ABSTRACT
Machine learning models have demonstrated a strong potential in the recent decade for streamflow prediction purposes. Even reaching considerable accuracy, machine learning models still present some limitations and little explore aspects on this topic: the accuracy decrease on high streamflow, uncertainty estimation applications, and hydrological interpretation of models. High streamflow values are the most relevant for early warning systems of flood mitigation. However, these records are scarce in the data. Hence, a decrease in accuracy is seen, which gets deeper in extrapolated scenarios.
This thesis proposes two methodologies that support machine learning models with the outputs of a physically based model to enhance the prediction capabilities in high streamflow prediction. The first methodology produces a machine learning model trained with a combination of observed and synthetic high streamflow events generated by a physically based model. The second methodology creates a hybrid model that combines outputs based on a physically based model and the prediction of its residuals by a machine learning model. Both methodologies have shown accuracy improvements compared to models trained only with observed data, reaching reductions of root mean square errors of more than 23% for streamflow values larger than the 3-year return period in the study area. However, the second methodology reaches closer approximations to the peak observed value with significant extrapolation capabilities. Despite the accuracy, the uncertainty related to the model is a main topic to explore.
A methodology to estimate the uncertainty of the output of a streamflow prediction model by employing machine learning techniques is developed in the thesis. The results of the methodology show that the distributions of the residuals can be acceptably described. Consequently, the uncertainty interval covers 87.9% of the observed values higher than the 3-year return period with a width of interval significantly smaller than the broadly used Box-Cox method. Finally, the hydrological interpretation of a machine learning model for streamflow prediction is undertaken.
The results show that the model may identify areas of the catchment whose runoff considerably contributes to the control point, as well as period of previous soil saturation, and the impact of the highest precipitation values in the model’s prediction. When the hydrograph rises steeply, the model acceptably considers the recent accumulated precipitation as the most relevant features along with previous saturation. As the hydrograph peak approaches and during the falling limb, previous streamflow acquires main relevance supported by runoff of distant regions. The machine learning model can suitably interpret the catchment system and provide valuable information.
PhD Advisors:
-
- Prof Ernest Bladé
- Prof Fernando Salazar
CANDIDATE
Mr Sergio Ricardo López Chacón is a PhD candidate in Civil and Environmental Engineering, part of CIMNE’s Machine Learning and Models in Hydro-Environmental Engineering research cluster.