基于长短时记忆和动态贝叶斯网络的序列预测

Time Series Prediction Based on Long-short Time Memory and Dynamic Bayesian Network

伴随着计算机视觉技术的迅猛发展,时间序列预测问题在算法优化中扮演着越来越重要的作用。由于数据不确定性的增加,多步预测遇到了巨大的挑战。针对传统预测模型中累积误差造成的预测精度低和算法复杂度等问题,提出了一种基于长短时记忆神经网络(LSTM)和动态贝叶斯网络(DBN)的时间序列预测模型,研究并证明了一种最优估计理论,并在此基础上得到了最优的预测估计。利用递归图模型,通过概率推理提高了预测性能,建立了一种由长短时记忆预测模型和动态贝叶斯网络组合成的新的图模型,称其为基于长短时记忆神经网络和动态贝叶斯网络的时间序列预测模型(LSTM-DBN),用于预测序列数据。仿真结果表明,该模型能够在提高序列预测精度和速度的同时,降低算法的复杂度。

With the rapid development of computer vision technology,time series prediction is playing an increasingly important role inoptimization of algorithms. Due to the increase in data uncertainty,the multi-step prediction has encountered great challenges. The pre-diction accuracy and complexity of the traditional prediction model are low,so we propose a time series prediction model based on combi-nation of the long-short time memory neural network model and the dynamic Bayesian network （DBN）. And we research and prove anoptimal estimation theorem,and on the basis we can get the optimal prediction estimation. The recursion-based graph model is used toenhance prediction performance through probability inference. A new graph model called LSTM-DBN generated from a combination ofLSTM prediction and DBN is developed to predict series data. The simulation shows that the model can improve the accuracy and speedof the sequence prediction and reduce the complexity of the algorithm.