Time Series Regression for Forecasting Flood Events in Schenectady, New York

Abstract

Floods typically occur due to ice jams in the winter or extended periods of precipitation in the spring and summer seasons. An increase in the rate of water discharge in the river coincides with a flood event. This research combines the time series decomposition and the time series regression model for the flood prediction in Mohawk River at Schenectady, New York. The time series decomposition has been applied to separate the different frequencies in hydrogeological and climatic data. The time series data have been decomposed into the long-term, seasonal-term, and short-term components using the Kolmogorov-Zurbenko filter. For the application of the time series regression model, we determine the lags of the hydrogeological and climatic variables that provide the maximum performance for the model. The lags applied in the predictor variables of the model have been used for the physical interpretation of the model to strengthen the relationship between the water discharge and the climatic and hydrogeological variables. The overall model accuracy has been increased up to 73%. The results show that using the lags of the variables in the time regression model, and the forecasting accuracy has been increased compared to the raw data by two times.