基于深度置信网络的时间序列预测

Time series prediction based on deep confidence network

针对传统计算机神经网络存在梯度弥散、局部最小值、非线性时间序列长期预测性能不佳和高维序列数据复杂度高等问题,提出时序深度置信网络模型(timing deep belief network model,T-DBN).该模型预训练阶段采用改进的贪婪预训练算法,在预训练过程中使用梯度修正并行回火(gradient fixing parallel tempering,GFPT)算法,采用重构误差确定网络深度,在反向调整阶段采用拟牛顿法(BFGS算法),以获得更加准确的预测精度.结合相空间重构理论和BP(back propagation)神经网络,对中国江西省2016—2020年农业机械总动力进行了预测.针对高非线性的股票数据,提取同花顺软件1990-12-20—2018-03-30时间段内的上证指数特征信息,分别采用T-DBN、DBN和长短期记忆(long short-term memory,LSTM)模型进行股票预测,预测准确率分别为79.3%、77.9%和74.6%,T-DBN模型的预测准确率高于DBN和LSTM模型.

The timing deep belief network model(T-DBN)is proposed to solve the problems of gradient dispersion,local minimum,poor long-term prediction performance of nonlinear time series prediction and high complexity of high-dimensional sequence data in traditional computer neural networks.In the pre-training stage,the improved greedy pre-training algorithm is adopted.In the pre-training process,the gradient fixing parallel tempering(GFPT)algorithm is used to determine the depth of network by reconstruction error.In the stage of reverse adjustment,the quasi-newton method(BFGS)is adopted to obtain more accurate prediction performance.Combined with the theory of phase space reconstruction and BP(back propagation)neural network,the total power of agricultural machinery in Jiangxi province of China from 2016 to 2020 is predicted.For the highly nonlinear stock data,the characteristics of the Shanghai stock exchange index from 1990-12-20 to 2018-03-30 are extracted from the stock data of Tonghuashun software.T-DBN,DBN and long short-term memory(LSTM)models areused for stock forecasting with the prediction accuracy of 79.3%,77.9%and 74.6%,respectively.The experimental results demonstrate the better prediction performance of T-DBN model in comparison with DBN and LSTM models.