A data analytics pipeline to optimize polymer dose strategy in a semi-continuous multi-feed dewatering system

Automated application of scientific data analytics is a critical element to enhance industrial productivity and minimize human errors. In this study a dynamic polymer dosing strategy is developed for an industrial centrifuge system receiving highly variable rates and quality of feed solids from a wastewater treatment plant. A four-section methodology is developed and tested at full-scale containing (i) data extraction for data collection and aggregation; (ii) Data wrangling including delay analysis and batch analysis; (iii) model development with predictive ability; (iv) model analysis for model evaluation and interpretation. A partial least squares and a random forest model were validated and used to predict polymer dosages. In contrast to PLS, the RF model was capable of learning structural information and describe which products related to increases or decreases in polymer dosage. An additional analysis investigating the impact of different product codes on the polymer dosage is presented, revealing that certain products generally lead to consistent changes in the polymer dosage. The proposed approach could potentially save operators 3–6 h a day in terms of time spent on manually adjusting polymer dosages. The presented methodology for data pipelining and analysis has a generic nature and easily exportable to other case studies.