Hybrid Water Quality Prediction with Frequency Domain Conversion Enhancement and Seasonal Decomposition

Water quality prediction can accurately reflect the development trend of water quality, and it is an important means to prevent the water environment from being polluted and maintain the health of the water environment. Existing prediction methods generally cannot accurately capture non-linear characteristics of water quality, and suffer from issues of gradient disappearance and gradient explosion. This work designs a water quality prediction model called SMF2 to effectively solve these problems and increase the accuracy of prediction. SMF2 combines the Savitsky-Golay filter, seasonal-trend decomposition using loess for multiple seasonal components, Fourier transform frequency-enhanced block and frequency-enhanced attention, serving for noise smoothing, extraction of exact seasonal components, time domain-frequency domain interconversion, feature extraction, and time series prediction by frequency domain low-rank approximation transform, respectively. Experimental results based on a real-life water environment data set show that the proposed SMF2 outperforms other advanced algorithms in terms of prediction accuracy.