Deep Learning with Long Short-Term Memory for Time Series Prediction
Deep Learning with Long Short-Term Memory for Time Series Prediction
复制标题
使用长短期记忆进行时间序列预测的深度学习
DOI：
10.1109/mcom.2019.1800155
复制
发表时间：
2019-06-01
影响因子：
11.2
通讯作者：
Zhang, Honggang
中科院分区：
文献类型：
--
作者：
Hua, Yuxiu;Zhao, Zhifeng;Zhang, Honggang
时间序列预测可以概括为从历史记录中提取有用信息，然后确定未来值的过程。学习嵌入在时间序列中的长程依赖关系通常是大多数算法的障碍，而LSTM解决方案作为深度学习中的一种特定方案，有望有效地克服这个问题。在本文中，我们首先简要介绍了LSTM的结构和前向传播机制。然后，为了降低LSTM的计算成本，我们提出了一个RCLSTM模型，通过引入随机连接到传统的LSTM神经元。因此，RCLSTM表现出一定程度的稀疏性，并导致计算复杂度的降低。在电信网络领域，流量和用户移动性的预测可以直接受益于这一改进，因为我们利用一个现实的数据集来表明，对于RCLSTM，与LSTM相当的预测性能是可用的，而所需的计算时间要少得多。我们强烈认为RCLSTM在延迟严格或功率受限的应用场景中比LSTM更有能力。
Time series prediction can be generalized as a process that extracts useful information from historical records and then determines future values. Learning long-range dependencies that are embedded in time series is often an obstacle for most algorithms, whereas LSTM solutions, as a specific kind of scheme in deep learning, promise to effectively overcome the problem. In this article, we first give a brief introduction to the structure and forward propagation mechanism of LSTM. Then, aiming at reducing the considerable computing cost of LSTM, we put forward a RCLSTM model by introducing stochastic connectivity to conventional LSTM neurons. Therefore, RCLSTM exhibits a certain level of sparsity and leads to a decrease in computational complexity. In the field of telecommunication networks, the prediction of traffic and user mobility could directly benefit from this improvement as we leverage a realistic dataset to show that for RCLSTM, the prediction performance comparable to LSTM is available, whereas considerably less computing time is required. We strongly argue that RCLSTM is more competent than LSTM in latency-stringent or power-constrained application scenarios.