Design and Implementation of a DAG-based Water Level Prediction Approach

Flood is a hazardous disaster which threatens the United States and its territories as it affects millions of people every year. In this context, having suitable approaches for predicting flood threats in the future may benefit humans in reducing property losses as they will be able to more easily prepare in advance. There has been much research centering on how to reduce the flood damage to human life and property, including research that centers on developing early warning systems which simulate the dynamics of water flows by using hydrological models. These models utilize complex mathematical and physics equations to forecast water levels. Moreover, many cities have developed sensor-based flood monitoring systems that collect flood-related data, such as the amount of rainfall and water levels along streams. The goal of this thesis is to take advantage of these data and develop a data-driven water level prediction approach that extrapolates the past into the future. In particular, a DAG-based multi-target prediction (DBMTP) approach is proposed for this purpose. The Directed Acyclic Graph (DAG) is used to model upstream-downstream dependencies between measuring points. DBMTP not only uses historical rainfall and water level information to predict water levels in the future, but also learns dependencies between neighboring sensors. This approach chains and feeds predecessors’ water level predictions into the prediction models for downstream locations. We evaluate the performance of this approach in two case studies involving watersheds in Harris County. The experiments show that DBMTP improves prediction accuracy in some cases over traditional regression approaches.