Multi-objective Pigeon-inspired Optimized feature enhancement soft-sensing model of Wastewater Treatment Process

Abstract

Under the increasingly severe fresh water supply pressure, wastewater treatment is considered to be the optimal strategy to satisfy the current and future water demand, thus being highly valued by most countries. As a complicated process, there are some hard-to-measure effluent indicators in wastewater treatment such as 5-day Biological Oxygen Demand (BOD5), which brings significant difficulties to the monitoring of key indicators in sewage disposal process, thus imposing massive constraints on evaluation of effluent quality. In order to realize the real time supervision of the water quality, data-driven artificial intelligence soft-sensing models have been widely considered as an active research field. However, the over-parameterization of artificial intelligence methods and the complication of sewage treatment environment lead to obvious non-Gaussian characteristics in the wastewater data, which makes it difficult to artificially set parameters to maintain the optimal values to satisfy the accuracy requirement in wastewater treatment process. In view of the aforementioned problems, a soft-sensing model for super-parameter intelligent setting of Broad Learning System based on Overcomplete Independent Component Analysis (OICA) is proposed in this paper. A two-stage multi-objective optimization algorithm is adopted in the model to set superparameters intelligently, which reduces human intervention and improves the accuracy of the model. Additionally, the adaptability of the proposed model to data is significantly enhanced through improvement of the feature extraction ability of the Broad Learning System (BLS) and the capture of peculiar non-Gaussianity in wastewater data with statistical methods. Comparative experiments are conducted on the sewage simulation platform BSM1 with state-of-the-art artificial intelligence soft measurement models, and the results show that the advantage of the presented model lies both in accuracy and in modeling speed, demonstrating the effectiveness of the proposed method.