Á¤º¸°úÇÐȸ ÄÄÇ»ÆÃÀÇ ½ÇÁ¦ ³í¹®Áö (KIISE Transactions on Computing Practices)
Current Result Document :
ÇѱÛÁ¦¸ñ(Korean Title) |
Ç×°ø±â ±ËÀû ¿¹ÃøÀ» À§ÇÑ LSTM-based Multi-head AttentionÀ» È°¿ëÇÑ Recurrent ³×Æ®¿öÅ© |
¿µ¹®Á¦¸ñ(English Title) |
Recurrent Network using LSTM-based Multi-head Attention for Aircraft Trajectory Prediction |
ÀúÀÚ(Author) |
±è¸í¼ö
ÃÖ¿øÀÍ
Myoungsoo Kim
Wonik Choi
|
¿ø¹®¼ö·Ïó(Citation) |
VOL 28 NO. 03 PP. 0146 ~ 0152 (2022. 03) |
Çѱ۳»¿ë (Korean Abstract) |
ÃÖ±Ù ¸¹Àº ¿¬±¸µé¿¡¼ µö·¯´× ±â¹ýÀ» Àû¿ëÇÏ¿© Ç×°ø ±³Åë È帧À» ºÐ¼®Çϰųª Ç×°ø±â ±ËÀûÀ» ¿¹ÃøÇÏ´Â ¹æ¹ýµéÀ» Á¦¾ÈÇÏ°í ÀÖ´Ù. º» ³í¹®¿¡¼´Â ½Ã°£¿¡ µû¸¥ À§µµ, °æµµ, °íµµ µ¥ÀÌÅ͸¦ È¿°úÀûÀ¸·Î ÇнÀÇÏ¿© Ç×°ø±â ±ËÀû ¿¹ÃøÀ» ¼öÇàÇϱâ À§ÇÏ¿© LSTM-based Multi-head AttentionÀ» È°¿ëÇÑ recurrent ³×Æ®¿öÅ©¸¦ Á¦¾ÈÇÑ´Ù. ÀÔ·Â µ¥ÀÌÅÍ¿¡ ´ëÇÏ¿© attentionÀ» Àû¿ëÇÏ´Â ºÎºÐ¿¡¼ LSTMÀ» »ç¿ëÇÔÀ¸·Î½á key, query, value¸¦ »ý¼ºÇÏ´Â °úÁ¤¿¡¼ °ú°ÅÀÇ À§µµ, °æµµ, °íµµ °ªÀÇ Á߿伺À» ¿¹Ãø ¸ðµ¨¿¡ °Á¶Çϵµ·Ï ÇÏ¿´´Ù. ½Ç¼¼°è µ¥ÀÌÅ͸¦ ÀÌ¿ëÇÑ ±¤¹üÀ§ÇÑ ½ÇÇè °á°ú, º» ³í¹®¿¡¼ Á¦¾ÈÇÏ´Â ¸ðµ¨ÀÌ Æ¯Á¤ Ç×°øÆí¿¡ ´ëÇؼ multi-head attention ±â¹ÝÀÇ Bi-LSTM ¸ðµ¨º¸´Ù °³¼±µÈ °á°ú¸¦ º¸ÀÓÀ» ¾Ë ¼ö ÀÖ¾ú´Ù. À§µµ ¿¹ÃøÀÇ °æ¿ì Á¦¾ÈÇÏ´Â ¸ðµ¨ÀÌ KAL1253Æí¿¡ ´ëÇؼ MAE ±âÁØ 29%ÀÇ ¿ÀÂ÷¸¦ ÁÙ¿´°í, KAL1209Æí¿¡ ´ëÇؼ RMSE±âÁØ 17%ÀÇ ¿ÀÂ÷¸¦ °¨¼Ò½ÃŲ °á°ú¸¦ º¸¿©ÁÖ¾ú´Ù. °æµµ ¿¹Ãø¿¡¼´Â KAL1253Æí¿¡ ´ëÇؼ MAE ±âÁØ 83%, KAL1257Æí¿¡ ´ëÇؼ RMSE±âÁØ 82%ÀÇ ¿ÀÂ÷°¡ °¨¼ÒÇÑ °á°ú¸¦ º¸¿©ÁÖ¾ú´Ù. |
¿µ¹®³»¿ë (English Abstract) |
Recently, many studies have proposed methods to analyze air traffic flow or predict aircraft trajectories by applying deep learning techniques. In this paper, we propose a recurrent network using LSTM-based Multi-head Attention to effectively learn latitude, longitude, and altitude data over time to perform aircraft trajectory predictions. Specifically, we exploit LSTM to focus on the importance of past latitude, longitude, and altitude values in the process of generating keys, queries, and values. Extensive experiments using real-world datasets show that the proposed model outperforms the multi-head attention-based Bi-LSTM model for specific flights. In the case of latitude prediction, the proposed model reduces the MAE error by 29% for the KAL1253 flight and the RMSE error by 17% for the KAL1209 flight. In the case of longitude prediction, the MAE error for the KAL1253 flight is reduced by 83%, and the RMSE error for the KAL1257 flight is reduced by 82%. |
Å°¿öµå(Keyword) |
Ç×°ø±â ±ËÀû ¿¹Ãø
Multi-Head Attention
LSTM
Bidirectional LSTM
aircraft trajectory prediction
multi-head attention
bidirectional-LSTM
|
ÆÄÀÏ÷ºÎ |
PDF ´Ù¿î·Îµå
|