Evaluating artificial intelligence models for flood and drought forecasting in tropical and arid climatic regions | |
| Author | Adikari, Adikari Appuhamilage Kasuni Erandika |
| Call Number | AIT Thesis no.WM-20-05 |
| Subject(s) | Artificial intelligence--Forecasting Flood forecasting rought forecasting |
| Note | A thesis submitted in partial fulfillment of the requirements for the degree of Master of Engineeing in Water Engineering and Management |
| Publisher | Asian Institute of Technology |
| Abstract | With the advancements in the field of computer science, Artificial Intelligence (AI) has been vastly used in many fields to increase performances and productivity. In the field of water engineering, finding methods for mitigation of disasters due to climatic changes is one of the key areas in focus. Therefore, it is essential to identify methods to accurately forecast disasters in order to take necessary disaster management steps. Therefore, this study focused on analyzing the performance of novel AI techniques in the problem of flood and drought forecasting. Three AI models; Convolutional Neural Networks (CNN), Long-Short Term Memory networks (LSTM) and Adaptive Neuro Fuzzy Inference System combined with Wavelet Decomposition functions (WANFIS) were used to simulate the forecasts. These simulations were carried out in two regions; Lower Darling river basin and Sekong river basin, that had different climatic conditions compared to each other. The drought forecasting was done using a key meteorological drought indicator, Standard Precipitation Index (SPI), whereas discharge and Antecedent Precipitation Index (API) were used to indicate floods in the models. The models were designed to forecast one-time-lag ahead forecasts. Hence, droughts were predicted for next month whereas river flow was predicted for the following day, using the most recent past data. The analysis was done with the use of output plots, distribution analytics, drought/flood indicators and five evaluation criteria. The analysis performed on drought forecasting of Lower Darling basin suggested the use of WANFIS compared to the other two models in drought forecasting in arid regions. The drought simulations by models in Sekong basins seconded this suggestion. Outputs of both regions proposed the use of CNN for pluvial floods regardless of the climatic conditions of the regions. It was identified that WANFIS produce results for forecasts that use inputs of a single property whereas CNN performed better when multiple properties (that are independent with each other) are used as inputs. Besides, an early warning tool for floods and droughts were developed to demonstrate the use of the outcomes in practical applications. |
| Year | 2020 |
| Type | Thesis |
| School | School of Engineering and Technology (SET) |
| Department | Department of Civil and Infrastucture Engineering (DCIE) |
| Academic Program/FoS | Water Engineering and Management (WM) |
| Chairperson(s) | Shrestha, Sangam; |
| Examination Committee(s) | Dailey, Matthew N.;Shanmugan, S. Mohana; |
| Scholarship Donor(s) | AIT Fellowship; |
| Degree | Thesis (M. Eng.) - Asian Institute of Technology, 2020 |