Development of a Soft Sensor Using Machine Learning Algorithms for Predicting the Water Quality of an Onsite...

Area of Science:

• Environmental Engineering
• Water Quality Monitoring
• Wastewater Treatment

Background:

• Onsite wastewater treatment systems (OWTS) require continuous monitoring for efficiency and compliance.
• Traditional off-line water quality analyses (COD, TSS, E. coli) are time-consuming and lack real-time data.
• Existing real-time COD analyzers are often too costly for OWTS.

Purpose of the Study:

• To design and implement a real-time remote monitoring system for OWTS.
• To develop multi-input, single-output soft sensors predicting key water quality parameters (COD, TSS, E. coli).
• To evaluate machine learning algorithms for predicting off-line parameters using in-line sensor data.

Main Methods:

• Developed soft sensors integrating data from in-line sensors (turbidity, color, pH, NH4+, NO3-, electrical conductivity).
• Utilized temporal and spatial water quality data from a field-tested OWTS (n=56).
• Evaluated four machine learning algorithms: PLS, SVR, CUB, and QRNN for prediction accuracy.

Main Results:

• Support Vector Regression (SVR) accurately predicted COD (MAPE 14.5%, R^2 0.96).
• Cubist regression (CUB) optimally predicted TSS (MAPE 24.8%, R^2 0.99).
• E. coli prediction was less accurate (CUB: MAPE 71.4%, R^2 0.22), requiring further development.

Conclusions:

• Machine learning-based soft sensors can provide valuable real-time estimates of COD and TSS in OWTS.
• Soft sensors offer a cost-effective alternative for real-time monitoring, enabling operational adjustments.
• Further research is needed to improve the accuracy of E. coli prediction in OWTS using soft sensors.