Examining ensemble machine-learning approaches to improve precipitation forecasting

我们星球上的水循环是由复杂的物理过程驱动的，很难为降水预测建立准确的模型。用于这一目的的天气模型必须根据时间和地点的考虑做出简化的假设，因此它们不能普遍适用。通常情况下，会使用多个模型来缓解该问题，但尚未找到将它们结合在一起的最佳方法。事实证明，机器学习技术在发现复杂函数之间的潜在关系方面很有用，超出了传统统计方法的能力，在某些情况下甚至是人类，前提是有足够的数据和计算资源可用。这两种技术都将进行深入的试验和比较。威斯康星大学的研究团队在机器学习方面拥有20年的研究和行业合作经验，其中包括最近与WeatherLogic合作的一个机器学习研究项目。我们的合作伙伴通过获取传统天气信息并将其转换为特定行业的数据，提供专业的产品和服务。产生的一些关键的专业数据集包括路况预报、冰雹跟踪、农业天气和政府管理冬季道路维护业务的平台。这些服务允许公司和政府使用可用的最佳数据做出关键的天气决定。随着硬件能力和产生的数据量的不断增长，我们将研究两种主要的机器学习方法：i)神经网络和ii)随机森林，以开发用于预测降水的集成天气预报模型。该项目的研究成果将导致更准确的降水预报，用于帮助合作伙伴的客户改进决策。改进决策的例子包括更好的农业决策，在道路预报中更好地预报雪或冰，以及更及时地对强降水事件发出警报。从公共安全和经济角度来看，所有这些用例都使加拿大人受益。

The water cycle on our planet is driven by complex physical processes, and difficult to be modelled accurately for precipitation prediction. The weather models used for this purpose have to make simplifying assumptions based on time and location considerations and thus, they cannot work universally. Typically, multiple models are utilized to mitigate the issue, but the best method to combine them is yet to be found. Machine learning techniques have proven useful to discover the underlying relations between complex functions, exceeding the abilities of traditional statistical methods, and in some cases even humans, provided that adequate data and computing resources are available. Both techniques will be experimented with and compared in-depth. The research team at UW brings twenty years of research and industry collaboration experience in machine learning which includes a recent machine learning research project with Weatherlogics. Our partner provides specialized products and services, by taking traditional weather information and transforming it into industry-specific data. Some key specialized datasets produced include road condition forecasts, hailstorm tracking, agriculture weather, and platforms for governments to manage winter road maintenance operations. These services allow companies and governments to make crucial weather-dependent decisions using the best available data. With ever-growing hardware capabilities and the amounts of data generated, we will examine two prominent machine learning approaches i) neural networks and ii) random forests to develop an ensemble weather forecast model for predicting precipitation. The research outcome of this project will lead to more accurate precipitation forecasts, which are used to assist the partner's client to improve decision-making. Example of improved decisions include better agricultural decisions, better predictions of snow or ice in road forecasts, and more timely alerts of heavy precipitation events. All of these use cases benefit Canadians, both from a public safety and economic point of view.