(328f) Design of a Supervisory Control System for Inter-Networked Facilities Operation

Operation of highly networked systems with interdependent influent streams can result in suboptimal network performance, where individual system operability and optimization are the primary concern of plant operators and engineers. The authors have designed a supervisory control system for operation of county-wide facilities transport through utilization of machine learning and connectionist system modeling for the prediction of unsteady state influent behavior. Black box neural network models were generated from available multi-year data collected by individual plant SCADA systems, and a supervisory control system was designed to manage both influent routing for inter-networked facilities as well as individual plant-based operation.

Time-series artificial neural network models were generated for each of the seven advanced wastewater treatment facilities in Hillsborough County, Florida, where each facility receives influent flow from a series of centralized pumping stations that are engineered to divert flow via opening and closing of flow gates from one facility to another during instances of abnormally high influent flow rates. Due to current gate operation, there is necessarily a large amount of flow variance at any one instance, which leads to significant operational challenges at the plant level to manage and treat that supply in real time. Sudden changes in influent—coupled with long residence times for the wastewater treatment process—can be detrimental to meeting EPA regulations, since the primary control strategy involves manipulating dissolved oxygen concentration for the nitrification-denitrification reaction via mechanical aeration (a very slow process). It is often necessary to know several hours in advance of any significant changes to plant operation in order to affect a change that will keep the plant effluent in compliance or prevent expensive fines due to excess wastage to retention ponds.

As a result of the pumping stations and gate operation, influent data reported by the SCADA system is generally fraught with noise; influent flow rates vary by ±2 million gallons per day (MGD) within a 5-minute span. In order to create any meaningful models from available data, it is necessary to pre-process the data via profile smoothing, data cleaning, feature extraction, and feature generation. Then, a closed-loop time-series neural net can be trained from a subset of existing data to predict future events—often up to a day in advance. In combination with margins of error calculations at each time step, the predictive capability of the neural network can account for non-standard operational behavior (e.g., holidays) and anomalous activity (e.g., major sporting events or illegal industrial dumping). Machine learning aspects of the project allow neural networks to be retrained over time to account for gradual changes in plant performance and shifts in population behavior.

Finally, information gained from neural net models was abstracted to design an energy efficient supervisory control system for operators at the individual plant level and engineers in charge of flow diversion for the entire facilities network. The information was conveyed to operators in the form of SCADA-interfaced GUI with real-time prediction and rules-based control strategies for engineers and operators for optimizing network-level and plant-level performance.