Comparison of LSTM network, neural network and support vector regression coupled with wavelet decomposition for drought forecasting in the western area of the DPRK

Drought forecasting is very important in reducing the drought damage and optimizing water resources. This paper focuses on confirming the advantage of wavelet long short-term memory network (WLSTMN) through comparison with wavelet artificial neural network (WANN) and wavelet support vector regression (WSVR) for drought forecasting in the west area of the Democratic People’s Republic of Korea. The standardized precipitation index with 6 and 12-month timescales (SPI-6 and SPI-12) was used in this study. In order to increase the number of training samples for the development of data-driven models, SPIs were calculated at ten days’ intervals and input data was lagged combinations of time series that decomposed using Haar wavelet mother function at 1–10 decomposition levels. The performances of the three models with several decomposition levels and lags at 1-month lead time were estimated with determination coefficient (R2), Lin's concordance correlation coefficient (LCCC), root-mean-square error (RMSE) and mean absolute error (MAE). Area-averaged performance measures of optimal models show that R2, LCCC, RMSE and MAE of WLSTMN for SPI-6 were 0.709, 0.806, 0.572 and 0.427, respectively, better than those of other models. And R2, LCCC, RMSE and MAE of WLSTMN for SPI-12 were 0.919, 0.950, 0.296 and 0.190, respectively. It has a better performance compared to the other models. Consequently, WLSTMN model for drought indices with two timescales outperformed traditional WANN and WSVR, which have smaller R2 and LCCC, larger RMSE and MAE.