Development of a TVF-EMD-based multi-decomposition technique integrated with Encoder-Decoder-Bidirectional-LSTM for monthly rainfall forecasting

Accurate forecasting of rainfall is extremely important due to its complex nature and enormous impacts on hydrology, floods, droughts, agriculture, and monitoring of pollutant concentration levels. In this study, a new multi-decomposition deep learning-based technique was proposed to forecast monthly rainfall in Himalayan region of India (i.e., Haridwar and Nainital). In the first stage, the original rainfall signals as the individual accessible datasets were decomposed into intrinsic mode decomposition functions (IMFs) through the time-varying filter-based empirical mode decomposition (TVF-EMD) technique, and then the significant lagged values were computed from the decomposed sub-sequences (i.e., IMFs) using the partial autocorrelation function (PACF). In the second stage, the PACF-based decomposed IMFs signals were again decomposed by the Singular Valued Decomposition (SVD) approach to reduce the dimensionality and enhance the forecasting accuracy. The machine learning approaches including the bidirectional long-short term memory reinforced with the Encoder-Decoder Bidirectional (EDBi-LSTM), Adaptive Boosting Regression (Adaboost), Generalized Regression Neural Network (GRNN), and Random Forest (RF) were used to construct the hybrid forecasting models. Also, several statistical metrics i.e., correlation coefficient (R), root mean square error (RMSE), and Nash-Sutcliffe efficiency (NSE) and graphical interpretation tools were employed to evaluate the hybrid (TVF-EMD-SVD-RF, TVF-EMD-SVD-EDBi-LSTM, TVF-EMD-SVD-Adaboost, and TVF-EMD-SVD-GRNN) and standalone counterpart (EDBi-LSTM, Adaboost, RF, and GRNN) models. The outcomes of monthly rainfall forecasting ascertain that the TVF-EMD-SVD-EDBi-LSTM in the Haridwar (R = 0.5870, RMSE = 118.4782 mm, and NSE = 0.3116) and Nainital (R = 0.9698, RMSE = 44.3963 mm, NSE = 0.9388) outperformed the benchmarking models.