Time Series Forecast Via Similar Fluctuate Pattern

The time series similarity measure was the research focus of time series data mining. Because of the complexity of the real system and the effect of observation condition, there was much similarity shape in time series, such as amplitude shift, amplitude scaling, and linear drift in real data set, which caused the insufficiency of similarity measure in time series data mining. However, the currently existing similarity measure methods can’t support the recognition of various similarity shape.Time series forecasting was another research focus of time series data mining. Traditional forecasting model forecast time series by building the model with analysis of historical time series data, which ignored the link between time series and the physical meaning of time series itself. Massive time series data has the characteristics of nonstationarity. The character parameter would change over time, the forecast model should keep track of the change so that it can adapt currently data. The above two points cause the problem of bad stability in the forecast when the traditional forecasting model is used in real.The main works of this dissertation are as follows:(1) In this dissertation, we proposed the fluctuate pattern (FP), FP could save trend information of original time series, we used longest common subsequence (LCS) to calculate the similarity of FP. FP could eliminate the influence of the similarity shape to similarity measure. The simulation experiment proves that the FP algorithm can identify the similarity shape effectively on dataset set by analogue simulation. The experiment on real-world dataset prove that it can also identify complexity mixture similarity shape effectively by evaluating the classification rate.(2) Time series sequence fragments are create by import sliding time window, so that the similarity measure between the long time series could turn to the similarity measure between the sequence fragments. The forecast result is complex by uniform scale the analog patterns which is mined in massive time series data by calculating the similarity. Experiment select the farm product price forecast as a target, set the accuracy and MSE as the evaluation index, the second experiment is set up in contrast by use the ARMA model and the second multinomial exponential smoothing model to forecast. It proves that our algorithm proposed by this dissertation has more accuracy and more stable than classical forecast model.