AAU Student Projects - visit Aalborg University's student projects portal
A master thesis from Aalborg University

River Flood Forecasting using Long-Short Term Memory Neural Networks

[Forudsigelse af Flodoversvømmelse ved Brug af Long-Short Term Neurale Netværk]

Author(s)

Term

4. Term

Education

Publication year

2022

Submitted on

2022-05-28

Pages

77 pages

Abstract

Dette projekt prøver at løse problemet med forudsigelse af flodoversvømmelser ved udvikling af en datadrevet metode baseret på Long-Short Term Memory (LSTM) neurale netværk som kan generelt bruges til alle afvandingsområder. Vi har udviklet en metode som vælger blokke til et glidende gennemsnit af nedbøren ved brug af Spearmans rangkorrelationskoefficienten. Vi bruger metoden til at forudsige vandspejlet ved Abelones Plads, Vejle. En Deep Residual LSTM med 8 skjulte lag blev trænet som viste sig at være den bedste model med hensyn til validerings MSE. Vi viser at modellen kan forklarer op til 67\% af variansen af vandstandsændringen med en forudsigelsestid på 3 timer. Desuden viser vi at alle situationer, hvor modellen fejler at være bedre end forsinkelsesmodellen som bruges som sammenligningsgrundlag, kan være forklaret at være forårsaget af rumligt heterogen nedbør som er ikke målt korrekt af regnmåleren som bruges af modellen. Vi konkluderer at LSTM modeller kan være bruges til forudsigelse af flodoversvømmelser.

This project attempts to solve the problem of river flood forecasting by providing a data-driven methodology that can be generalised to any catchment using Long-Short Term Memory (LSTM) neural networks. A feature selection method to automatically select moving average blocks for a tapped delay line of rainfall measurements using the multiple Spearman rank correlation coefficient was developed. We applied our method on Abelones Plads, Vejle. A Deep Residual LSTM with 8 layers was found to obtain the best validation performance of all trained networks. We show that it explains up to 67\% of the variance in the water level change with a prediction time of 3 hours. We then show that all events where the LSTM fails to outperform the zero-order forecast benchmark model can be attributed to highly spatially variable rainfall which we fail to accurately capture given the location of the rain gauge available to us. We conclude that LSTM models can be used to forecast floods.

Keywords

Documents


Colophon: This page is part of the AAU Student Projects portal, which is run by Aalborg University. Here, you can find and download publicly available bachelor's theses and master's projects from across the university dating from 2008 onwards. Student projects from before 2008 are available in printed form at Aalborg University Library.

If you have any questions about AAU Student Projects or the research registration, dissemination and analysis at Aalborg University, please feel free to contact the VBN team. You can also find more information in the AAU Student Projects FAQs.