Rainfall–Runoff Prediction Using Spatiotemporal Precipitation Pattern Mining and Pattern-Guided Spatial Attention with GRU

Accurate monthly runoff prediction remains challenging because precipitation exhibits substantial spatial heterogeneity within a watershed. Conventional basin-averaged rainfall may obscure spatial differences, whereas direct use of grid-cell rainfall can introduce high-dimensional and redundant inputs. This study develops a pattern-guided spatial attention gated recurrent unit (PG-SA-GRU) model to better represent and utilize spatial precipitation information for monthly runoff prediction. Principal component analysis and spatiotemporal spectral clustering were first employed to identify spatial precipitation patterns by jointly considering precipitation variability and spatial proximity. Pattern-guided aggregation was then used to organize regional precipitation features, while spatial attention adaptively learned their relative contributions, followed by GRU-based temporal dependency modeling. Application to Yichang station in the Upper Yangtze River Basin identified five spatial precipitation regions with distinct precipitation characteristics and different lagged correlations with runoff. A ten-month historical input window yielded the best prediction performance, with an NSE of 0.905, R of 0.962, RMSE of 5.926 mm/month, and MAE of 3.967 mm/month. PG-SA-GRU outperformed seven benchmark models under identical evaluation conditions. Compared with the standard GRU, it increased NSE and R by 14.70% and 5.71%, respectively, while reducing RMSE and MAE by 32.54% and 44.21%. These results demonstrate the value of spatial precipitation pattern identification and regional information organization for monthly runoff prediction.