In wastewater treatment process (WWTP)

existence of several important but difficult-to-measure process variables hinders not only monitoring of the production process but also adjustment or optimization of process control strategies. Even though soft-sensor models are reasonably constructed

it will still suffer degradation problem and high maintenance cost in real-time operation. Additionally

selection of proper secondary variables directly affects subsequent modeling. Therefore

a self-adaptive multi-output soft sensor model based on deep neural network was proposed for simultaneous online prediction of multiple target variables in wastewater treatment. Deep neural network was constructed from a stacked auto-encoder

which had satisfactory performance of online prediction under extremely complex scenarios. In order to overcome degradation problem and select proper secondary variables

time difference modeling and VIP (variable importance in projection) methods were added. Finally

validation on a true WWTP process shows that the proposed soft-sensor model has good performance on multiple output prediction and satisfactory prediction on single target.