Using machine learning techniques for influent flow forecasting at water resource reclamation facilities

Machine Learning (ML) techniques can be used to algorithmically predict states of complex systems where mechanistic approaches are not available or are impractical. This paper explores the use of the Random Forest (RF) technique to make short-term predictions of future influent flow to a wastewater treatment plant. Large datasets of influent flow at high temporal resolution (one data point every 5 minutes) are used to train the algorithm. The system uses a set of inputs (e.g. temperature, precipitation, month, time-of-day, etc.) to predict influent flow for several days into the future. These inputs can then be used as inputs for a mechanistic models of a wastewater treatment systems, allowing for future predictions of plant performance. The Random Forest method was applied to a dataset from a Southern Ontario wastewater treatment plant (WWTP). The algorithm was trained using a one-year dataset with hourly datapoints, and successfully reproduced typical influent flow patterns for different time periods.