Multi-step flood forecasting in urban drainage systems using time-series data mining techniques

Piadeh, Farzad, Behzadian Moghadam, Kourosh ORCID: https://orcid.org/0000-0002-1459-8408 and Alani, Amir (2022) Multi-step flood forecasting in urban drainage systems using time-series data mining techniques. In: Water Efficiency Conference 2022, 14-16 Dec 2022, Trinidad and Tobago.

[img]
Preview
PDF
Piadeh et al. 2022-WATEF manuscript.pdf - Accepted Version

Download (511kB) | Preview

Abstract

While early warning systems are recognised as the most cost-effective solution in urban flood risk management, highly accurate flood forecasting is limited to short-term timesteps, usually less than a few hours especially for prediction of overflowing in urban drainage systems. This study aims to provide a framework for more accurate overflow predictions for longer lead times by using data mining models applied to time series data for multi-step flood forecasting. The framework including event identification, feature analysis and developing models is demonstrated by its application to a pilot study in London. All numerical rainfall data and water levels in urban drainage systems are first turned to the categorical events on which 6 common weak learner models are developed. Then, three new time-series models, including overflowing-based, non-overflowing-based, and accuracy-based, are developed based on these models to predict overflow states among all identified events. Three weak learner models, i.e. discriminant analysis, naive Bayes, and decision tree are considered as the best models based on accuracy, total overflowing detection and total non-overflowing detection. Furthermore, while the accuracy of these models is changed between 95 to 85% from 1 to 12-step ahead of prediction, these models can detect the non-overflow conditions better than overflow detection. To cover this gap, new time series developed models could significantly reduce the overestimation and underestimation of water levels, including correct predicting of 50% of the total events after 12-step ahead by overflow-based model. This result shows the potential of using time-series data-demanding models for effective and highly accurate predictions of overflow events.

Item Type: Conference or Workshop Item (Paper)
Keywords: Data mining; Drainage system; Flooding classification; Multistep prediction Overflow prediction
Subjects: Construction and engineering
Related URLs:
Depositing User: Farzad Piadeh
Date Deposited: 16 Dec 2022 08:05
Last Modified: 05 Jan 2023 14:11
URI: https://repository.uwl.ac.uk/id/eprint/9690

Downloads

Downloads per month over past year

Actions (login required)

View Item View Item

Menu