Research And Implementation Of Time Series Prediction Method Based On Feature Fusion

The method of time series prediction is to predict the development trend of things in the future by analyzing historical data,and is widely used in various fields such as meteorology and agriculture.Because time series data has the characteristics of large amount of data,non-linearity and many influencing factors,time series prediction has become the difficulty of research.The method of time series prediction mainly has traditional time series prediction method and time series prediction method based on machine learning.The traditional time series prediction method is mainly based on statistical theory,the modeling technology is single,and the prediction accuracy is not high.The time series prediction method based on machine learning not only improves the accuracy of prediction,but also has a strong generalization ability.Although there are many studies on time series prediction based on machine learning,there are still some problems to be solved,and feature extraction is one of them.At present,there are two main problems in feature extraction: the first is the feature extraction method.The method of feature extraction is mainly manual,which has certain subjectivity and incompleteness;the second is the selection of effective features.The quality of features directly affects the level of modeling.How to select effective features when there are many features is also one of the research hotspots of machine learning.In order to solve these two problems,this paper proposes a prediction method which fuses artificial and recessive features generated by deep learning,and integrates the fused feature sets into ensemble model for prediction.The method uses the improved Convolutional Neural Network(CNN)to automatically extract the recessive features,then fuses the recessive features with the artificial features to form the total feature set.Finally,the effective features selected by the random forest are set into the LightGBM algorithm for supervised training.Based on the proposed method,this paper uses two time series data to verify,and establishes the prediction of the amount of vehicle registration and the prediction of the price of agricultural products.Compared with other prediction methods,the results show that the prediction accuracy of the method is higher,which indicate the feasibility and effectiveness of the proposed method.