<![CDATA[Accurate prediction of necessary water quality parameters such as Dissolved Oxygen (DO) is very critical in precision aquaculture and is essential for performance-based decision-making. This thesis fills the gap between reactive monitoring and predictive intelligence through the construction of a solid machine learning infrastructure. We convert high frequency multivariate time series data into a supervised learning problem by an advanced feature engineering process that generates temporal predictions including lag features and rolling window statistics. A Light Gradient Boosting machine (LightGBM) algorithm trained using the above-mentioned engineered dataset has an extreme predictive power. Results of single-variable interpretation analysis showed that short term data, especially the 5-minute rolling statistics of DO and turbidity variability, are the main driving factors for the model prediction. This research confirms that a feature-engineered LightGBM approach is a computationally efficient, but highly accurate approach to supporting the development of early warning systems in modern aquaculture as a computationally scalable approach.]]>
Copyrights © 2026
