Research And Application Of Time Series Classification And Prediction Algorithms

Time series classification and forecasting are important parts of time series analysis.By analyzing a large number of observation data,it is possible to predict the future development trend of data in order to control the events that will occur.In this thesis,we study three problems of time series classification,multi-step traffic flow forecasting and early time series classification in time series analysis.In response to these problems,we propose three improved algorithms and apply them to the recognition of excessive concentration of smoke in the tunnel.Problems with data distortion and noise effects in time series classification,we propose an improved time series classification algorithm based on multi-view canonical correlation analysis.Firstly,we use training data to define a template sequence for each category.Secondly,we use the dynamic time warping algorithm to calculate the dynamic time warping path between each time series and each template sequence and use the path as the dynamic time warping feature of the sequence,and then extract the Histogram of Oriented one Dimension temporal Gradient feature of the original sequence.Finally,we use these two the features as the two views of the original time series and merged them by a multi-view canonical correlation analysis method,and then use the fused features to classify.Besides,we apply the algorithm to the recognition of excessive concentration of smoke in the tunnel,and the experimental results show that the improved time series pattern classification algorithm based on the multi-view canonical correlation analysis has a better classification effect than the time series classification algorithm based on single feature.In view of the strong periodicity and volatility of traffic flow time series,we propose an improved multi-step traffic flow forecasting algorithm based on empirical mode decomposition.In this algorithm,the data at the same moment in different cycles are used to compose the longitudinal sequence and the longitudinal data is subjected to empirical mode decomposition.Then,the fluctuating part and the base traffic flow corresponding to each cycle are obtained.The forecast result of the fluctuating part plus the base traffic flow of its corresponding moment is taken as the longitudinal prediction result.The final forecast result is added the longitudinal forecast results to the forecast results based on the direct strategy.Then,using the Vanet Mobi Sim simulation software,the macroscopic movement model and the microscopic movement model of the urban road traffic flow simulation system are constructed.Then the urban road traffic flow is simulated and the traffic flow time series is extracted.Experimental results show that the improved multi-step traffic flow prediction algorithm based on empirical mode decomposition has better prediction accuracy.For the problem of early time series classification,the existing K-Nearest-neighbor algorithm is used to improve the existing time-series classification algorithm based on cost aware.A framework for early time series classification is proposed,where the cost of delaying the decision is included in the optimization function.Firstly,the cost of each training time series at each time point is calculated in the training process.In the test process,The K nearest neighbors algorithm is used to obtain the K nearest neighbors of the test time series in the training set,and then the future cost of the K neighbors is used to estimate the future cost of the time series to be classified.According to the calculated cost,the optimal classification moment is obtained and then the classification is performed.The experimental results show that the improved early time series classification algorithm has better classification results.