[1I4-J-2] Machine learning: new modeling

[1I4-J-2-03] Evaluation Method of Major Factors in Long-term Prediction with Exogenous Variables

Yuki Nakatsuka¹, 〇Susumu Shirayama¹ (1. Univ. of Tokyo)

Keywords:Long-term Prediction, Evaluation Method of Major Factors, Attention Mechanism

In many fields, long-term prediction on time series data with exogenous variables has been performed. However, it is difficult to deal with some noise in input data and to qualitatively understand the obtained results. In this paper, first, we extend the DA-RNN proposed by Qin et al so that it can be used in the different time range from the original method. Also, the DA-RNN is extended to be applicable to long-term prediction. Second, we develop a new method of long term prediction based on the extended DA-RNN and the subsequence time-series clustering. An evaluation method of major factors in the exogenous variables is proposed by visualizing the weight of attention mechanism. We tested our method using dataset named SML2010. It is shown that our method has high prediction ability and robustness against noise, accountability of the results.