A multi-output hybrid prediction model for key indicators of wastewater treatment processes

Abstract

The fluctuating working conditions in wastewater treatment processes, influenced by various factors, result in highly nonlinear characteristics in online monitoring data. This presents challenges for accurately estimating water quality. Addressing the issue of single-model performance degradation under changing data distributions, this paper proposes a two-stage hybrid prediction scheme based on clustering. Firstly, historical data is divided and features are extracted and clustered based on different time periods. Subsequently, distinct prediction models are applied to data within each working mode, thereby enhancing overall prediction performance. The selection and combination of two classical models with different characteristics, namely the partial least squares random weight neural network (PLS-RWNN) and the multi-output correlation vector machine (MRVM), enable better adaptation to the complex wastewater treatment data source. The proposed approach is validated using the wastewater treatment platform BSM2. On average, clustering modeling combined with models provides better predictions for all three variables. The comprehensive index RMSSD of the mixed model is 0.6189, which is 42.17 % higher than that of a single model used before clustering. Results indicate that the proposed network architecture significantly improves prediction performance, highlighting its effectiveness in dealing with the nonlinear and fluctuating nature of wastewater treatment data.