A novel hybrid BPNN model based on adaptive evolutionary Artificial Bee Colony Algorithm for water quality index prediction

Abstract

With the accelerated industrialization and urbanization process, water pollution in rivers is being increasingly worsened, and has caused a series of ecological and environmental issues. The prediction of river water quality index (WQI) is a prerequisite for river pollution prevention and management. However, the water quality data series is non-smooth and non-linear, and a strong coupling relationship between different water quality parameters that influence each other is observed, making it an inevitable problem to accurately predict water quality parameters. To this end, a combination of machine learning and intelligent optimization algorithms was hereby used to break this dilemma. Specifically, a Back Propagation Neural Network (BPNN) model was established using the Artificial Bee Colony (ABC) algorithm, with the three adaptive evolutionary strategies, i.e., dynamic adaptive factors, probability selection and gradient initialization combined to form the Adaptive Evolutionary Artificial Bee Colony (AEABC) algorithm. The experimental results of this algorithm demonstrate that the AEABC-BPNN model only requires 14 iterations to converge in this case. The predictions of WQI can reduce the error evaluation indicators of mean square error (MSE) to 0.2745, which is at least 25.2% lower than those of the rest algorithms compared, and the mean absolute percentage error (MAPE) is lower than 7.58%. In four WQIs, the prediction interval coverage percentage (PICP) reaches 100%. Besides, robustness testing experiments were also designed to verify that the AEABC-BPNN model still outperforms the rest of the algorithms in terms of prediction accuracy when guided by historical error data. The proposed model plays a pivotal role in water pollution management in rivers and lakes, and has scientific significance for future water environmental protection.