A comprehensive evaluation of boosting algorithms for artificial neural network-based flow forecasting models

The severity of urban flooding is expected to increase in coming years due to drivers such as climate change and rapid urban growth. Flood damage can be mitigated using early warning systems (EWS), which provide emergency services with an advanced notice of flood likelihood, allowing them to take preventative measures.

Machine learning methods have proven to be suitable approaches for modelling the complex relationships between hydrological variables and the flow forecasts needed for EWS. However, these models perform poorly during infrequent, extreme flow events.

In recent years, the boosting meta-algorithm has emerged as a powerful tool for improving model performance for these types of extreme events. Generally speaking, boosting techniques involve iteratively recalibrating a model, where each subsequent calibration focuses on improving the worst performing observations in the previous model.

The objective of this research is to provide a comprehensive comparison of different boosting techniques for artificial neural network (ANN) based flow forecasting EWS models. Two variants of boosting are compared in this research to evaluate their efficacy in improving flow forecasts for EWS: adaptive boosting and gradient boosting. Adaptive boosting is implemented both as a resampling technique and for weighting in the model calibration process. Gradient boosting is implemented using a variety of different cost functions.

The effects of each boosting technique are assessed by evaluating the ANN performance using several performance measures to quantify the overall performance improvement, and specifically high-flow performance and ensemble diversity. Preliminary results show how boosting techniques improve overall ensemble mean performance yet do little to improve peak flow accuracy. Additionally, boosting tends to reduce ensemble diversity. These effects are undesirable for EWS: despite improving mean performance of the models, the prediction envelopes of the boosted models contain less of the observed flows compared to the non-boosted models.