A novel coupling preprocessing approach for handling missing data in water quality prediction

Abstract

Predicting water quality accurately is of great significance in water resource management. Data-driven models (e.g., neural networks) perform effectively in this task. High quality original datasets are necessary to build higher accuracy prediction models. As missing data commonly exists, this paper presents a novel piecewise multivariate imputation (PWIMP) method for dealing with missing data in water quality series. To obtain a prediction model with higher accuracy, wavelet shrinkage denoising based on the maximal overlap discrete wavelet transform (MODWT) is used to reduce the noise of the original data after imputation. Four datasets, which were generated by global multivariate imputation (GLIMP), PWIMP, coupling GLIMP and MODWT, and coupling PWIMP and MODWT, are evaluated by the outcome of the long short-term memory (LSTM) neural networks in terms of six statistical indices. The results show that the effect of PWIMP is significantly better than that of GLIMP and is competitive with the method of coupling GLIMP and MODWT. The method of coupling PWIMP and MODWT significantly outperforms the other three models.