Research On Time Series Prediction Algorithm Based On Machine Learning

Time series data generally refers to a sequence of observations of a particular variable changing its value in a system according to time.The integrated result is influenced by various factors in the system.The time series data reflects the changes of characteristics,development trends and motion rules of the particular object.Time series data can be high-dimensional or univariate.Time series data of different characteristics must be modeled by suitable methods to obtain acceptable performance.The ultimate purpose of this dissertation is to study the research strategy using time series data with different characteristics.The main research contents of this dissertation include:1)For the univariate small size time series data,the method Adaboost.RT is used to integrate different learning models to improve the prediction accuracy of small sample single-dimensional time series,since deep learning models are hardly applicable under this situation.In addition,individual machine learning model is not effective.The integrated learning algorithm is proposed and different classic machine learning models are used as weak learners.The Adaboost.RT algorithm is utilized for integration.When the number of data samples increases,more complex models are constructed.However,the proposed models may have higher requirements of the efficiency improvement,dimension reduction,complexity analysis,and deep learning neural network implementations.2)For univariate time series data with sufficient number of data samples,a method combining static wavelet transform with long-term and short-term memory neural network(LSTM)is proposed.Univariate time series data with large volatility,high uncertainty and high noise is difficult to predict.It is proposed to apply multi-scale single-step static wavelet decomposition method,and use the appropriate wavelet basis function to decompose the original signal into low frequency signal and high.The frequency signal is then separately trained and predicted for each signal after decomposition using the hybrid LSTM model.Finally,the almost lossless inverse wavelet transform of the wavelet transform is used to obtain the predicted value of the original signal.3)For high-dimensional time series data,feature engineering and feature extraction have great influence on the regression prediction effect.Therefore,after feature engineering of the original data,feature extraction is performed on the multi-dimensional feature using the inception-Resnet structure,and input to the GRU cyclic neural network model for prediction.The end-to-end model of deep learning is built to predict the time series data with multi-dimensional and multi-features,avoiding complex feature engineering and realizing the mining and extraction of the characteristics of the deep learning model.