A Two-Stage Multi-Target Domain Adaptation Framework for Prediction of Key Performance Indicators Based on Adversarial Network

In wastewater treatment processes, building performance evaluation models to predict key indicators under uncommon operating conditions is difficult due to the lack of labeled data. Domain adaptation can be used to solve this problem through leveraging the knowledge of common conditions to construct prediction models for uncommon conditions. Considering the costs of labeling data, it is reasonable to assume that only data from the most common condition are labeled. Therefore, all domain adaptation tasks share a source domain. Under this assumption, most domain adaptation methods require mapping the same source data multiple times and training multiple task-specific predictors in different tasks, resulting in additional computational costs. To give a solution, a stepwise domain alignment strategy is proposed, which consists of two steps. First, the latent features of source domain are extracted, and the features are fixed after this step. Second, target domains from different tasks are mapped to the fixed feature space to achieve domain alignment. Based on the strategy, a two-stage multi-target adversarial adaptation network (TS-MAAN) for predicting effluent quality index is proposed, which consists of an autoencoder and a generative adversarial network. Additionally, parameter initialization and multi-kernel maximum mean discrepancy optimization are further proposed to improve the stability and prediction accuracy of the TS-MAAN, respectively. Experiments conducted on datasets generated by the Benchmark Simulation Model No.1 demonstrate that TS-MAAN exhibits excellent prediction accuracy and stability, while enabling efficient multi-target domain adaptation. Moreover, these experiments verify the effectiveness of parameter initialization and MK-MMD optimization.