ABSTRACT

Water demand forecasting is a crucial task in the efficient management of the water supply system. This paper compares classical and adapted machine learning algorithms used for water usage predictions including ARIMA, support vector regression, random forests and extremely randomized trees. These models were enriched with human mobility data to improve the predictive power of water demand forecasting. Furthermore, a framework for processing mobility data into time-series correlated with water usage data is proposed. This study uses 51 days of water consumption readings and over 7 million geolocated mobility records from urban areas. Results show that using human mobility data improves water demand prediction. The best forecasting algorithm employing a random forest method achieved 90.4% accuracy (measured by the mean absolute percentage error) and is better by 1% than the same algorithm using only water data, while classic ARIMA approach achieved 90.0%. The Blind (copying) prediction achieved 85.1% of accuracy.

摘要

需水预测是有效管理供水系统的关键任务。本文比较了用于用水量预测的经典和适应性机器学习算法，包括ARIMA，支持向量回归，随机森林和极随机树。这些模型充斥着人类流动性数据，以提高需水量预测的预测能力。此外，提出了一种用于将流动性数据处理为与用水量数据相关的时间序列的框架。这项研究使用了51天的耗水量读数和超过700万条来自市区的地理位置流动记录。结果表明，使用人类流动性数据可以改善需水量预测。采用随机森林方法的最佳预测算法达到90。与仅使用水数据的相同算法相比，精度为4％（由平均绝对百分比误差衡量），比同一算法高1％，而经典ARIMA方法则达到了90.0％。盲（复制）预测达到了85.1％的准确性。