Convolutional Neural Networks (CNN)-Signal Processing Combination for Daily Runoff Forecasting

Aims: The main aim of this study was to assess the efficacy of two important signal processing approaches i.e., wavelet transform and ensemble empirical mode decomposition (EEMD) on the performance of the convolutional neural network (CNN). Materials Methods: The study was performed in two watersheds i.e., the Kasilian and Bar-Erieh Watersheds. In the first step, the CNN-based runoff modeling was done in its single form i.e., using the original data as input. In the next step, the input data was decomposed into several different sub-components i.e., approximation and details using Wavelet transform and Intrinsic Mode Functions (IMFs) using EEMD. Then the decomposed data were imported to the CNN model as input and combined Wavelet-CNN and EEMD-CNN models were provided. Findings: The results showed that CNN in its single form could not estimate the one-day-ahead runoff with acceptable accuracy. CNN in its original form had a moderate performance (with NRMSE of 83 and 66%). However, the application of Wavelet transform and EEMD in combination with CNN produced acceptable results. It was shown that Wavelet transform had a higher impact (with NRMSE of 48 and 26%) on the performance of CNN in comparison to EEMD (with NRMSE of 52 and 61%). Conclusion: This study showed that signal processing approaches can enhance the ability of deep learning methods such as CNN in predicting runoff values for one-day-ahead. However, the impact of signal processing methods on the performance of deep learning methods is not equal.