Multi-Model Ensemble Annual Rainfall Forecasts for Mountainous Lincang, Southwest China: Implications for Water-Resource Management and Climate Resilience

Accurate prediction of annual rainfall is critical for water-resource management in mountainous areas. However, annual series typically exhibit weak persistence and pronounced interannual oscillation, fundamentally constraining predictability. Based on 47-year (1979–2025) records from 34 stations in Lincang, Yunnan, this study developed site-specific Autoregressive Integrated Moving Average (ARIMA), Prophet, and long short-term memory (LSTM) models, plus a combined weighted model with an equal-weight fallback mechanism. To prevent data leakage, data were partitioned chronologically into training, validation, and testing subsets; the extreme wet year 2025 (regional mean: 1580.5 mm, +17.5% vs. climatology) was strictly isolated as the test set. Performance ranked: ARIMA > Combined Model ≈ Prophet > LSTM. ARIMA was optimal at 28 stations (Mean Absolute Error (MAE): 339.0 mm; Mean Absolute Percentage Error (MAPE: 21.2%). Equal-weight fallback activated at 29 stations, confirming that most base models underperformed climatology during dry validation and that annual predictability is inherently limited. The combined model prioritized error cancellation over single-model optimality. All models systematically underestimated 2025 rainfall; nevertheless, Pearson correlation across stations ranged from 0.87 to 0.90, indicating that spatial wet–dry patterns were well reproduced and errors were mainly level shifts rather than spatial distortion. Univariate models inherently struggle to capture upward extremes; after climatological bias correction, predictions remain useful for regional water allocation and operational forecasting.