A Comparative Study Of The Multi-step Prediction Of Time Series With Different Sampling Frequencies

Time series is arranged in accordance with the chronological sequence data. Time series exist in all areas of our lives, due to incomplete randomness and easy accessibility of the time series, more and more data mining researchers hoping to find some law from specific time-series data and to do more accurate predictions. Time series analysis method is a hot spot in Statistics and Economics. Through time series analysis and forecasting, people can conducive to grasp the dynamic changes of the time series data and trends, which bears law, resource management and economic reality forecasts and other aspects has important practical significance.Focusing on time series different prediction model research, this thesis proposes a comparative study of single-step and multi-step prediction time series for different sampling frequencies. The main contributions are as the following:Firstly, this paper research the background of the development and current status of the time-series data mining technology, introduced the concept of the time series and time series data mining, as well as the definition of the time series and the basic steps of time series prediction. The time series prediction methods can be divided into single-step prediction and multi-step prediction, the multi-step prediction can be divided into direct and indirect multi-step prediction. Time series prediction basic steps include:data acquisition, data preprocessing, data analysis and model selection, modeling, predicting data.(See1.1,1.2,2.1-2.3)Secondly, this paper researched the two basic methods of time series prediction, there is smoothing method and fitting method. The basic idea of smoothing is moving average, and which the paper mainly studied is the Seasonal Exponential Smoothing Method and Adaptive Filtering Method; The fitting method is that use the best model fit the existing time-series data, this paper mainly studied the ARMA model. In addition, this paper researched the neural network in time series prediction. In addition, for the application of neural network in time series prediction was constructed BP Neural Network time series forecasting model.(Chapter2)Thirdly, this paper introduced the concept of direct multi-step prediction and indirect multi-step prediction, which is based on the classic time series model. This paper made direct multi-step prediction with ARMA, Seasonal Exponential Smoothing Method, Adaptive Filtering Method and BP Neural Network, for Hourly, Daily, Weekly, Monthly and Quartly data-five groups of different frequencies data-and made a comparative analysis of the prediction error.(Chapter3)Finally, on the basis of single-step prediction, using ARMA, Seasonal Exponential Smoothing Method, Adaptive Filtering Method and BP Neural Network to make indirect multi-step prediction on five groups of different frequencies data, and made a comparative analysis of the prediction error. In addition, this paper made a comparative analysis of direct multi-step prediction and indirect multi-step prediction performance. And made a comparative analysis of the results of different models forecast of the same frequency.(Chapter4)...