDEM) is developed in this study. A decision criterion is suggested to decide whether to evacuate the population at risk or to delay the decision. The optimum time for evacuating the PAR is obtained by minimizing the expected total loss, which integrates the time related probabilities and flood consequences. The time-related probability of dam failure is taken as a stochastic process and estimated using a time-series analysis method. The time-related flood consequences, including evacuation costs, flood damage and monetized loss of life, are taken as functions of warning time and evaluated using HURAM. When a delayed decision is chosen, the decision making can be updated with available new information. DYDEM is used to help rational decision in the emergency management of the Tangjiashan landslide dam, which was caused by the Wenchuan earthquake in 2008 and posed high risks to the people downstream of the dam. Three stages are distinguished with different levels of information along the timeline of the landslide dam failure event. The dam breaching parameters are predicted with an empirical model in stage 1 when no soil property information is available, and a physical model in stages 2 and 3 when knowledge of soil properties and a division channel have been obtained. Dynamic decision analysis is conducted using DYDEM to find the optimal time to evacuate the population at risk with minimum expected total consequences (MTC) in each of these three stages. The variance of dam-failure probability as a time series increases with prediction lead time. The flood consequences as functions of warning time from HURAM show that the monetized loss of life dominates when the warning time is short. However, if the warning time is relatively long, the evacuation costs would become the largest expenditure. The minimum expected total consequences (MTC), which integrates the time-related probability and flood consequences, do not vary significantly before the predicted optimal time if the PAR is small (e.g. in Beichuan Town) as they are dominated by the evacuation costs. The MTC would increase rapidly a short period after the predicted optimal time as the monetized loss of life gradually dominates the total consequence. Therefore, the times for evacuating the PAR may be brought forward a certain period for more effective evacuation (e.g. at daytime) if the PAR is relatively small. However, if the PAR is relatively large (e.g. in Mianyang City), the expected total consequences are sensitive to the time for evacuating the PAR. Thus, the time for issuing evacuation warning in areas with a large PAR should be considered carefully. The methods of risk analysis and dynamic decision making developed in this thesis are equally applicable to the management of other types of disasters such as levee breaks, flash floods, hurricanes, tsunamis, landslides and debris flows