The complex and unique topography of karst regions highlights the weaknesses of traditional hydrological models which fail to fully generalize them. The successive proposals of standard artificial intelligence (AI) models, pre-processing techniques, and post-processing methods have provided new opportunities to enhance the accuracy of runoff prediction in karst areas. In this study, first, the BP neural network model and the Elman neural network model were used for runoff prediction. Then, the performance of four coupled models—formed by combining two AI pre-processing techniques, Empirical Modal Decomposition (EMD) and Ensemble Empirical Modal Decomposition (EEMD), with the previously mentioned AI models—was investigated. Finally, the accuracy of triple-coupled models, formed by applying the post-processing method of quantile mapping (QM) to the previous coupled models, was estimated. The Nash–Sutcliffe efficiency (NSE), the mean absolute percentage error (MAPE), the root mean square error (RMSE), and the peak percentage of threshold statistics (PPTS) were selected to evaluate and analyze the forecasting results of the above models. The results demonstrated that the BP model had the best prediction effect of the standard AI models, the coupled forecasting models had better prediction accuracy than the standard AI models, and the triple-coupled QM–EMD–Elman model had the best forecasting effect with an NSE value of 0.73, MAPE value of 0.75, RMSE value of 34.60, and PPTS value of 2.36.

To improve reservoir flood control and scheduling schemes under changing environmental conditions, we established an adaptive reservoir regulation method integrating hydrological non-stationarity diagnosis, hydrological frequency analysis, design flood calculations, and reservoir flood control optimization scheduling and applied it to the Chengbi River Reservoir. The results showed that the peak annual flood sequence and the variation point of the annual maximum 3-day flood sequence of the Chengbi River Reservoir was in 1979, and the variation point of the annual maximum 1-day flood sequence was in 1980. A mixed distribution model was developed via a simulated annealing algorithm, hydrological frequency analysis was carried out, and a non-stationary design flood considering the variation point was obtained according to the analysis results; the increases in the flood peak compared to the original design were 4.00% and 8.66%, respectively. A maximum peak reduction model for optimal reservoir scheduling using the minimum sum of squares of the downgradient flow as the objective function was established and solved via a particle swarm optimization algorithm. The proposed adaptive scheduling scheme reduced discharge flow to 2661 m³/s under 1000-year flood conditions, and the peak reduction rate reached 60.6%. Furthermore, the discharge flow was reduced to 2661 m³/s under 10,000-year flood conditions, and the peak reduction rate reached 65.9%.