Long Short Term Memory Water Quality Predictive Model Discrepancy Mitigation Through Genetic Algorithm Optimisation and Ensemble Modeling

A predictive long short-term memory (LSTM) model developed on a particular water quality dataset will only apply to the and may fail make an accurate prediction another dataset. This paper focuses improving LSTM tolerance by mitigating discrepancies in capability that arises when is applied different datasets. Two models are from datasets, Baffle Burnett, optimised using metaheuristic genetic algorithm (GA) create hybrid GA-optimised subsequently combined with linear weight-based technique develop tolerant ensemble model. The successfully predict river terms of dissolved oxygen concentration. After GA-optimisation, RMSE values Burnett decrease 42.42% 10.71%, respectively. Furthermore, two GA-hybrid models, namely average optimal weighted ensemble. GA-Baffle 5.05% for 6.06% ensemble, GA-Burnett 7.84% 8.82%, When tested unseen unrelated predictions, indicating applicability domains outside sector. consistent similar performance any illustrates successful mitigation capacity individual proposed scheme. observed highlights datasets which could potentially predictions.