Hybrid singular spectrum analysis and LSTM modeling for daily precipitation forecasting

To address the limitations of traditional univariate time series methods in achieving accurate daily precipitation forecasts, this contribution proposes a hybrid predictive model that integrates singular spectrum analysis (SSA) with a long short-term memory (LSTM) network, termed SSA-LSTM. The model first employs SSA to decompose the original precipitation series, effectively extracting key temporal features such as trends and periodic components. It then utilizes the LSTM network to capture complex nonlinear relationships, thereby enhancing prediction accuracy. Daily precipitation data from four urban stations (Ganzhou, Shangrao, Jiujiang, and Yichun) served as experimental datasets. Comparative experiments were conducted among several models: single models including the temporal convolutional network (TCN) and LSTM; comparisons between single and hybrid models; and evaluations of different hybrid approaches, namely CEEMDAN-TCN, CEEMDAN-LSTM, VMD-LSTM, and the proposed SSA-LSTM. The results showed that the LSTM model outperformed the TCN, while hybrid models demonstrated significantly better performance than individual models. Among the hybrid architectures, the proposed SSA-LSTM achieved an R 2 value between 0.988 and 0.990, exceeding both CEEMDAN-LSTM ( R 2 = 0.710–0.802) and VMD-LSTM ( R 2 = 0.980–0.983). In terms of RMSE and MAE, SSA-LSTM achieved reductions of 77.48%–79.69% and 75.31%–79.85%, respectively, compared to CEEMDAN-LSTM, and 16.59%–25.78% and 26.62%–31.67%, respectively, compared to VMD-LSTM. These findings confirmed the effectiveness of SSA decomposition in extracting informative temporal features. Furthermore, SSA-LSTM exhibited superior computational efficiency compared to VMD-LSTM, providing an efficient solution for high-accuracy daily precipitation forecasting. This approach can offer a reliable basis for meteorological disaster warning and water resource management.