Research On Prediction Algorithm Of Urban Daily Water Demand Based On Feature Combination

With global warming,population growth and urban expansion,the daily water demand of cities is increasing,resulting in more and more cities facing the problem of water resource shortage.It is important to make reasonable planning and effective management of limited water resources.To this end,accurate and reliable daily water demand prediction algorithm is essential.However,the traditional forecasting models cannot meet the growing demand of urban daily water demand forecasting.On the one hand,it is difficult to select reasonable influencing factors as characteristics thanks to the complex and changeable factors affecting the daily water demand of cities.On the other hand,the prediction accuracy of the current urban daily water demand prediction model is strongly dependent on data,and the learning ability is not enough.More recently,how to accurately predict urban daily water demand by machine learning method has become a research hotspot of urban intelligent water supply system.Therefore,this thesis firstly systematically analyze the influencing factors of urban daily water demand and carry out feature extraction.Based on this,constructs the urban daily water demand forecasting model based on machine learning method.The main research contents are as follows.Aiming at the characteristics of high correlation and large fluctuation of short-term water demand,this thesis proposes an urban daily water demand forecasting algorithm combines the local weighted regression seasonal trend decomposition procedure based on loess(STL)and temporal convolutional network(TCN),named STL-TCN.In STL-TCN,the original time series is firstly decomposed into trend term,season term and residual term by STL,then each sub series is predicted by TCN,and the final prediction result is obtained by adding the prediction results.Wherein,STL significantly improve the stability of the data by decomposing data,and TCN can effectively extract the characteristics of daily water demand time series due to its good time domain modeling ability and feature extraction ability.Finally,the extensive simulation results verify the prediction effect of STL-TCN.However,in some special weather conditions,the STL-TCN algorithm may have a large prediction error for the urban daily water demand,owing to it only uses historical water demand data for prediction.To solve this issue,this issue further establishes a modified model of water demand prediction.Firstly,the correlation coefficient analysis is used to verify the correlation between urban daily water demand and external factors.Then,the STACK model has been established.The combined data of STL-TCN prediction results,phase space reconstruction data and climate data are used as the input of STACK model.Finally,to further improve the prediction accuracy,the prediction results of STL-TCN are modified by STACK model.The simulation results show that there are significant correlations between climate factors,historical daily water demand and urban daily water demand,and the prediction accuracy of urban daily water demand can be improved to a certain extent with STACK model.