Multi-agent system for flood forecasting in tropical river basin

Abstract

It is well known, the problems related to the generation of floods, their control, and management, have been treated with traditional hydrologic modeling tools focused on the study and the analysis of the precipitation-runoff relationship, a physical process which is driven by the hydrological cycle and the climate regime and that is directly proportional to the generation of floodwaters. Within the hydrological discipline, they classify these traditional modeling tools according to three principal groups, being the first group defined as trial-and-error models (e.g., "black-models"), the second group are the conceptual models, which are categorized in three main sub-groups as "lumped", "semi-lumped" and "semi-distributed", according to the special distribution, and finally, models that are based on physical processes, known as "white-box models" are the so-called "distributed-models". On the other hand, in engineering applications, there are two types of models used in streamflow forecasting, and which are classified concerning the type of measurements and variables required as "physically based models", as well as "data-driven models". The Physically oriented prototypes present an in-depth account of the dynamics related to the physical aspects that occur internally among the different systems of a given hydrographic basin. However, aside from being laborious to implement, they rely thoroughly on mathematical algorithms, and an understanding of these interactions requires the abstraction of mathematical concepts and the conceptualization of the physical processes that are intertwined among these systems. Besides, models determined by data necessitates an a priori understanding of the physical laws controlling the process within the system, and they are bound to mathematical formulations, which require a lot of numeric information for field adjustments. Therefore, these models are remarkably different from each other because of their needs for data, and their interpretation of physical phenomena. Although there is considerable progress in hydrologic modeling for flood forecasting, several significant setbacks remain unresolved, given the stochastic nature of the hydrological phenomena, is the challenge to implement user-friendly, re-usable, robust, and reliable forecasting systems, the amount of uncertainty they must deal with when trying to solve the flood forecasting problem. However, in the past decades, with the growing environment and development of the artificial intelligence (AI) field, some researchers have seldomly attempted to deal with the stochastic nature of hydrologic events with the application of some of these techniques. Given the setbacks to hydrologic flood forecasting previously described this thesis research aims to integrate the physics-based hydrologic, hydraulic, and data-driven models under the paradigm of Multi-agent Systems for flood forecasting by designing and developing a multi-agent system (MAS) framework for flood forecasting events within the scope of tropical watersheds. With the emergence of the agent technologies, the "agent-based modeling" and "multi-agent systems" simulation methods have provided applications for some areas of hydro base management like flood protection, planning, control, management, mitigation, and forecasting to combat the shocks produced by floods on society; however, all these focused on evacuation drills, and the latter not aimed at the tropical river basin, whose hydrological regime is extremely unique. In this catchment modeling environment approach, it was applied the multi-agent systems approach as a surrogate of the conventional hydrologic model to build a system that operates at the catchment level displayed with hydrometric stations, that use the data from hydrometric sensors networks (e.g., rainfall, river stage, river flow) captured, stored and administered by an organization of interacting agents whose main aim is to perform flow forecasting and awareness, and in so doing enhance the policy-making process at the watershed level. Section one of this document surveys the status of the current research in hydrologic modeling for the flood forecasting task. It is a journey through the background of related concerns to the hydrological process, flood ontologies, management, and forecasting. The section covers, to a certain extent, the techniques, methods, and theoretical aspects and methods of hydrological modeling and their types, from the conventional models to the present-day artificial intelligence prototypes, making special emphasis on the multi-agent systems, as most recent modeling methodology in the hydrological sciences. However, it is also underlined here that the section does not contribute to an all-inclusive revision, rather its purpose is to serve as a framework for this sort of work and a path to underline the significant aspects of the works. In section two of the document, it is detailed the conceptual framework for the suggested Multiagent system in support of flood forecasting. To accomplish this task, several works need to be carried out such as the sketching and implementation of the system’s framework with the (Belief-Desire-Intention model) architecture for flood forecasting events within the concept of the tropical river basin. Contributions of this proposed architecture are the replacement of the conventional hydrologic modeling with the use of multi-agent systems, which makes it quick for hydrometric time-series data administration and modeling of the precipitation-runoff process which conveys to flood in a river course. Another advantage is the user-friendly environment provided by the proposed multi-agent system platform graphical interface, the real-time generation of graphs, charts, and monitors with the information on the immediate event taking place in the catchment, which makes it easy for the viewer with some or no background in data analysis and their interpretation to get a visual idea of the information at hand regarding the flood awareness. The required agents developed in this multi-agent system modeling framework for flood forecasting have been trained, tested, and validated under a series of experimental tasks, using the hydrometric series information of rainfall, river stage, and streamflow data collected by the hydrometric sensor agents from the hydrometric sensors.