Water Quality Prediction Based on SSA-MIC-SMBO-ESN

Abstract

Water pollution threatens the safety of human production and life. To quickly respond to water pollution, it is important for water management staff to predict water quality changes in advance. Drawing on the temporality of water quality data, the leaky integrator echo state network (ESN) was introduced to construct the water quality prediction models for dissolved oxygen (DO), permanganate index (CODMn), and total phosphorus (TP), respectively. First, the missing values were filled by the linear trend method of adjacent points, and the outliers were detected and corrected by the Z-score method and the linear trend method. Second, the singular spectrum analysis (SSA) was performed to denoise the original monitoring data, such that the predicted data catch up with the real data, and the model accuracy is not affected by the hidden noise in the data. Third, the correlation between water quality indices was measured by the maximum information coefficient (MIC), and the strongly correlated indices were imported to the prediction model. Finally, according to these strong correlation indicators, the water quality prediction models based on multiple features were constructed, respectively, using the offline and online learning algorithms of the ESN. The hyperparameters of the models were optimized through the sequential model-based optimization (SMBO). Experimental results show that the proposed water quality prediction models, namely, SSA-MIC-SMBO-Offline ESN and SSA-MIC-SMBO-Online ESN, predicted DO, CODMn, and TP accurately, providing suitable tools for practical applications.