Deep learning based soft sensor for microbial wastewater treatment efficiency prediction

Abstract

With the swift industrial development, the pollution of water bodies by industrial effluent is becoming more widespread and serious. To achieve a more effective and intelligent control strategy to deal with the risk of standard-exceeding of water quality, it is crucial to develop a soft sensor to meet the demand of water quality prediction of IETP. Therefore, this paper proposes a CEEMD-ReliefF-CNNGA soft sensor for water quality prediction. The soft sensor firstly combines CEEMD and ReliefF algorithms to optimize the feature structures. CEEMD decomposes mode mixing between signals efficiently, and ReliefF intensifies the weights of signals with positive effects adaptively. Then, the soft sensor selects feature combinations as the input of CNNGA model. CNNGA calibrates the signal features most relevant for the predicted target by means of multi-layer iterative convolutions, embeds the pooling layers to prevent overfitting, and introduces Bidirection-GRU to adaptively capture the bidirectional dependencies of different temporal scales in the signal features. To avoid Bidirection-GRU holds the same attention on all signal features, an attention mechanism is introduced to strengthen time-dependent influences of key information, so as to improve the soft sensor performance effectively. To evaluate the soft sensor performance, the experiment adapts microbial wastewater data from IETP to forecast the concentrations of COD, NH3-N, TN, and TP in the real effluent. Results show the proposed soft sensor performs better than the other advanced models, and the soft sensor has a certain validity and stability in the actual prediction requirements of IETP.