Accurate river flow forecasting is vital for flood warning and water resource management, yet hourly-scale prediction in small catchments remains underexplored despite its importance for rapid response flood control. To address this gap, this study proposes an enhanced temporal attention module xLSTM (TAM-xLSTM) model that combines temporal feature extraction with timestep-level attention to better capture dynamic variations and dependencies. Case studies in the Qiandongnan region demonstrate that TAM-xLSTM substantially outperforms baseline models during wet season forecasting at Panghai Station, reducing RMSE by 9.6%, MAE by 24.1%, and Theil’s U by 6.6%, while increasing NSE by 4.8%. These results highlight the model’s ability to improve short-term river flow prediction in complex mountainous terrain and its potential to support effective flood warning and water resource management.

Runoff forecasting is an essential application in the management of water resources and sustainable development. In practice, there are limitations in the forecast results because of factors such as data unavailability, noise interference, and spatiotemporal variation in multi-site data. To overcome the limitations, this paper proposes a hybrid forecast model based on Autoencoder (AE), Sparsified Dynamic Graph Convolution (SDGC), and Autoformer. The AE cleans noise and sharpens feature representation, the SDGC constructs dynamic adjacency matrices via the Multidimensional Dynamic Time Warping (MDTW) and sparsifies with a parameterized Multi-Layer Perceptron (MLP) to capture time-varying spatial correlations among stations, and the Autoformer decomposes features to model long-term nonlinear runoff trends through its autocorrelation mechanism. The experiment was carried out in six locations in the southeastern part of Guizhou province during the wet and dry periods and was contrasted with different mainstream models and supplemented with hydrological mechanism consistency analysis. Experimental results show that the hybrid model performs better than all the other models. In the short-term runoff simulation at XingHua Station during the wet season, NSE attains the maximum value of 0.891, with RMSE decreased by 6.5% to 24.1% and MAE by 20.2% to 35.5%. This model provides accurate runoff data to support flood early warning, dry-season water scheduling, and ecological flow protection, offering a reliable tool for sustainable water resource management in complex karst basins.

Quantitative evaluation of artificial precipitation enhancement effects remains a critical challenge in meteorological science. Traditional methods rely on idealized assumptions and control area selection, with limited applicability under complex terrain conditions. To address this problem, this paper proposes physics‐informed simple video prediction (SimVP) (PiSim), a hybrid physics‐informed disentangled framework designed to disentangle precipitation evolution. By employing the advection–diffusion equation (ADE) as an inductive bias within the latent space, we transform spatial comparative evaluation into a precise spatiotemporal sequence prediction task. The method adopts a dual‐branch architecture: The physics branch explicitly models deterministic macroscopic motion (advection and diffusion), whereas the data‐driven branch learns complex nonlinear residuals, effectively compensating for microphysical processes and local variations. Experimental results on the Hubei Province Swan radar dataset demonstrate that PiSim achieves a 5.5% improvement in MSE compared to the SimVP baseline, with pronounced advantages in heavy precipitation forecasting. Evaluation of 10 typical artificial precipitation enhancement operations shows hourly net rainfall increments ranging from 0.16 to 4.60 mm, which are highly consistent with historical records, validating the method’s effectiveness.