Author(s): Piciaccia Luca; Croce Danilo; Basili Roberto; Pettersen Jonas; Ryfors Pia
Linked Author(s):
Keywords: No Keywords
Abstract: VEAS is the largest WWTP in Norway, where inflow is collected through a combined sewer system, i. e., storm water runoff is combined in a common conduit with wastewater from homes, businesses, and industry and delivered to the plant. From a process perspective this already high degree of variability is further compounded by return flows from the plant itself. The VEAS plant is fully located in cavern and is operated 24/7. Cavern location requires low footprint and consequently high surface load. The VEAS process features a “single-shot” sedimentation and has a record-low water retention time of 3 hours. This highly efficient configuration is sensitive to variation in the inflow water parameters and internal plant recirculation flows, 25 measured parameters have been identified as impacting the effectiveness of the sedimentation process. Due to the high non-linearity of the parameters influence, even extensive use of classic non-linear statistical analysis has failed to clearly identify the main performance drivers of the process. In this paper we investigate the use of Kernel-based and Neural methods for the learning of the optimal control parameters in the context of industrial plants. The main objective is to define an automatic way to identify and tune the most relevant parameters of the plant (e. g., dosage of chemicals, sump level setting) to minimize the final water turbidity. The adopted machine learning framework enables the automatic analysis of the evolution of the plant behavior over time, i. e. exploits sensors readings stored for a long time period (one year), to develop a predictive model of the future behavior of the system.
Year: 2018