Comparison of machine learning and rainfall-runoff models for runoff prediction in urban drainage systems
Abstract
This thesis presents a comprehensive investigation into the effectiveness of machine learning
models in comparison to traditional rainfall-runoff models for flood prediction, utilizing the
Hydrologic Engineering Center's Hydrologic Modeling System (HEC HMS) and Storm Water
Management Model (SWMM) as the benchmark because they are established model that are
widely recognized and used hydrological modelling system. The case study focuses on the
Lubigi catchment, an important drainage channel within Uganda. The main objective of this
study is to develop and evaluate the performance of a machine learning model for predicting
runoff in the Lubigi catchment and compare it with the rainfall-runoff model implemented
through HEC-HMS and SWMM.
The study utilized historical daily rainfall data spanning from 1981 to 2022, obtained from the
Uganda National Meteorological Association. The data encompassed rainfall data and then
generated runoff measurements using HEC-HMS, which were divided into a training dataset
(80% of the data) for model development and a testing dataset (20% of the data) for model
validation. Additionally, digital elevation models (DEM), land use, and soil data specific to the
Lubigi catchment were acquired from the Kampala Capital City Authority (KCCA) to facilitate
the implementation of HEC HMS.
For the HEC HMS-based rainfall-runoff model, various components were employed. The
subbasin loss employed the Soil Conservation Service (SCS) Curve Number method, the
subbasin transform utilized the unit hydrograph approach, the reach routing employed the
Muskingum method, and the subbasin precipitation used a specified hyetograph. The model
was calibrated and validated using the available rainfall and runoff data for the Lubigi
catchment.
In parallel, the machine learning model was developed using the Random Forest Regression
algorithm. The model was trained on the historical rainfall and runoff data, and subsequently,
it was utilized to predict runoff based solely on the daily rainfall data and some output of runoff
from HEC HMS for the year 2022.
The performance of the machine learning model and the rainfall-runoff model implemented
through HEC HMS were assessed and compared. Evaluation metric root mean square error
4
(RMSE) was used to determine the effectiveness and reliability of each model in predicting
runoff within the Lubigi catchment.
The findings of this study provide valuable insights into the capabilities and limitations of
machine learning models compared to traditional rainfall-runoff models for flood prediction.
The results contribute to the advancement of flood forecasting techniques and offer guidance
for decision-makers in selecting appropriate modeling approaches for similar hydrological
applications.