Multi-Step Water Quality Prediction with Series Decomposition and Auto-Correlation

Water quality prediction provides timely management to solve possible water environmental problems, which is of great importance. However, the following challenges exist: 1) The existence of noise in the water quality time series can lead to overfitting of nonlinear models; 2) It is difficult to capture temporal dependencies in complex time series data; 3) Long-term forecasting is difficult to achieve. To address the above difficulties, this work proposes a multi-step water quality prediction model, called SG-Autoformer, which combines the Savitzky-Golay filter, the inner series decomposition, and an auto-correlation mechanism. First, SG-Autoformer performs noise reduction on the water quality time series to suppress overfitting of nonlinear models. Second, it embeds series decomposition inside the encoder and decoder, which obtains more predictable components from complex time series for long-term prediction. Third, SG-Autoformer utilizes the auto-correlation mechanism to capture the time dependence and improve information utilization. Extensive experiments with real-world datasets show that SG-Autoformer outperforms other advanced prediction methods in terms of prediction accuracy.