Research Of Incremental Learning In Water Demand Prediction Of Smart Water Affairs

Intelligent water system constructed using advanced technologies such as Internet of Things and Big Data has greatly improved the level of automation and information,with the progress of science and technology and our country's strong support for smart water construction.The research of this thesis is based on the self-developed intelligent water management system,which monitors the operation status of pipe network in real time through data collection meters such as pressure gauge,flow meter and water quality monitor.And all the monitoring data are transferred and stored together to form a unified data center.Finally a series of services is set up on the data center,which strives to manage the whole production,marketing and decision-making process of the water supply in a meticulous and dynamic way.The construction of intelligent water resources can greatly improve the utilization of water resources,and prediction of water demand is the key part of smart water affairs.It is the necessary content for the city's rational planning and construction.Also it is an effective means to ensure the safe operation of water supply system and realize the scientific management and optimal scheduling.Water demand forecast is divided into long term prediction and short term prediction,and short term prediction is mainly discussed here.Water demand is influenced by macroeconomic and human activities,and the distribution of samples is characterized by periodicity,trend and randomness.Historical data can be used to predict the trend of water demand and predict its value.Firstly,the characteristics of water demand data are analyzed and several traditional forecasting methods for small sample processing are introduced.However,traditional methods are no longer applicable as the number and dimension of samples increase rapidly.So the concept of incremental learning is introduced in this thesis.In the thesis,the incremental regression prediction method based on support vector is studied.After analyzing the advantages and disadvantages of this method,the self-organizing incremental neural network is applied to learn the topological node information of samples and then the weighted support vector regression machine is used to train the data.Incremental learning based on support vector regression could not solve the impact of new samples on data distribution.A prediction model is established by studying the regularity of historical series data.The method proposed in this thesis not only synthesizes the advantages of incremental prediction based on support vector theory,which can retain the information of old batch samples,but it may make up for the shortcoming of incremental model based on general support vector by using the distribution density of learned samples.Finally,the validity of the method is verified by the data collected on the spot.