Data-driven Water Demand Forecast For Distributed Water Supply Network

As the era of big data is coming,amount of data is produced during production process in the industrial system,such as distributed water supply network.This treasure-trove of data is able to provide basis for decision making.The research focus on water network in Shanghai and a data-driven method is introduced to forecast water demand.Accurate prediction is important for intelligent scheduling and efficient running in the water supply system.And this article explores the use of multi-scale analysis in the water demand forecast.Firstly,a method of improved empirical mode decomposition(L-EMD)based on local extremum is proposed in the research.By means of extending time serise at the both side of the data set,the method ensures that the endpoint is on the extreme envelope while cubic spline interpolation is used to calculate the IMF.This method is likely to overcome endpoint effect in EMD,and the decomposition results are improved.Secordly,a hybrid modeling approach based on L-EMD is put forward to forecast short-term demand in Shanghai.The approach decomposes time series in several IMFs by method of L-EMD and analyzes the exogenous variables’ influence on water demand at each time scale independently.The model of each IMF is built by DANN and the experimental results show that the approach will greatly improve the forecasting accuracy.Finally,in order to meet real-time the requirements of the system,a multiple model-based adaptive online prediction method is proposed under the framework of multi-scale hybrid modeling.In the distributed system,the research utilizes L-EMD to reconstuct the data which is decomposed into 3 time scales.Sub-models of relationship between historical data and predicted value are established offline.And the exogenous variables and other nodes in the space are taken into consideration in this research.Then predicted output is constructed by weighted summation of sub-models,and the weights are calculate at every sampling time.The on-line method gets relatively accurate experimental results.