Forecasting shellfish contamination by marine biotoxins based onmultivariate time series

• Autor(a) principal: Cruz, Rafaela Carreira Eleutério Gregório da
• Data de Publicação: 2022
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10362/139558
• topic: Shellfish Contamination; Time Series; Forecasting; Artificial Neural Networks; Feed-Forward Neural Networks; Convolutional Neural Networks; Domínio/Área Científica::Engenharia e Tecnologia::Engenharia Eletrotécnica, Eletrónica e Informática

Shellfish production has been growing in recent years, having a high impact in several Portuguese coastal regions. However, this resource can be contaminated by marine biotoxins produced by toxic phytoplankton. The consumption of contaminated shellfish can cause serious health problems, and hence the harvesting and commercialisation of this product are prohibited whenever biotoxin concentration exceeds the safety limits. Since this prohibition leads to severe economic losses, it becomes necessary to develop strategies that predict shellfish contamination. Biotoxin concentration in bivalve molluscs can be predicted using univariate and multivariate time series, by modelling past information to predict the future. These time series include historical data on in-situ measurements of biotoxin concentration in several shellfish species, as well as other biological and meteorological data. In this thesis, multiple time series were acquired from different sources, integrated and pre-processed. Afterwards, various univariate andmultivariate time series forecasting methods were developed to predict mussel contamination in multiple production areas. In this context, autoregressive models and artificial neural networks (ANNs), such as feed-forward, convolutional and long short-term memory (LSTM) networks, were tested. Additionally, various data preparation and feature engineering methods were explored to improve these models. The forecasting models were evaluated and compared in order to determine which are the most suitable to solve the problem at hand. The results showed that the ANNs, namely networks trained on data whose dimension had previously been reduced using an autoencoder and networks trained on univariate time series, outperformed the classic autoregressive models. Moreover, among the ANN models, the LSTMs were very accurate, especially at one-week ahead predictions. Finally, the multivariate models did not outperform the univariate models, which may be explained by the fact that the additional variables used in this thesis did not provide relevant information to forecast shellfish contamination. These results might be regarded as the first pivotal steps towards the development of a model-based forecasting tool, which will allow the production sector to anticipate the harvesting prohibition, enabling the development of strategies to mitigate the economic losses inherent to this situation.