摘要
针对标准长短期记忆(long short-term memory,LSTM)神经网络用于时间序列预测具有耗时长、复杂度高等问题,提出简化型LSTM神经网络并应用于时间序列预测.首先,通过耦合输入门与遗忘门实现对标准LSTM神经网络的结构简化;其次,从门结构控制方程中消除输入信号与偏差实现进一步精简;然后,采用梯度下降算法更新简化型LSTM神经网络的参数;最后,通过2个时间序列基准数据集及污水处理过程出水生化需氧量(biochemical oxygen demand,BOD)质量浓度预测进行实验验证.结果表明:在不显著降低预测精度的情况下,所设计的模型能够缩短训练时间,减少LSTM神经网络的计算复杂度,实现时间序列的预测.
To solve the problem that the standard long short-term memory(LSTM)neural network is time consuming and has high complexity for time series prediction,a simplified LSTM neural network was proposed and it was applied to time series prediciton.First,the structure of the standard LSTM neural network was simplified by coupling input gate and forget gate.Second,the inputs and bias were removed from dynamic equation of the gates to further simplify the parameters.Third,the gradient descent algorithm was utilized to update the parameters of the simplified LSTM neural network.Finally,the validity of the proposed model was demonstrated by two time series benchmark problems and the prediction of biochemical oxygen demand(BOD)mass concentration in the wastewater treatment process.The experimental results show that the training time is shortened and the computational complexity is reduced without significantly reducing the prediction accuracy,which makes an efficient time series prediction.
作者
李文静
王潇潇
LI Wenjing;WANG Xiaoxiao(Faculty of Information Technology,Beijing University of Technology,Beijing 100124,China;Beijing Key Laboratory of Computational Intelligence and Intelligent System,Beijing 100124,China;Beijing Laboratory for Intelligent Environmental Protection,Beijing 100124,China;Beijing Artificial Intelligence Institute,Beijing 100124,China)
出处
《北京工业大学学报》
CAS
CSCD
北大核心
2021年第5期480-488,共9页
Journal of Beijing University of Technology
基金
国家自然科学基金资助项目(61603009,62021003,61890930-5)
国家重点研发计划资助项目(2018YFC1900800-5)
北京市教育委员会科技计划资助项目(KM201910005023)。
