Mathivha FHUMULANI, University of Zululand, South Africa
Mbatha NKANYISO, University of Zululand, South Africa
Sigauke CASTON, University of Venda, South Africa
Chikoore HECTOR, North West University, South Africa
Drought forecasting play a significant role in drought early warning system for drought risk assessment and management. Due to increasing spatial and temporal variability of precipitation and the reported increasing frequency and magnitude of hydrological extreme, forecasting of these extremes have never been more significant. More so in developing nation such as those located in sub-Saharan Africa. The aim of this study is to assess and forecast drought conditions in Luvuvhu River Catchment in north eastern South Africa using machine learning neural network models. The Standardised Precipitation Evaporation Index (SPEI) was used as a drought quantifying parameter at three timescales (i.e. 1-, 6- and 12- months). Long Short-Term Memory (LSTM) artificial recurrent neural network (RNN) architecture was used to forecast the SPEI time series. The study further utilised Quantile Regression Averaging (QRA) for forecast combination. Prediction intervals were used for uncertainty analysis of the forecasts. An undecomposed LSTM mimicked the test data better than the decomposed LSTM while the forecast QRA results performed better than the decomposed LSTM. Model performance showed an improved RSME for all SPEI timescales with 12- month timescale showing the best score for fQRA (i.e. RMSE of 0.0184). The PICP (Prediction Interval Coverage Probability) and PINAW (Prediction Interval Normalised Average Width) results indicates that LSTM-fQRA was the best model to forecast SPEI time series at all timescales. This therefore indicates that deep learning neural network models are better in drought forecasting in semi-arid catchments. The models used in this study can therefore be incorporated into early warning systems for drought risk reduction in rural communities in semi-arid regions.
Mots clés : Drought risk|environmental degradation|early warning system|forecasting|water resources
A105390FM