深圳大学学报理工版

工程应用中的时间序列多为非线性、非平稳序列,直接对其进行预测难度较大.本研究通过经验模态分解算法将原始时间序列分解为多个相对平稳,并具有不同特征尺度的本征模态函数及趋势项,在一定程度上降低时间序列的复杂程度; 同时,在预测过程中,针对递归神经网络模型难以训练及梯度消失等问题,引入长短期记忆网络算法.利用长短期记忆网络算法对分解的本征模态函数分量及趋势项进行分别预测,叠加预测结果得到最终预测结果.以中国北京市PM2.5浓度为例进行预测分析,并将本预测算法与单一预测算法进行比较,结果表明,所提方法具有更高的模型预测精度,达到预测要求.

The time series in engineering applications are mostly non-stationary and non-linear, which are difficult to be directly predicted. Based on the empirical model decomposition(EMD)method, we decompose the original time series into a number of intrinsic mode functions(IMFs)and trend series with the different feature scales in order to reduce the complexity of time series. Meanwhile, in the prediction process, in order to solve the problems of training difficulty and the gradient disappearance in recurrent neural network(RNN)model, we introduce a long-short term memory(LSTM)network algorithm to predict both the results of the decomposed IMF components and trend series respectively, and obtain the final prediction result by surposing the respective prediction results. Taking the PM2.5 concentration in Beijing as an example for prediction and analysis, we compare our prediction algorithm with the single prediction algorithm. The results show that the proposed prediction model has higher accuracy and can meet the prediction requirements.

引言
1 长短期记忆网络
2 EMD-LSTM算法
3 仿真分析
4 结语

图1 递归神经网络神经元[9]<br/>Fig.1 Structure of recurrent neurons[9]

图1 递归神经网络神经元[9]
Fig.1 Structure of recurrent neurons[9]

图2 长短期记忆网络<br/>Fig.2 Long-short term memory network

图2 长短期记忆网络
Fig.2 Long-short term memory network

图3 EMD-LSTM预测流程图<br/>Fig.3 Flow chart of EMD-LSTM

图3 EMD-LSTM预测流程图
Fig.3 Flow chart of EMD-LSTM

图4 PM2.5浓度数据序列<br/>Fig.4 The data series for the concentration of PM2.5

图4 PM2.5浓度数据序列
Fig.4 The data series for the concentration of PM2.5

图5 原始序列的EMD分解结果<br/>Fig.5 The decomposition result of original series by EMD

图5 原始序列的EMD分解结果
Fig.5 The decomposition result of original series by EMD

图6 序列的排列熵<br/>Fig.6 Permutation entropy of series

图6 序列的排列熵
Fig.6 Permutation entropy of series

表1 各序列的预测误差<br/>Table 1 Prediction errors of series

表1 各序列的预测误差
Table 1 Prediction errors of series

图7 真实值及对比预测结果<br/>Fig.7 The original data and the contrast of prediction results

图7 真实值及对比预测结果
Fig.7 The original data and the contrast of prediction results

表2 四种方法的预测指标<br/>Table 2 Prediction errors of four algorithms

表2 四种方法的预测指标
Table 2 Prediction errors of four algorithms

[1] 陈艳华. 基于机器学习的时间序列预测方法研究与应用[D]. 兰州:兰州大学, 2017.
[2] 李祥, 彭玲, 邵静, 等. 基于小波分解和ARMA模型的空气污染预报研究[J]. 环境工程, 2016, 34(8):110-113.
[3] 牟敬锋, 赵星, 樊静洁, 等. 基于ARIMA模型的深圳市空气质量指数时间序列预测研究[J]. 环境卫生学杂志, 2017, 7(2):102-107.
[4] 王威, 陈熙源. 一种基于多尺度和改进支持向量机的光纤陀螺温度漂移建模与补偿方法[J]. 中国惯性技术学报, 2016,24(6): 793-797.
[5] LI Guohui, MA Xiao, YANG Hong. A hybrid model for monthly precipitation time series forecasting based on variational mode decomposition with extreme learning machine[J]. Information, 2018, 9(7): 1-13.
[6] 丁亮, 武林, 刘清民. 基于边缘保持中值滤波和BP神经网络的自然手势识别研究[J]. 现代计算机, 2017(25): 10-13.
[7] LUKOEVIACˇG3IUS M, JAEGER H. Reservoir computing approaches to recurrent neural network training[J]. Computer Science Review, 2009, 3(3):127-149.
[8] 毛景慧. 基于LSTM深度神经网络的股市时间序列预测精度的影响因素研究[D]. 广州:暨南大学, 2017.
[9] LEV G, SADEH G, KLEIN B, et al. RNN fisher vectors for action recognition and image annotation[J]. Computer Science, 2016: 833-850.
[10] FU Rui, ZHANG Zuo, LI Li. Using LSTM and GRU neural network methods for traffic flow prediction[C]// The 31st Youth Academic Annual Conference of Chinese Association of Automation. Wuhan, China: IEEE, 2017: 324-328.
[11] ZHANG Boning. Foreign exchange rates forecasting with an EMD-LSTM neural networks model[J]. Journal of Physics Conference, 2018, 1053(1): 5-12.
[12] 杨训政, 柯余洋, 梁肖, 等. 基于LSTM的发电机组污染物排放预测研究[J]. 电气自动化, 2016, 38(5): 22-25.
[13] HUANG M L, LONG S E, SHIH H H, et al. The empirical mode decomposition and the Hilbert spectrum for nonlinear and non-stationary time series analysis[J]. Proceedings: Mathematical, Physical and Engineering Sciences, 1998, 454(1971): 903-993.
[14] HUANG Yiming, WU Di, ZHANG Zhifen, et al. EMD-based pulsed TIG welding process porosity defect detection and defect diagnosis using GA-SVM[J]. Journal of Materials Processing Technology, 2017, 239: 92-102.
[15] 秦喜文, 刘媛媛, 王新民, 等. 基于整体经验模态分解和支持向量回归的北京市 PM2.5预测[J]. 吉林大学学报地球科学版, 2016, 46(2): 563-568.

_2.5

[16] 刘媛媛. 基于整体经验模态分解和随机森林的城市PM2.5预测[D]. 长春:长春工业大学, 2017.

_2.5

[17] WILLMOTT C J, MATSUURA K. Advantages of the mean absolute error(MAE)over the root mean square error(RMSE)in assessing average model performance[J]. Climate Research, 2005, 30(1): 79-82.
[18] 冯辅周, 饶国强, 司爱威, 等. 排列熵算法的应用与发展[J]. 装甲兵工程学院学报, 2012, 26(2): 34-38.

备注

引言

1 长短期记忆网络

2 EMD-LSTM算法

3 仿真分析

4 结语

期刊信息

备注

引言

1 长短期记忆网络

2 EMD-LSTM算法

3 仿真分析

4 结 语

期刊信息

4 结语