Predicting Flood Streamflow with Auto Regressive Integrated Moving Average Models

Abstract

Flooding is the most common natural disaster and continues to increase in frequency and intensity due to climate changes [7]. Currently, there is a lack of efficient tools to predict flooding. This research is aimed to create a Time Series Machine Learning (ML) program using Auto Regressive Moving Average (ARIMA) models to forecast streamflow, one of the most prominent factors in flood prediction. A streamflow dataset from the Ganges River, Bangladesh was used to plot several graphs of the river Log Volume to observe possible trends. Another plot was graphed to check and quantify how much the distribution of the stream volume changed over the course of 10 years using KL Divergence. The plot analyses and Partial Autocorrelation Function (PACF) and Autocorrelation Function (ACF) tests were used to help obtain the ARIMA parameters of (p, d, q) as (1, 1, 1). However, the forecasted streamflow of the ARIMA function was not accurate when compared with previously recorded data because of heavy seasonality. As a result, the final program was redesigned with Seasonal ARIMA (SARIMA) to account for the inaccuracy. The SARIMA model was used to forecast the streamflow of subsequent years and was close to the actual recorded data. Such accuracy indicates that this method can be a useful tool in navigating and preparing for floods.