A novel hybrid approach based on outlier and error correction methods to predict river discharge using meteorological variables

Abstract

A new hybrid approach for the river discharge prediction is proposed by integrating the Hampel filter (HF) with an autoregressive distributed lag (ARDL) model and multi-model error correction method. This study applied the HF to detect and correct outliers present in the data. Then, the HF-treated data variables were employed in the ARDL model to obtain discharge predictions and errors were obtained. Next, a multi-model approach (named ASR) was used based on a combination of artificial neural networks (ANN), support vector machines (SVM), and random forest (RF) models to predict errors. The ASR-predicted errors were aggregated with HF-ARDL prediction to determine the final HF-ARDL-ASR hybrid model predictions. The effectiveness of this approach was explored and compared with different models on the discharge data of four rivers of the Indus River basin of Pakistan. The root mean squared error (RMSE) of the HF-ARDL-ASR hybrid model in Jhelum River (Domel station) is 96.88 m³/s in the testing phase that is 53.92%, 50.0%, 48.7%, 50.0%, 13.4%, 53.2%, 50.3%, 46.4%, and 49.1% lower than the RMSE of the multiple linear regression (MLR), SVM, ANN, RF, ARDL, HF-MLR, HF-SVM, HF-ANN, and HF-RF models respectively. On test data, the Nash–Sutcliffe Efficiency (NSE) values of the suggested HF-ARDL-ASR hybrid model in Jhelum River (Chattar Kallas station) is 0.8571, Jhelum River (Domel) is 0.8294, Kabul River (Nowshera) is 0.8291 and Kunhar River (Talhata) is 0.8506. Therefore, the proposed HF-ARDL-ASR model has shown superior performance, lower errors, and higher prediction accuracy than all comparative models in the study.