An ensemble learning framework for long-term flood forecasting

考虑到洪水风险，河流洪水预报对水资源管理和风险防范具有重要意义。尽管长期洪水预报是一项艰巨的任务，但长期洪水预报非常有用，例如让市政当局有足够的时间进行准备和行动。最终建立长期预报框架的努力通常面临来自天气动力学的挑战。机器学习技术最近被广泛应用于可持续基础设施的复杂问题建模，特别是在极端事件预测方面。特别是，机器学习技术已经成功地应用于洪水预报，提供了改进的预测技术和相对更准确的结果。最近，集成学习受到了极大的关注。与单一模型相比，集成学习提供了更稳定的预测性能，推动了集成学习的不确定性行为的减少。本研究的主要目标是开发基于集成的机器学习（EML）模型，用于在不同信息标准和有限极端事件历史下的河流流量长期预测。新的、高度准确和可靠的长期预测模型将为美国及其在加拿大的客户提供非常有用的模型，这些模型将大大改进长期预测，从而有助于有效和可持续地规划和管理极端事件应对策略。

Given flood risks, forecasting river flood is important for water resource management and risk preven-tion.Even though long-term flood forecasting is a difficult task, log-term flood forecasting is very useful forinstance to let municipalities to have enough time for preparation and action. The efforts in ultimately creatinga long-term forecasting framework are usually faced with the challenges stemming from weather dynamics.Machine learning techniques have recently been recognized and widely adopted for modeling complexproblems in sustainable infrastructures, especially in forecasting extreme events. In particular, machinelearning techniques have been successfully applied for flood forecasting and provided improved forecastingtechniques and relatively more accurate results. More recently, Ensemble learning has re-ceived a significantamount of interest. Ensemble learning provides a more stable prediction performance compared to singlemodel, driving the diminishing uncertainty behaviour of ensemble learning. The main goal of this researchproject is to develop ensemble based machine learning (EML) models for the long-term forecasting of riverflow under different information criterion and limited history of extreme events. The new, highly accurate andreliable long-term forecasting models will provide US, and in turn their clients across Canada, with very usefulmodels that will allow significantly improved long-term forecast-ing and will consequently help in effectivelyand sustainably plan and manage extreme events response strategies.