Short-term forecasting of hourly wa

Short-term forecasting of hourly water consumption by using automatic metering readers data
Abstract
A completely data-driven,fully adaptive self-learning algorithm for water dermand forecasting in the short-term and with hourly periodicity is proposed, according to the renewed interest generated by the availability of new technological solutions such as Automatic Metering Readers (AMR) , a key enabler of the Smart Water paradigm. The approach is based on two sequential stages: at the first stage (time-series clustering) the daily water demand patterns (i.e.,time-series of hourly data) are analysed to identify alimited set of typical behaviours. At the second stage Support Vector Machine regression is used to obtain one specific forecasting model (consisting of a regression madel for each hour) for each cluster identified at the first starge. The approach has been validated on real data acquired by AMRs deployed on the Italian pilot site of IceWater, computing the widely adopted error measure MAPE (Mean Absolute Percentage Error).