LSTM-Based Prediction and Evaluation of Meteorological Drought Indices Considering Cumulative Precipitation Timescale Combinations

This study applied a deep learning-based time series prediction model to drought index forecasting by comparing two input configurations: a univariate model using a single-timescale drought index and a multivariate model integrating multiple timescales. The Standardized Precipitation Index (SPI) and the Standardized Precipitation Evapotranspiration Index (SPEI) were selected as target indices, and Gwangju Metropolitan City served as the study area. Using climate time-series data, a Long Short-Term Memory (LSTM) model was developed and evaluated using R<sup>2</sup>, RMSE, MSE, MAE, and NSME. Results demonstrate generally stable predictive performance, confirming the model’s applicability to drought index forecasting. To examine practical usability, the predicted drought stages were compared with the official drought warning system of the National Drought Information Statistics System. For some indices, the multivariate structure achieved more stable performance than the univariate structure, and the SPI exhibited relatively stronger agreement with official drought stages. The results of this study can serve as a reference for input design and the operational application of deep-learning-based drought prediction models.