A machine learning framework to improve effluent quality control in wastewater treatment plants

Area of Science:

• Environmental Engineering
• Wastewater Treatment
• Machine Learning Applications

Background:

• Wastewater treatment plant (WWTP) processes are complex, making real-time effluent quality control challenging and costly.
• Conventional mechanistic models have limitations; Machine Learning (ML) offers an alternative for process modeling.
• Existing ML applications often overlook operational factors and time lags, hindering understanding of effluent quality drivers.

Purpose of the Study:

• To present a novel ML-based framework for enhancing WWTP effluent quality control.
• To clarify the relationships between operational variables and effluent parameters, including time-lag effects.
• To identify key operational factors influencing effluent Total Suspended Solids (TSS_e) and Phosphate (PO4_e).

Main Methods:

• Development of a framework integrating Random Forest (RF) and Deep Neural Network (DNN) models.
• Application of Variable Importance Measure (VIM) and Partial Dependence Plot (PDP) analyses.
• Incorporation of a novel approach to account for time lags between WWTP process steps.

Main Results:

• Influent temperature identified as the most influential variable for both TSS_e and PO4_e, with distinct effects.
• Aeration basin Total Suspended Solids (TSS) strongly influence PO4_e; higher TSS generally aids removal, but excess is detrimental.
• The impact of aeration basin TSS on effluent quality increases with basin distance from the outlet; excessive return sludge negatively affects TSS_e.

Conclusions:

• The proposed ML framework effectively clarifies operational factor impacts on WWTP effluent quality, considering time lags.
• Findings provide actionable insights for optimizing control strategies at the Umeå WWTP and similar facilities to reduce costs.
• The framework's applicability extends to other WWTP parameters and industrial processes with sufficient high-resolution data.