Optimizing the treatment of drinking water using reinforcement learning

该项目的目标是为饮用水处理中膜超过滤的自动化开发机器学习技术。我们将在试点规模的实验平台上研究这些方法的可行性和有效性，最终实现全面部署。****目前，这个复杂的过程是由人工操作员管理的。因此，现代工厂水处理过程控制的成本效益在很大程度上依赖于操作人员的技能、判断力和直觉，而这些只能通过多年的经验来获得。因此，水处理的有效性将受到以下因素的限制：经验丰富的操作人员的可用性；由于其他工作要求，操作人员能够调整工艺参数的频率有限；以及可供他们做出决定的可用信息有限。随着运营大型工厂的复杂性增加，招聘和培训人工操作员变得越来越具有挑战性，尤其是对小型社区而言。****我们将研究使用强化学习来自动化部分水处理过程。这些方法使用恒定的传感器信息流，包括有关水质、流量、需求和电价的信息，在不影响水质或可用性的情况下将成本降至最低。例如，这些学习算法可以确定如何优化泵的运行以降低电力成本。另一个例子是，清洁过滤器是一项巨大的成本；这些算法可以提供更细粒度的控制，可以在几秒钟或几分钟内做出决定，而不是几小时或几天，从而降低清洁成本并延长过滤器的使用寿命。即使是很小的优化，该项目也有可能显著节省水处理费用

The goal of the project is to develop machine learning techniques for the automation of membrane ultra- filtration in drinking water treatment. We will investigate the feasibility and efficacy of these approaches on a pilot-scale experimental platform, towards eventual full-scale deployment.****Currently, this complex process is managed by human operators. Thus, the cost-effectiveness of water- treatment process control in modern plants relies heavily on operator skill, judgment and intuition, which can only be gained over many seasons of experience. As a result, the effectiveness of water treatment will be limited by the availability of experienced human operators, the bounded frequency with which human operators are able to adjust process parameters due to their other job requirements, and the limited amount of information available to them on which to base their decisions. As the complexity of operating large plants increases, recruiting and training human operators is getting ever more challenging, especially for small communities. ****We will investigate the use of reinforcement learning for automating parts of the water treatment process. These approaches uses a constant stream of sensor information, including information about water quality, flow, demand and electricity prices, to minimize costs without impacting water quality or availability. For example, these learning algorithms could determine how to optimize operation of pumps to reduce electricity costs. As another example, cleaning the filters is a significant cost; these algorithms could provide more fine-grained control, that makes decisions over seconds or minutes rather than hours or days, reducing cleaning costs and increasing the life-span of the filters. With even small optimizations, this project has the potential to produce significant savings for water treatment.**