Predictive AI Models in WTP Operations

Predictive AI Models in WTP Operations

Overview

CWT was engaged to develop an artificial intelligence model to predict and optimise coagulant dosing at a NSW water treatment plant. Using historical water quality and dosing records, the project focused on improving dosing accuracy and embedding the tool into the operator’s existing workflow for practical, on-site use. 

For more information, check out our detailed Case Study.

Services Offered

End-to-end model delivery

  • data review and cleaning
  • machine learning model development
  • validation against historical performance
  • deployment through a low-cost cloud hosting

Need

Coagulant dosing varies with changing raw water conditions, and manual dosing decisions can lead to inconsistent outcomes and unnecessary chemical use. The plant required a reliable prediction tool that could use routine water quality inputs (e.g., UV254, turbidity, alkalinity and other parameters) and fit seamlessly into existing operational practices. 

Solution

CWT delivered a four-step approach: preparing and cleaning the training dataset, training a Random Forest Regression model, validating predictions through residual analysis, and deploying the model via a Microsoft Azure serverless environment. 

The final solution was integrated into the operator’s Excel workbook (via Power Query/API), returning real-time dosing predictions as operators update water quality parameters. 

Benefit

The model achieved strong predictive performance (R² of 0.883 and RMSE of ~1.59 mg/L), supporting more consistent dosing decisions within typical operating ranges. Serverless hosting provided a highly cost-effective, scalable deployment method (with fast response times and minimal ongoing cost) while also enabling future retraining and version control as plant conditions and datasets evolve.