Research On Forecasting Method And Application Of Urban Water Demand Based On Deep Learning

With the acceleration of China’s urbanization process and the rapid expansion of urban scale,it is imperative to promote the construction of smart city.As an important infrastructure,smart water plays an indispensable role in smart city.Moreover,the urban water demand forecasting is the base of smart water planning and optimal dispatching of water supply network system.Its forecasting accuracy not only determines the rationality and scientificity of the planning and design,but also directly affects the economics,reliability,and water quality of the water supply network system.Therefore,it is of great theoretical significance and practical value to model the forecasting of urban water demand and to improve the forecasting accuracy of the model.The forecasting of short-term daily and hourly water demand provides data support for the optimal dispatching of the urban water supply network system,guides the optimal dispatching of the pumping station,establishes reasonable water distribution management system,and ensures users in the water supply area have sufficient water and enough water pressure,and thus reduce energy consumption in water production and save energy.Domestic and foreign scholars have carried out extensive remarkable theoretical research,but there are very few models and methods applied to practical engineering.Therefore,taking the daily and hourly water demand of waterworks in the smart city Zhuzhou as the research object,this thesis deeply studies the construction of forecasting model and practical engineering application.The main research works are as follows:（1）In terms of the fact that the time series of urban daily water demand includes random disturbances and deterministic components,and have chaotic characteristics,a novel forecasting model of daily water demand is proposed.The model is based on chaos theory and Continuous Deep Belief Neural Network（CDBNN）,which has strong capacity of nonlinear learning.This paper studies the chaotic characteristics of the time series of daily water demand and reconstructs the phase space of the time series.The CDBN model,which is constructed by Continuous Restricted Boltzmann Machines（CRBMs）,is utilized to extract the potential characteristics of the original data of daily water demand.In addition,the Neural Network（NN）model is applied for feature regression to construct the CDBNN model.Finally,the historical data of daily water demand of two waterworks in Zhuzhou are applied for modeling and forecasting.In comparison with Support Vector Regression（SVR）,Generalized Regression Neural Network（GRNN）and Feed Forward Neural Network（FFNN）,the proposed forecasting model of daily water demand achieves better forecasting accuracy.（2）Based on the idea of local modeling,this paper proposes a Dual-Scale Deep Belief Network（DSDBN）model for forecasting daily water demand.The original time series of daily water demand is decomposed into several Intrinsic Mode Functions（IMFs）and a residue component by Ensemble Empirical Mode Decomposition（EEMD）.Employing the generalized Fourier transform to analyze the frequency characteristic of each component,this study reconstructs the stochastic and deterministic terms.Then,the double DBN model is built to predict two feature items,respectively.The final forecasting result of historical data of daily water demand is obtained by the combination of the aforementioned two forecasting results.The forecasting model is constructed and verified by using the historical time series of daily water demand of the A waterworks in Zhuzhou.Compared with the Autoregressive Integrated Moving Average model（ARIMA）,FFNN,SVR,and a single DBN model,the DSDBN model improves the pertinence of forecasting and achieves more excellent forecasting performance.（3）According to the linear and nonlinear characteristics of the hourly water demand time series,a linear-nonlinear forecasting framework is proposed.Due to the fact that the ARIMA model has good linear learning ability,and the CDBNN model has an excellent nonlinear fitting ability,this study constructs a hybrid forecasting model based on ARIMA and CDBNN.The original time series of hourly water demand is decomposed by EEMD.IMF₁ denotes the random components of the original time series and is predicted by CDBNN model.The recombination of the other IMFs and the residue component represents the linear features of the original series and is predicted by the ARIMA model.The final forecasting result of historical hourly water demand is obtained by combining the results of the two models.By using the hourly water demand data of the B waterworks in Zhuzhou,the proposed model improves the forecasting accuracy in comparison with EEMD-BPNN and EEMD-SVR models.（4）For the traditional DBN or CDBNN model,BP algorithm is used to adjust the parameters.It always leads to slow convergence and become easy to fall into local optimum,and thus the model forecasting accuracy is often not satisfactory.To solve this issue,this study introduces the echo state network（ESN）and to construct the CDBESN model.The proposed model uses CDBN as the feature extraction algorithm and ESN as the regression algorithm.The model was tested by the historical data of hourly water demand of the C waterworks in Zhuzhou.Compared with the ESN,CDBNN and SVR models,the CDBESN model performs better in the forecasting of hourly water demand.（5）Based on deep learning,this study designs a prototype system of a cloud platform for urban water demand forecasting.According to the analysis of user demand from urban water demand forecasting and combined with the big data of water demand,this study introduces cloud computing technology,designs the system structure of cloud plat for urban water demand forecasting,describes the workflow of the system and expounds the implementation scheme of CDBNN in Matlab on Hadoop platform.Taking the daily water demand forecasting as an example,this paper introduces the practical application of the prototype system in the D waterworks in Zhuzhou.