Research And Application Of Multi-task Learning In Time Series Prediction

Time series is a common data structure,which widely exists in human society and objective world.It aims at establishing the mathematical model to predict the future trends of time sequence according to the implicit intrinsic changes in the time series and the implied time evolution.Some available information from the time series data of different observations can be mined by in-depth analysis,and it has a great significance on identifying the development trends of objects,describing the relevant characteristics and controlling the change between things.At present,the time series forecasting method based on data driven can be divided into global model and local model.The time series data obtained from the real world often have the characteristics of strong nonlinearity and uncertainty,which makes the modeling process more difficult and inefficient,thus limiting the prediction accuracy of the model.The global model is simply modeling based on all historical data,but it is very sensitive to the outliers in the sequence.Aiming at the above problems,this thesis proposes a local modeling method based on the cloud model similarity measurement.A prediction model combined with BPNN and LS-SVM is also built to improve the prediction accuracy of sequence.Although the local model has a certain contribution to the prediction accuracy,it still belongs to the category of single task learning.The generalization performance of the model is affected because the related information is not fully excavated in single task.However,multi task learning takes into account the relevance and differences between the tasks.And the learning performance is ultimately improved due to mining the common information between the shared knowledge structures of all tasks.Considering the advantages of local model and multi task learning,this thesis proposes a local modeling method based on multi task learning.In order to improve the generalization ability of the model,the rich information between adjacent time points of time series is regard as the different tasks.The main work of this thesis includes the following aspects:(1)The advantages and disadvantages of several common time series similarity measurement methods are analyzed,and two different types of time series prediction methods are compared;(2)In order to solve the problems such as nonlinearity,complexity and uncertainty of the original time series data,the cloud model theory is used to represent the original sequence and the sequence processed by first order difference simultaneously;(3)In view of the traditional distance function cannot effectively measure the uncertainty data,in this paper,a local modeling method based on cloud model similarity measure is proposed,and BPNN and LS-SVM are used to build the forecast model;(4)Aiming at the problem that the information mining in single task learning is not sufficient and the prediction accuracy is low,we propose a local modeling method based on multi-task learning.In the method,the multi-task learning is used in the time series prediction under the framework of the proposed local modeling method;(5)The proposed method was validated on six real engineering datasets selected from the Applications of Machine Learning(AML)research group of the Aalto University School of Science.