Real-time rainfall and runoff prediction by integrating BC-MODWT and automatically-tuned DNNs: Comparing different deep learning models

The development of reliable rainfall and runoff prediction models holds significant importance in the domains of flood forecasting, early warning systems, and sustainable water resources planning and management. This research successfully enhances the accuracy of rainfall and runoff predictions by integrating the BC-MODWT (boundary-corrected maximal overlap discrete wavelet transform) preprocessing technique with various univariate and multivariate automatically tuned DNNs (deep neural networks). To do so, this research utilizes distinct Daubechies mother wavelets, namely db1, db2, and db3, at different levels of decomposition, to enhance the accuracy of rainfall and runoff prediction in an urban catchment with low time of concentration. The aforementioned framework is applied to the EDC (East Drainage Catchment) of Tehran city. Random search is used as an automatic hyperparameter tuning technique for univariate and multivariate DNNs. The results illustrate that the utilization of the BC-MODWT technique along with the automatically-tuned DNNs significantly improves the prediction performance compared to the automatically-tuned DNNs (i.e., increases NSE values from 0.54 to 0.97). Furthermore, the performance of top automatically-tuned BC-MODWT-DNNs is compared in terms of their accuracy in predicting rainfall hyetograph and peak flow. Therefore, it can be concluded that the automatically-tuned BC-MODWT-DNNs, especially univariate ConvLSTM and CNN-Bi-LSTM integrated with BC-MODWT, can be effectively used for rainfall and runoff prediction in urban areas with low time of concentration.