Powered by OpenAIRE graph
Found an issue? Give us feedback
addClaim

AI-based river discharge forecasting

Authors: Silva, Luís Filipe Guedes Borges da;

AI-based river discharge forecasting

Abstract

Neste documento introduzimos o problema de previsão de caudal de rios que tem um papel central no controlo de qualidade de água, prevenção de cheias, avaliações de risco, planeamento de barragens hidroelétricas e gestão de reservatórios e escassez de água. Há um custo financeiro elevado associado a esta tarefa uma vez que é dependente de dados hidrológicos, tais como o caudal, que são caros de obter. No contexto deste problema, conduzimos o nosso trabalho com dois objetivos em mente. Primeiramente, alcançar um modelo capaz de prever as variações temporais do caudal de um rio de forma fiável e precisa e, em seguida, determinar se é possível treinar um modelo que rescinda da necessidade de dados hidrológicos. Para esse efeito selecionámos quatro modelos: Regressão Linear, utilizado como referência pela sua simplicidade e baixo custo computacional. Long Short-Term Memory, pela capacidade de introduzir memória no processo de previsão; Híbrido Convolutional Neural Network - Long Short-Term Memory, pela sua capacidade de reconhecer e extrair padrões; Bidirectional Long Short- Term Memory, por ter contexto completo de todos os pontos de informação ao fazer previsões. Tínhamos à disposição quatro conjuntos de dados relativos a rios distintos. Monitorizámos as performances dos modelos propostos em três conjuntos de preditores, a que chamámos cenários: caudal (Univariate); temperatura e precipitação recente (Multivariate); temperatura, precipitação recente e, o recurso criado por nós, precipitação não-recente (Multivariate_nrp). Constatámos que o cenário Univariate produz os melhores resultados e é preferível sempre que for uma opção viável. As melhores performances neste cenário foram obtidas com o modelo de Regressão Linear. Constatámos também que integrar precipitação não-recente nas nossas previsões melhorou a performance dos nossos modelos, fazendo assim com o cenário Multivariate_nrp seja mais aconselhável do que Multivariate. No entanto, descobrimos que o conhecimento adquirido de um rio não é transferível para outro através de aprendizagem por transferência, significando que para ter performances adequadas num dado rio, o treino deve ser feito num conjunto de dados relativo ao mesmo e requer conhecimento acerca do seu caudal. Considerámos então termo-nos sucedido em atingir um modelo fiável e preciso mas mal-sucedido em desenvencilhar-nos da dependência de dados hidrológicos.

In this document we introduce the problem of river discharge forecasting which plays a central role in water quality control, flood prevention, risk evaluation, design of hydroeletric damns, and the management of reservoirs and water scarcity. There is a high financial burden associated with this task as it relies heavily on hydrological data, such as the river discharge, which is costly to collect. In the context of this problem we conducted this work with two goals in mind. Firstly, to engineer a model capable of accurately and reliably portraying river discharge temporal variations and, secondly, to determine if it is possible to train a model that forgoes the need for hydrologic data. To that end we selected four models: Linear Regression, to serve as a benchmark for its simplicty and low computational cost; Long Short-Term Memory, capable of introducing memory into the prediction process; Convolutional Neural Network - Long Short-Term Memory Hybrid, for the added ability to recognize and extract patterns; Bidirectional Long Short-Term Memory, complete context of all points of information when making predictions. Available to us were four datasets pertaining to distinct rivers. We monitored the proposed models’ performances on three different sets of predictors, which we called scenarios: river discharge (Univariate); temperature and recent precipitation (Multivariate); temperature, recent precipitation and, the engineered feature, nonrecent precipitation (Multivariate_nrp). We found that the Univariate scenario produced the best results and is to be preferred whenever it is a viable option. The best performances within this scenario were obtained using the Linear Regression model. We also found that integrating non-recent precipitation into our predictions boosted the performance of our models, making Multivariate_nrp more advisable than Multivariate scenario. However, we found that the knowledge acquired from one river was not transferable to others through Transfer Learning, meaning that to have adequate performances on a given river, training must be done on a dataset pertaining to that river and requires knowledge about its river discharge. We therefore considered to have succeeded in achieving an accurate and reliable model but failing to rid ourselves of the dependence on hydrological data.

Mestrado em Robótica e Sistemas Inteligentes

Country
Portugal
Related Organizations
Keywords

River discharge forecasting, Long short-term memory, Deep learning, Convolutional neural network, Linear regression, Bidirectional long short-term memory, Transfer learning

  • BIP!
    Impact byBIP!
    selected citations
    These citations are derived from selected sources.
    This is an alternative to the "Influence" indicator, which also reflects the overall/total impact of an article in the research community at large, based on the underlying citation network (diachronically).
    0
    popularity
    This indicator reflects the "current" impact/attention (the "hype") of an article in the research community at large, based on the underlying citation network.
    Average
    influence
    This indicator reflects the overall/total impact of an article in the research community at large, based on the underlying citation network (diachronically).
    Average
    impulse
    This indicator reflects the initial momentum of an article directly after its publication, based on the underlying citation network.
    Average
Powered by OpenAIRE graph
Found an issue? Give us feedback
selected citations
These citations are derived from selected sources.
This is an alternative to the "Influence" indicator, which also reflects the overall/total impact of an article in the research community at large, based on the underlying citation network (diachronically).
BIP!Citations provided by BIP!
popularity
This indicator reflects the "current" impact/attention (the "hype") of an article in the research community at large, based on the underlying citation network.
BIP!Popularity provided by BIP!
influence
This indicator reflects the overall/total impact of an article in the research community at large, based on the underlying citation network (diachronically).
BIP!Influence provided by BIP!
impulse
This indicator reflects the initial momentum of an article directly after its publication, based on the underlying citation network.
BIP!Impulse provided by BIP!
0
Average
Average
Average
Upload OA version
Are you the author of this publication? Upload your Open Access version to Zenodo!
It’s fast and easy, just two clicks!