关键词: 月径流概率预测, 非交叉分位数回归, 来水情景生成, 水量平衡校核, 跨流域调水

Abstract: Presently, in annual inter-basin water transfer planning, the representation of monthly inflow uncertainty is insufficient, and it is usually difficult to determine the dry-scenario water shortage risks and the wet-scenario regulation pressure. Aimed at these issues, this study construct a novel framework for integrating monthly runoff probabilistic forecasting, inflow scenario generation, and water-balance checking. Observed monthly runoff is used as a unified evaluation target, and monthly mean runoff estimated from the training phase is adopted as a seasonal baseline; then probabilistic residual modelling is conducted. We develop a bidirectional gated recurrent network with an attention mechanism and non-crossing quantile constraints so as to generate the dry, normal, and wet inflow scenarios. A case study of the Sanhekou reservoir shows that the Nash-Sutcliffe efficiency coefficient in the test period is 0.36, and the coverage probability of the 90% prediction interval is 0.93. With an illustrative annual transfer target of 500 million m3, the dry scenario supplies 181.2 million m3 transferable water with a shortage of 318.8 million m3. The normal and wet scenarios both meet the target, while the spill or forced-release volumes are 114.1 million m3 and 1.4 billion m3 respectively. The results indicate that our new framework is satisfactory in converting the runoff forecasting into the inflow boundaries and risk indicators that are needed in annual water transfer planning and its rolling adjustment.

Key words: monthly runoff probabilistic forecasting, non-crossing quantile regression, inflow scenario generation, water balance checking, inter-basin water transfer