Forecasting salinity time series using RF and ELM approaches coupled with decomposition techniques

Abstract

Using historical salinity data from nine drought periods in the Pearl River Delta of China, this study utilized two machine learning approaches to forecast the salinity time series for multistep lead times: random forest (RF) models and extreme learning machine (ELM) models. To improve conventional RF and ELM models, three signal decomposition techniques were applied to preprocess the input time series: empirical mode decomposition (EMD), wavelet decomposition (WD) and wavelet packet decomposition (WPD). The study results indicated that in contrast to conventional RF/ELM, a hybrid RF/ELM method accompanied by decomposition techniques displayed better forecasting performance and yielded reasonably accurate prediction results. More specifically, hybrid models coupled with WPD displayed the best performance for all three forecast lead times of one, three and five days, whereas EMD underperformed both WPD and WD because of the limited predictability of the components. Both the WPD and WD hybrid models using the $$coif5$$ wavelet basis performed better than those using the other two bases (db8 and sym8). In addition, ELM method performed better for conventional and WD/WPD hybrid models, whereas the RF method worked better for EMD hybrid model. The findings of the study showed that the nonstationary salinity series could be transformed into several relatively stationary components in the decomposition process, which provided more accurate salinity forecasts. The developed hybrid models coupling RF/ELM method with decomposition techniques could be a feasible way for salinity prediction.