基于集成深度学习的时间序列预测模型

基于高斯过程的条件受限玻尔兹曼机(GCRBM)时序模型可以很好的预测单一种类时序数据,但是该模型难以预测多类别的真实高维数据。针对这个问题,提出基于集成深度学习的时间序列预测模型,对多类时序对应训练多个深可信网络(deep belief networks, DBN)模型来学习低维特征,利用低维特征对应训练多个GCRBM时序模型。预测时序时先通过训练出的一组DBN模型对目标数据进行降维并通过重建误差识别类别,然后通过识别到的类别所对应的GCRBM模型预测目标数据的后期时序。在CASIA-A步态数据集上的试验结果表明:本方法能够准确识别出步态序列,而且预测结果能够模拟出真实的步态序列,证实了本模型的有效性。

The time series prediction model based on integrated deep learning

The conditional restricted Boltzmann machine time series model based on the Gaussian process(GCRBM)could efficiently predict single type of time series data, but the model could not make accurate predictions for multi-category data and real high-dimensional data. To solve the problem above, the time series prediction model based on integrated deep learning was proposed. Multiple deep belief networks(DBN)corresponding to the multi-category timing data was trained to study low dimensional feature. The low dimensional feature of multi-category data was used to train multiple GCRBM models. When the time series was predicted, the dimensionality of the model was reduced and categories of target data were identified by DBN model's reconstruction error, and the sequence of target data was predicted by the GCRBM model. The experimental results based on CASIA-A gait data set showed that the method could accurately recognize the categories of gait sequences and the predicting result could simulate the true gait sequences, which demonstrated the validity of the model.