Comparison of ANN-NARX, SMAR-LTF and LPMLTF Flood Forecasting Models in the Nzoia Basin, Kenya

Abstract

Flooding is a perennial problem that occurs along several riverine environments, causing loss of property and human lives. The Nzoia Basin in Kenya is one of the areas prone to annual flooding, especially the lower reaches such as Budalang’i. This study assesses three flood forecasting techniques with a view of determining the most efficient model for flood forecasting in the Nzoia Basin. The three techniques considered are SMAR-LTF, ANN-NARX and LPM-LTF techniques. Each of the three simulation models was calibrated individually, initially without updating, then subsequently updated based on either of the two updating techniques (LTF or NARX). The study found that ANN-NARX forecasts had a slightly higher correlation to the observed discharge data when compared to the SMAR-LTF technique, although both had high coefficient of determinations (�������� > 0.9) for up to 6 day lead forecasts. For LPM-LTF technique, forecasts had a high decay rate beyond one lead-day forecasts, with �������� being less than 0.8, well below SMAR-LTF and ANN-NARX. This is consistently evident for the flood event periods considered during the study. It can thus be concluded that for flood forecasting in the Nzoia Basin, ANN-NARX is the best technique, although SMAR-LTF had close results to the ANN-NARX technique for all the 6 lead-day forecasts.