A Fine Grained Forecasting Method for Minute-Level Urban Water Demand

Abstract

Forecasting of one day’s Minute-level urban water demand is essential for urban water dispatch. Minute-level forecasting in one day requires super long steps which are around 300. Common multi-step forecasting method is almost impossiable to achieve the prediction of ultra-long step length. In this paper, we propose a multi-step forecasting method named Fine Grained Forecasting (FGF). Specifically, this method firstly forecasts cumulative flow of urban water demand which is daily water demand. Then fine grained division, searching and prediction are applied for minute-level urban water demand. The principle that FGF is effective has been proved to be valid by careful analysis. And a LSTM-Self-Attention deep neural network is proposed to realize fast and accurate forecasting of daily urban water demand. Forecasting results of daily and minute-level urban water demand have been proved to achieve high accuracy through experiments. This shows the effectiveness of fine grained forecasting.