A Transferable Ensemble Additive Network for Interpretable Prediction of Key Performance Indicators

In the wastewater treatment process, unsupervised domain adaptation (UDA) enables cross-condition prediction for key performance indicators. However, the lack of interpretability in predictions can compromise the reliability of the model. This work proposes a transferable ensemble additive network (TEAN) that is both capable of solving domain adaptation tasks and providing interpretable predictions. TEAN consists of an ensemble additive network (EAN) and a transferable additive network (TAN), which are essentially a transfer feature learning model and a domain adaptation model based on multikernel maximum mean discrepancy (MK-MMD), respectively. To improve the performance of the model in terms of domain adaptation, the EAN is pretrained to learn transfer features, and the latent feature dimensions in the TAN are augmented to better learn and capture interdomain discrepancies. To enhance interpretability, EAN conducts feature selection based on importance to obtain sparse feature representations. To avoid inconsistent selection results across multiple runs compromising the interpretability of the model, TEAN constructs weighted variance to measure the importance of features and applies an ensemble strategy in building EAN. Experimental results conducted on data generated from benchmark simulation model No. 1 (BSM1) demonstrate that TEAN outperforms the other comparison methods. TEAN achieves more consistent feature selection results under multiple runs and exhibits excellent prediction accuracy for key performance indicators, while its predictions are interpretable.