PREDICTIVE MODEL OF 2-MIB AND GEOSMIN CONCENTRATIONS IN A RIVER WATER BY FEED FORWARD ARTIFICIAL NEURAL NETWORK AND LONG SHORT-TERM MEMORY NETWORK INPUTTING HOURLY WATER QUALITY DATA

Abstract

 Recently, 2-methylisoborneol (2-MIB), which are mold odorants, has been confirmed in the Ara River, and powdered activated carbon has been added for their reduction. However, measurement of the concentration of the odor is difficult, and in many water purification plants, powdered activated carbon has been added empirically.

 In this study, we developed models for short-term predictions of odor concentration by FFANN (Feed Forward Artificial Neural Network) and LSTM (Long Short-term Memory) using the big data held by water purification plants. In FFANN, the odor concentration could be estimated from the basic water quality items at the current time. In addition, in LSTM, a short-term estimate of the concentration of the odor after 3 hours was obtained from the water quality including the concentration of the odor until the current time.

 This suggests that powdered activated carbon can be controlled by inputting the odorant concentrations obtained from FFANN into LSTM.