Forecasting Operational Conditions: A case-study from dewatering of biomass at an industrial wastewater treatment plant

In this paper, we present a data-driven approach to predicting polymer dosages for industrial decanters based on upstream production data. First, a data extraction algorithm using on-line sensors is developed to identify when the operational mode is changed with a 99 % accuracy. Next, an investigation of process delays in the collected data is carried out by analysing partial autocorrelation matrix eigenvalues upon which is it concluded to transform the data by summarising the data by batch and including lagged summaries to account for a time delay of 2 hours. Finally, a random forest forecasting model is trained capable of learning structured information from the lagged summaries producing decent predictions for both low and high polymer dosages (RMSE 14.89). The proposed approach could potentially save operators 3-6 hours a day.