Accurate precipitation forecasting is essential for water resource management, flood early-warning systems, and agriculture, but remains difficult because of the nonlinear and highly variable spatiotemporal nature of rainfall. This paper compares four deep learning architectures—a standalone LSTM, a standalone CNN, a hybrid CNN–LSTM, and a Transformer encoder—against three classical baselines (persistence, day-of-year climatology, and per-grid-point ARIMA) for daily precipitation forecasting over Washington State at lead times of one to four days. A 40-year ERA5 dataset (1985–2024) of near-surface air temperature, mean sea-level pressure, and total precipitation is split into training (1985–2012), validation (2013–2015), and test (2016–2024) periods, with the test years held out completely. Each (model, horizon) is trained with three random seeds and evaluated in physical units (mm/day). On the held-out test period, the hybrid CNN–LSTM achieves the lowest RMSE at every horizon h≥2, with R2=0.576±0.007 and RMSE =15.08±0.07 mm/day at h=4. Diebold–Mariano tests, paired t-tests, and bootstrap 95% confidence intervals confirm that the CNN–LSTM advantage over the LSTM is statistically significant at horizons 2–4 (but not at h=1), while CNN–LSTM is significantly better than every classical baseline and the Transformer at every horizon. The headline result is reproduced under a rolling-origin temporal cross-validation across three non-overlapping splits (R2∈0.576,0.590). Practically, the sub-millisecond inference cost of the CNN–LSTM makes it directly deployable in operational forecasting pipelines used for flood early-warning, irrigation scheduling, and reservoir management, where even modest improvements in 3–4-day-ahead RMSE translate into measurable risk reduction and improved decision lead time for water managers and emergency planners.
Al-Omari et al. (Fri,) studied this question.