Deep Learning Optimization for Soft Sensing of Hard-to-Measure Wastewater Key Variables

Soft sensors can be an essential part of a digital twin to acquire critical wastewater information for operation optimization. Soft sensor predictions have been successfully applied in nitrogen compounds, but hard-to-measure variables such as biochemical oxygen demand (BOD) and total suspended solids (TSS) have been a major challenge partially due to difficulty in capturing complex nonlinearity and needed information acquisition. This study pinpointed the bottlenecks by developing advanced hyperparameter optimized (HPO) deep learning (DL) models and testing different groups of data. By comparing two DL algorithms [multilayer perceptron and deep belief network (DBN)] with three HPO methods (genetic algorithm, particle swarm optimization (PSO), and grey wolf optimization), we found that DBN-PSO showed performance superior to other hybrid methods for both CBOD5 and TSS predictions based on 11 years of operational data. While the hybrid models exhibit complex topography, better results can be achieved with a slow learning process and a combination of aggressive pre-training and smooth fine-tuning for CBOD5 and TSS, respectively. Additional precipitation data did not provide additional benefits, whereas metal concentration data helped further improve the prediction accuracy (testing error index: 1.9 mg CBOD5/L and 1.5 mg TSS/L), suggesting that more diverse data acquisition is valuable for a better soft-sensor practice.