Generalised Linear Models for Prediction of Dissolved Oxygen in a Waste Stabilisation Pond

Abstract

Due to simplicity and low costs, waste stabilisation ponds (WSPs) have become one of the most popular biological wastewater treatment systems that are applied in many places around the globe. Increasingly, pond modelling has become an interesting tool to improve and optimise their performance. Unlike process-driven models, generalised linear models (GLMs) can deliver considerable practical values in specific case studies with limited resources of time, data and mechanistic understanding, especially in the case of pond systems containing vast complexity of many unknown processes. This study aimed to investigate the key driving factors of dissolved oxygen variability in Ucubamba WSP (Ecuador), by applying and comparing numerous GLMs. Particularly, using different data partitioning and cross-validation strategies, we compared the predictive accuracy of 83 GLMs. The obtained results showed that chlorophyll a had a strong impact on the dissolved oxygen (DO) level near the water surface, while organic matter could be the most influential factor on the DO variability at the bottom of the pond. Among the 83 models, the optimal models were pond- and depth-specific. Specifically, among the ponds, the models of MPs predicted DO more precisely than those of facultative ponds; while within a pond, the models of the surface performed better than those of the bottom. Using mean absolute error (MAE) and symmetric mean absolute percentage error (SMAPE) to represent model predictive performance, it was found that MAEs varied in the range of 0.22–2.75 mg L−1 in the training period and 0.74–3.54 mg L−1 in the validation period; while SMAPEs were in the range of 2.35–38.70% in the training period and 10.88–71.62% in the validation period. By providing insights into the oxygen-related processes, the findings could be valuable for future pond operation and monitoring.