An innovative hybrid drought modelling approach using wavelet-long-short term memory (W-LSTM) with standardised precipitation index (SPI) data for Badeggi district, Niger State-Nigeria

Abstract This study introduces discrete wavelet transform (DWT) coupled with the long-short-term memory (LSTM) model for drought prediction. We evaluate the W-LSTM model against the traditional ARIMA and conventional LSTM methods, based on the monthly precipitation data from January 1968 to January 2018 of Badeggi District in Niger State – Nigeria; which were evaluated into standardized precipitation index (SPI) of various timescales of 3, 6, 9, and 12 respectively. To understand any of the underlying features in the data, each of the SPI series was plotted graphically. Subsequently, the SPIs undergoes decomposition using Wavelet resulting in (three Details and an Approximation) components. The next step involved determining the input data based on the number of significant lags observed in PACF of each component, thereby restructuring the data to conform to LSTM’s network architecture, thereby making the modelling and forecasting of the input series by the LSTM network. Statistical evaluation metrics which include the root-mean-square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), R 2 and percentage of bias (PBIAS) are employed to assess the model accuracy. The W-LSTM model demonstrated superior performance over ARIMA and LSTM in forecasting, significantly reducing RMSE by up to 83.4% and improving R² by up to 147.5%. For SPI-12, W-LSTM achieved a 56.2% RMSE reduction over ARIMA and a 74.8% reduction over LSTM, with a 46.7% increase in MaPE, a 1.1% increase in R², and recorded the best PBIAS value of 2.078, compared to the other models. Similar trends were observed across SPI-6, SPI-9, and SPI-12, confirming W-LSTM’s effectiveness in minimizing errors and enhancing predictive accuracy when compared to standalone ARIMA and LSTM models.