Forecasting the future water height of a reservoir with temporal data mining

Abstract

Throughout last few decades, flooding has been one of the most expensive natural disasters, which has had an increased impact on human causalities, property damage and rehabilitation. Thus, there is a crucial need in taking necessary actions to avoid or minimize the impact caused by floods to human lives as well as to the stability of the economy of the country. As to the most recent devastating flood experience faced in Sri Lanka in 2016, one of the root causes identified was the unpredictable decision making strategy, which was used to control and manage excess water in reservoirs due to heavy rain. Moreover, reservoirs are a key water storage source in water management, where they are utilized by various sectors for different purposes. Therefore, there is an essential need in taking the best decision in releasing water from reservoirs to minimize the damage caused by the floods and to optimize the utilization of water as a scarce resource. Many researches have been carried out on the field of data science based on collected data from rivers, reservoirs & tanks to support decision making in water management. Those are mainly focused on classifying water release rules and ranges of a reservoir or tank. Research on forecasting the future water height of reservoirs, with both rainfall and uncertain water inflow due to human intervention, are very limited. Hence, this research focuses on predicting the future water height of a reservoir, when the water inflow is uncertain and the reservoir receives a significant amount of rainfall. This would allow to minimize the risk of deadly floods by opening the sluice gates in time and to manage the water in an optimum manner for irrigation purposes. This research proposes the most effective set of features to forecast the water height of a reservoir on next three days. Furthermore, it presents the comparison of the performance of different regression models and the effectiveness of applying clustering techniques on top of the regression models. The result obtained from this research demonstrates that, K-Medoids clustering with feed forward artificial neural network model has the best performance in forecasting the reservoir water height when there is a significant amount of rainfall and the water inflow is uncertain.