Hybrid Model for Dissolved Oxygen Prediction Using Ensemble MINE-BISRU-Attention and LightGBM-BiSRU-Attention Approaches

Abstract

Aquaculture productivity in most cases depends on the quality of water which is one of the determinants of water health. One of the most important variables needing monitoring and control is the dissolved oxygen (DO). Even with complex relationships typical in aquaculture processes, some traditional methods of prediction have more often than not been ineffective. In this work, the LightGBM-BiSRU-Attention hybrid approach is proposed which uses LightGBM for feature selection, BiSRU for sequence learning and Attention for parameter tuning. Further, an enhanced model called Ensemble MINE-BISRU-Attention is proposed which further explores the use of the Maximal Information Coefficient (MIC) for more advanced feature selection. In this study, these models were evaluated against the Kaggle water quality dataset and consistently with good accuracy. The models predicted the average mean square error with the LightGBM-BiSRU-Attention model to be 0.178 while the Ensemble MINE-BISRU-Attention further lowered the number to 0.104. This also acts as a shift towards the intelligent aquaculture systems which provide a solution to the gaps which existed in water quality prediction models.