A Stream Fuzzy Time Series Forecasting Model

Abstract

Time series classification has been studied for over a decade and is now widely used in the sphere of data mining to increase the forecasting accuracy In recent years the evolution of information technology has caused a change in the data-storage approach As volume data is collected and stored continuously and rapidly a data analyzer is not able to efficiently retrieve the information from it over time Thus a new data-processing approach called ‘streaming’ was proposed which entails inputting data elements as sequences The advantage of streaming data is that data points can still be used to forecast future values while the total length is unknown Streaming data is diverse continuous rapid and time-varying thus it is not compatible with the conventionally stored data model To construct a novel approach that differs from a traditional forecasting model we used on-line learning to process data instantly We used a dynamic-adjusting mechanism to detect when to add a rule update a rule or delete a rule With these steps we can make our fuzzy time-series forecasting model conform with streaming data well In this research we focused on the combination of streaming data and fuzzy time series The recursive density updating algorithm is used in our model to decide the rule-updated or rule-added timing The purpose of using a dynamic-adjusting mechanism is to raise the rule-usage ratio and to remove redundant rules In addition it is also our goal to improve the forecasting accuracy of our model In doing so we refer to the on-line learning strategies of Mill?n-Giraldo S?nchez and Traver (2011) who proposed classifying the incoming data with missing attributes We used two strategies to simulate how to forecast value when parts of the data are delayed

Date of Award	2016 Jul 25
Original language	English
Supervisor	Sheng-Tun Li (Supervisor)