Forecasting Daily Flood Water Level Using Hybrid Advanced Machine Learning Based Time-Varying Filtered Empirical Mode Decomposition Approach

Accurate water level forecasting is important to understand and provide an early warning of flood risk and discharge. It is also crucial for many plants and animal species that needs specific ranges of water level. This research focused on long term multi-step ahead forecasting of daily flood water level in duration of (2005–2021) at two stations (i.e., Baryulgil and Lilydale) of the Clarence River, in Australia, introducing a novel hybrid framework coupling time varying filter-based empirical mode decomposition (TVF-EMD), classification and regression trees (CART) feature selection, and four advanced machine learning (ML) models. The implemented ML approaches are including Long-Short Term Memory (LSTM), cascaded forward neural network (CFNN), gradient boosting decision tree (GBDT), and multivariate adaptive regression spline (MARS). Here, original time series of WL in each station was decomposed into the optimal intrinsic mode functions (IMFs) using the TVF-EMD technique and the significant lagged-time components for two desired horizons (t + 1 and t + 7 time ahead) in each station was extracted by using the CART-feature selection method. Then, the IMFs and corresponded residual obtained from the pre-processing procedure were separately implemented to feed the ML models and produce the C_ART-T_VF-EMD-L_STM, C_ART-T_VF-EMD-C_FNN, C_ART-T_VF-EMD-M_ARS, and C_ART-T_VF-EMD-G_BDT by assembling all the individual sub-sequences outcomes. Several goodness-of-fit metrics such as correlation coefficient (R), Mean absolute percentage error (MAPE), and Kling-Gupta efficiency (KGE) and the infographic tools and diagnostic analysis were employed to evaluate the robustness of the provided techniques. The outcomes of developed expert systems ascertained that C_ART-T_VF-EMD-C_FNN for one- and seven-day horizons in both stations outperformed the C_ART-T_VF-EMD-M_ARS, C_ART-T_VF-EMD-L_STM, C_ART-T_VF-EMD-G_BDT, and all the standalone counterpart models (i.e., CFNN, MARS, LSTM, and GBDT) respectively. As one of the most important achievements of this research, the LSTM did not lead to superior and promising results in the long-term highly nonstationary time series.