Precise
control of biological wastewater treatment for nitrogen
removal is difficult because of the nonlinearity, time-varying, and
time-consuming nature of the process. With due emphasis on addressing
the challenges involved in its effective implementation, this study
developed an artificial neural network (ANN) based soft sensor (SS)
with a set of proposed thumb rules for online forecasting of the concentrations
of hard-to-measure parameters (NH4
+ and NO2
−) from secondary easy-to-measure variables,
(reactor volume, dissolved oxygen, suspended solids, pH, temperature,
and ORP) in an Anammox based pilot plant. Four hybrid neural networks
(PCA-Kalman NN, PCA NN, Kalman NN, and Non NN) were applied to identify
net optimum input vectors for the SS, using an appropriate quantity
of samples from the set of secondary variables. The proposed hybrid
SS was tested on a sewage wastewater treatment plant operated using
a Matlab R2018a framework and validated using operational plant data.
The results showed that the PCA-Kalman neural network with R
2 values of 0.9985 and 0.9263 for NH4
+ and NO2
–, respectively,
is potentially a valuable tool for plant operators in the selection
of operational states to minimize cost and to efficiently predict
important parameters that are prone to errors due to a failure in
online sensors.