Flood Frequency Modeling and Prediction of Beki and Pagladia Rivers Using Deep Learning Approach

摘要

Flood is one of the major reasons of death and despair in the North East Indian state of Assam during the monsoon months each year. The flooding in adjacent areas through which the Beki and Pagladia rivers in Assam flow has been a primary source of damage and destruction during the monsoon season. Traditional and data aided approaches are combined to study and predict the occurrence of flood involving these two rivers. Though traditional approaches have been extensively used, the recent trend is towards the use of learning aided methods. Primarily, a few deep learning approaches like Stacked Auto Encoder (SAE) connected with Tapped Delay Line (TDL) blocks are combined and integrated with layers of TDL-Long Short Term Memory (LSTM) units and bidirectional LSTM (BLSTM) cells to form a configuration which has been extensively trained. This learning involves use of data sets with monthly, weekly, daily and hourly records of significant river flow parameters of Beki and Pagladia to predict these for subsequent periods of time and generate decision states of flooding caused during the rainy season. Data collected from secondary sources and historical records have been combined to train the above mentioned deep learning based approaches to estimate the probability of occurrences of flood. Also a look ahead predictor approach has been configured and used as a part of the system which uses the recorded data to forecast occurrence of flood within certain time intervals.