Abstract
Accurate forecasting of key hydrological processes, such as rainfall, generally requires the use of auxiliary predictive hydrological variables. Data requirements can be reduced by using artificial intelligence models that are able to successfully capture the information contained in the historic observations of the target variable of interest. In this study, a novel Wavelet-coupled Multi-order Time Lagged Neural Network (WMTLNN) model is developed to accurately forecast rainfall by using previous rainfall records only. The study is conducted using daily rainfall data recorded in the period 2015–2017 at three meteorological stations (Astore, Chillas, and Gilgit) located in Upper Indus Basin (UIB), Pakistan. WMTLNN models are developed by introducing time lags up to ten days, Symlets 4 (sym4) wavelets and Daubechies wavelets with three vanishing moments (db3), and Maximal Overlap Discrete Wavelet Transformation (MODWT) to account for boundary effects in the forecasting mode. The performance of WMTLNN models is compared with that of Time Lagged Neural Network (TLNN) models, Wavelet-coupled Time Lagged Neural Network (WTLNN), and deep learning Long Short-Term Memory (LSTM) models. Comparative analysis indicates that WMTLNN models overcome the other models, with more than 80% forecasting accuracy for most of the cases, and a typical range of 0.85–0.95 accuracy in terms of NSE. The highest NSE value is 0.97 at Astore with LSTM model, 0.96 at Chillas with WMTLNN model, and 0.95 at Gilgit with WMTLNN model. Overall, wavelet transformation of time series data enhances efficiency and accuracy of rainfall forecast.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have