Speaker
Description
Recent advancements in computing power and the availability of online resources have fueled the increasing popularity of machine learning in modeling wastewater treatment plants. This study capitalizes on this trend by developing neural networks and ensemble models for predicting influent BOD using real data from a Sequential Batch Reactor (SBR) plant in the Philippines. The study employed a series of preprocessing techniques and introduced the Yeo-Johnson transformation, which improved the distribution of the data and significantly decreased bias while lowering variance. Through this process, three influent optimal features—COD, NH4-N, and PO4-P—were identified. Subsequently, DNN, LSTM, and CNN models were developed and fine-tuned using Bayesian optimization. The DNN model demonstrated superior generalization compared to LSTM and CNN models, achieving an R2 score of 0.35. Additionally, XGBM emerged as having better generalization than LGBM, with an R2 score of 0.25. Despite the success in model development, these models present opportunities for further improvement and may eventually be integrated into real-time operation for proactive decision-making to optimize performance and resource allocation in municipal wastewater treatment plants utilizing SBR technology in the Philippines.
