Mobile QR Code QR CODE : Journal of the Korean Institute of Illuminating and Electrical Installation Engineers

Journal of the Korean Institute of Illuminating and Electrical Installation Engineers

ISO Journal TitleJ Korean Inst. IIIum. Electr. Install. Eng.

  1. (Researcher, Data Science Lab., KEPCO, Seoul, Korea)
  2. (Assistant Vice President, Head of Distribution Planning Team, KEPCO, Naju, Korea)



ARIMA, Graph neural networks, Load forecasting, LSTM, Switch

1. ์„œ ๋ก 

์ „ ์„ธ๊ณ„์ ์œผ๋กœ ์—๋„ˆ์ง€ ์ „ํ™˜ ์‹œ๊ธฐ๊ฐ€ ๋„๋ž˜ํ•จ์— ๋”ฐ๋ผ ๋ฐฐ์ „๊ณ„ํ†ต์— ์ ‘์†๋˜๋Š” ํƒœ์–‘๊ด‘, ํ’๋ ฅ ๋“ฑ ์žฌ์ƒ์—๋„ˆ์ง€๊ฐ€ ๋งŽ์•„์ง€๊ณ  ์žˆ๋‹ค. ์ด์— ๊ณต๊ธ‰๊ณ„ํ†ต์˜ ๊ทœ๋ชจ๊ฐ€ ํ™•๋Œ€๋˜์–ด ๊ฐ€๊ณ  ์žˆ์œผ๋ฉฐ, ๊ณ„ํ†ต ๊ฐ„ ์—ฐ๊ณ„ ๋“ฑ์˜ ์ด์œ ๋กœ ๋ถˆํ™•์‹ค์„ฑ๊ณผ ๊ฐ„ํ—์„ฑ์ด ํฐ ์žฌ์ƒ์—๋„ˆ์ง€ ๋ฐœ์ „์˜ ํŠน์„ฑ์ƒ ๋ฐฐ์ „๊ณ„ํ†ต ์šด์˜์˜ ์•ˆ์ „์„ฑ์ด ์ค„์–ด๋“ค๊ณ  ๋ณต์žก์„ฑ์ด ์ฆ๊ฐ€ํ•˜๊ณ  ์žˆ๋‹ค. ๋”ฐ๋ผ์„œ ํšจ์œจ์ ์ด๊ณ  ์•ˆ์ •์ ์ธ ์ „๋ ฅ๊ณต๊ธ‰์„ ์œ„ํ•œ ๋น ๋ฅด๊ณ  ์ •ํ™•ํ•œ ์‹œ๊ฐ„๋ณ„ ๊ฐ์‹œ/์ œ์–ด๊ฐ€ ์š”๊ตฌ๋œ๋‹ค. ์ด๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ๋ฐฐ์ „์ž๋™ํ™”์Šคํ…œ(DAS)์„ 1997๋…„๋ถ€ํ„ฐ ์ „๋ ฅ์—ฐ๊ตฌ์›์— ์˜ํ•ด ๊ฐœ๋ฐœ ๋ฐ ์šด์˜๋˜๊ณ  ์žˆ๋‹ค(1,2). ์ด ์‹œ์Šคํ…œ์€ ๊ณ„ํ†ต์˜ ๋ฐฐ์ „์„ ๋กœ(22.9kV) ๊ฐœํ๊ธฐ ์ „์••ยท์ „๋ฅ˜ ์„ผ์„œ์—์„œ ์ทจ๋“ํ•œ ์‹ค์‹œ๊ฐ„ ์ •๋ณด๋ฅผ ์ „์„ ๋กœ ๋‹จ๋ง์žฅ์น˜ FRTU(Feeder Remote Terminal Unit)๋ฅผ ํ†ตํ•ด ์ค‘์•™์ œ์–ด์žฅ์น˜์— ์ „์†กํ•˜๊ณ , ์ธก์ •๋œ ์ •๋ณด๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ตฌ๊ฐ„๋ถ€ํ•˜๋ฅผ ๊ด€๋ฆฌํ•œ๋‹ค. ํŠนํžˆ, ๊ฐ ๊ตฌ๊ฐ„์˜ ํ”ผํฌ๋ถ€ํ•˜์˜ ๋ฐœ์ƒ์‹œ๊ฐ„์€ ์„œ๋กœ ์ƒ์ดํ•˜์—ฌ, ์‹œ๊ฐ„์— ๋”ฐ๋ฅธ ๋ถ€ํ•˜๊ณก์„ ์˜ ํ˜•ํƒœ๋กœ ๊ด€๋ฆฌํ•˜๋Š” ๊ฒƒ์€ ์ค‘์š”ํ•˜๋ฉฐ(3), ๋‚˜์•„๊ฐ€ ๋น ๋ฅด๊ณ  ์ •ํ™”ํ•œ ์˜ˆ์ธก์„ ํ†ตํ•œ ์„ ์ œ์  ์˜ˆ๋ฐฉ๋Šฅ๋ ฅ๊ณผ ๊ฐœํ๊ธฐ๋ฅผ ๊ธฐ์ค€์œผ๋กœ ๊ตฌ๊ฐ„๋ถ€ํ•˜๋ฅผ ์‚ฐ์ •ํ•  ํ•„์š”๊ฐ€ ์žˆ๋‹ค. ์ด๋ฅผ ์œ„ํ•ด ๊ฐœํ๊ธฐ๋ณ„ ๋‹จ๊ธฐ๋ถ€ํ•˜์˜ˆ์ธก์— ๊ด€ํ•œ ์—ฐ๊ตฌ๊ฐ€ ํ•„์š”ํ•˜๋‹ค. ์ด์— ๊ด€ํ•œ ์œ ์‚ฌ ์—ฐ๊ตฌ์—๋Š” ์‹ ๊ฒฝ๋ง ๊ธฐ๋ฐ˜ ํ‰๊ท ๋ถ€ํ•˜์™€ ์ตœ๋Œ€๋ถ€ํ•˜๋ฅผ ํ†ตํ•œ ๊ตฐ์ง‘๋ณ„ ์˜ˆ์ธก ์—ฐ๊ตฌ(4-6), ๊ทธ๋ฆฌ๊ณ  ARIMA์™€ LSTM์„ ํ™œ์šฉํ•œ ๊ธฐ๋ฒ•(7,8) ๋“ฑ์œผ๋กœ ์ง€์—ญ ๋˜๋Š” ํ™˜๊ฒฝ์— ๋”ฐ๋ผ ๋…๋ฆฝ์ ์ธ ์˜ˆ์ธก๋ชจ๋ธ์ด ํ•„์š”ํ•˜๋ฉฐ, ๋ณธ ์—ฐ๊ตฌ์˜ ๊ฐœํ๊ธฐ๋ณ„ ์ ‘๊ทผ๊ณผ๋Š” ์ฐจ์ด๊ฐ€ ์žˆ๋‹ค.

๋ณธ ๋…ผ๋ฌธ์€ ์—ฐ์‚ฐ์ž์›์„ ํšจ์œจ์ ์œผ๋กœ ํ™œ์šฉํ•˜์—ฌ, ๋ฐฐ์ „๋ง์—์„œ D/L ๋‚ด ๋ชจ๋“  ์ž๋™ํ™” ๊ฐœํ๊ธฐ๋ณ„ ๋ถ€ํ•˜์ •๋ณด์— ๋Œ€ํ•ด ํ•™์Šต์†๋„์™€ ์˜ˆ์ธก์„ฑ๋Šฅ์„ ๊ณ ๋ คํ•œ ๋‹จ๊ธฐ์˜ˆ์ธก ๋ชจ๋ธ๊ฐœ๋ฐœ์„ ๋ชฉํ‘œ๋กœ ํ•œ๋‹ค. ์ด๋ฅผ ์œ„ํ•ด ์ตœ๊ทผ ๊ตํ†ต๋Ÿ‰, ์‚ฌํšŒ๊ด€๊ณ„๋ง ๋“ฑ์˜ ๋ถ„์•ผ์—์„œ ํ™œ์šฉ๋˜๋Š”(9-11) ๊ทธ๋ž˜ํ”„ ๊ธฐ๋ฐ˜ ์˜ˆ์ธก๋ชจ๋ธ์„ ํ™œ์šฉํ•œ๋‹ค. ๋ฐฐ์ „๊ณ„ํ†ต์—์„œ ์ƒํ˜ธ์—ฐ๊ฒฐ๋œ ๋ถ€ํ•˜์ •๋ณด๋ฅผ ํ™œ์šฉํ•˜์—ฌ, ๊ทธ๋ž˜ํ”„ ์‹ ๊ฒฝ๋ง ๋ชจ๋ธ์ธ, ST-GCNs ๊ธฐ๋ฐ˜ ๋‹จ๊ธฐ๋ถ€ํ•˜์˜ˆ์ธก ๋ชจ๋ธ์— ์ ์šฉํ•œ๋‹ค. ๋˜ํ•œ, ํ•™์Šต์„ ์œ„ํ•œ ๋ฐ์ดํ„ฐ ์ƒ์„ฑ์„ ์œ„ํ•ด ์ „๋ ฅ๊ณ„ํ†ต์—์„œ D/L ๋‚ด ๊ฐœํ๊ธฐ ๊ฐ„ ์—ฐ๊ฒฐ์ •๋ณด์—์„œ ๋ถ€ํ•˜์ •๋ณด ์ถ”์ถœ์ด ๊ฐ€๋Šฅํ•œ ์ž๋™ํ™” ๊ฐœํ๊ธฐ ๊ฐ„ ๋ถ€๋ถ„์—ฐ๊ฒฐ์ •๋ณด๋กœ ์ถ”์ถœํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์ œ์‹œํ•œ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ, ๋น„๊ต ๋ชจ๋ธ์„ ํ†ตํ•ด ์„ฑ๋Šฅ๊ณผ ํ•™์Šต์†๋„๋ฅผ ํ™•์ธํ•œ๋‹ค.

๋…ผ๋ฌธ์˜ ๊ตฌ์„ฑ์€ ๋‹ค์Œ๊ณผ ๊ฐ™๋‹ค. 2์žฅ์—์„œ๋Š” ์—ฐ๊ตฌ์— ํ•„์š”ํ•œ ๋ฐ์ดํ„ฐ ์ถ”์ถœ์„ ์œ„ํ•ด ์‚ฌ์šฉํ•œ ๋ฐฐ์ „์ž๋™ํ™” ์‹œ์Šคํ…œ(DAS) ๋ฐ ๋ฐ์ดํ„ฐ ์ „์ฒ˜๋ฆฌ ๊ณผ์ •์„ ๋‹ค๋ฃฌ๋‹ค. 3์žฅ์—์„œ๋Š” ๊ทธ๋ž˜ํ”„ ์‹ ๊ฒฝ๋ง ๊ธฐ๋ฐ˜ ์ถ”๋ก  ๋ชจ๋ธ์ธ ST-GCNs๊ณผ ๋น„๊ต ๋ชจ๋ธ์„ ์„ค๋ช…ํ•œ๋‹ค. 4์žฅ์—์„œ๋Š” ๊ทธ๋ž˜ํ”„ ๊ธฐ๋ฐ˜ ๋ถ€ํ•˜ ์˜ˆ์ธก๋ชจ๋ธ์— ์ ์šฉํ•œ ๊ฒฐ๊ณผ๋ฅผ ๋น„๊ต ๋ชจ๋ธ๊ณผ ๋น„๊ตํ•˜๋ฉฐ, ๋ชจ๋ธ์˜ ์žฅ์ ๊ณผ ํ•œ๊ณ„์ ์— ๊ด€ํ•ด ์„œ์ˆ ํ•œ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ 5์žฅ ํ–ฅํ›„ ์—ฐ๊ตฌ ๋ฐฉํ–ฅ์— ๋Œ€ํ•ด ๋…ผ์˜ํ•œ๋‹ค.

2. DAS ๋ฐ์ดํ„ฐ ์ „์ฒ˜๋ฆฌ

ํ˜„์žฌ DAS๋Š” ๊ณ„ํ†ต์— ์„ค์น˜๋œ FRTU์™€ ๊ฐœํ๊ธฐ ํ†ต์‹  ์—ฐ๊ณ„๊ธฐ๋Šฅ ๋ชจ๋“ˆ์ธ ์ „๋‹จ์ฒ˜๋ฆฌ๊ธฐ๋ฅผ ํ†ตํ•ด์„œ ์ „์••ยท์ „๋ฅ˜ ๊ณ„์ธก๊ฐ’์„ ์ทจ๋“ํ•˜๊ณ , ์ œ์–ด์‹ ํ˜ธ๋ฅผ ์ „๋‹ฌํ•œ๋‹ค. ์ด๋ฅผ HMI๋ฅผ ํ™œ์šฉํ•˜์—ฌ, ๋ฐฐ์ „๊ณ„ํ†ต์˜ ์—ฐ๊ฒฐ์ •๋ณด๋ฅผ ๋‹จ์„ ๋„์™€ ๊ณ„ํ†ต๋„ ํ˜•ํƒœ๋กœ Fig. 1๊ณผ ๊ฐ™์ด ์ œ๊ณตํ•œ๋‹ค. ๋˜ํ•œ, ์ทจ๋“๋œ ๊ณ„์ธก๊ฐ’์€ ๋ณ„๋„์˜ ์„œ๋ฒ„์— ์ €์žฅํ•˜๊ณ  ์žˆ๋‹ค.

Fig. 1. Distribution system HMI (single-line diagram)
../../Resources/kiiee/JIEIE.2022.36.1.037/fig1.png

Fig. 2๋Š” 2019๋…„ 8์›”๋ถ€ํ„ฐ 2021๋…„ 8์›” ๊ฐ„ ์ถฉ๋ถ/์˜ค์†ก/๊ณต๋ถ-D/L ๋‚ด ๊ฐœํ๊ธฐ๋“ค์˜ ์—ฐ๊ฒฐ์ •๋ณด์™€ ๊ฐ๊ฐ์˜ ์‹œ๊ฐ„๋‹น A์ƒ ์ „๋ฅ˜ ์‹œ๊ณ„์—ด ๋ฐ์ดํ„ฐ๋ฅผ ๋‚˜ํƒ€๋‚ธ๋‹ค. ์—ฌ๊ธฐ์„œ ๊ทธ๋ฆผ ์ขŒ์ธก์˜ ์ˆซ์ž๋Š” ๊ฐœํ๊ธฐID, ํŒŒ๋ž€์›์€ ๋ฐ์ดํ„ฐ๋ฅผ ์ทจ๋“ํ•  ์ˆ˜ ์žˆ๋Š” ์ž๋™ํ™” ๊ฐœํ๊ธฐ๋ฅผ ์˜๋ฏธํ•˜๋ฉฐ, ๋‹ค์„ฏ ๊ฐœ์˜ ์ ์ƒ‰์›์œผ๋กœ ํ‘œ์‹œ๋œ ๊ฐœํ๊ธฐ์˜ A์ƒ ์ „๋ฅ˜ ํŒจํ„ด์„ ๊ทธ๋ฆผ ์šฐ์ธก ๊ทธ๋ž˜ํ”„๋กœ ์ƒํ•˜ ์ˆœ์ฐจ์ ์œผ๋กœ ๋ณด์—ฌ์ค€๋‹ค.

Fig. 2. Graph representation of connection information in D/L and load per hour of automatic switches
../../Resources/kiiee/JIEIE.2022.36.1.037/fig2.png

์ „์ฒ˜๋ฆฌ ๊ณผ์ •์€ ์‹œ๊ณ„์—ด ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ๋ˆ„๋ฝ ์ •๋ณด๋ฅผ ์ฒ˜๋ฆฌํ•˜๊ณ , ๊ทธ๋ž˜ํ”„ ๊ตฌ์กฐ๋ฅผ ๋„์ž…ํ•˜๋Š” ๊ณผ์ •์œผ๋กœ, ์ตœ์ข…์ ์ธ ๋ฐ์ดํ„ฐ์˜ ํ˜•ํƒœ๋Š” Fig. 3์˜ ๊ทธ๋ž˜ํ”„ ๊ตฌ์กฐ์˜ ์‹œ๊ณ„์—ด ๋ฐ์ดํ„ฐ์™€ ๊ฐ™์ด ํ‘œํ˜„๋œ๋‹ค. $I_{t}$๋Š” $t$์‹œ๊ฐ„์—์„œ ์ „๋ฅ˜ ์‹œ๊ณ„์—ด์€ ๊ทธ๋ž˜ํ”„ ํ˜•ํƒœ์˜ ๋ฐ์ดํ„ฐ ํ–‰๋ ฌ(matrix) ํ˜•ํƒœ์ด๋‹ค.

Fig. 3. Graph structure time series data
../../Resources/kiiee/JIEIE.2022.36.1.037/fig3.png

์—ฐ๊ตฌ์— ์‚ฌ์šฉํ•˜๋Š” D/L์€ ๋‹จ์„ ๋„ ๋‚ด ๊ฐœํ๊ธฐ๋ฅผ ๋…ธ๋“œ(node)๋กœ, ๊ฐœํ๊ธฐ๋งˆ๋‹ค ์—ฐ๊ฒฐ๋œ ๋ฐฐ์„ ์€ ๊ฐ ๋…ธ๋“œ๋ฅผ ์—ฐ๊ฒฐํ•˜๋Š” ๊ฐ„์„ (edge)์œผ๋กœ ๊ตฌ์„ฑ๋œ๋‹ค. ๊ฐœํ๊ธฐ๊ฐ€ ์—ฐ๊ฒฐ๋œ ํ˜•ํƒœ๋ฅผ ๊ฐœํ๊ธฐID๋ฅผ ํ†ตํ•ด, source์™€ target์œผ๋กœ ์ •๋ฆฌํ•œ๋‹ค. ์—ฐ๊ฒฐ๋ง์— ํฌํ•จ๋œ ์ˆ˜๋™๊ณผ ์ž๋™ํ™” ๊ฐœํ๊ธฐ๋ฅผ ๋ถ„๋ฆฌํ•˜์—ฌ ์—ฐ๊ฒฐ์„ ๊ตฌ์„ฑํ•œ๋‹ค. Fig. 4์—์„œ source์™€ target์€ ์ž๋™ํ™” ๊ฐœํ๊ธฐ์˜ ์—ฐ๊ฒฐ์„ ๋‚˜ํƒ€๋‚ด๋ฉฐ, con_pts๋Š” source์™€ target ์‚ฌ์ด์— ํฌํ•จ๋œ ์ˆ˜๋™ ๊ฐœํ๊ธฐID๋ฅผ ๋‚˜ํƒ€๋‚ธ๋‹ค. Fig. 4๋Š” ์ถฉ๋ถ/์˜ค์†ก/๊ณต๋ถ-DL ๋‚ด ์„ค์น˜๋œ ๊ฐœํ๊ธฐ ์—ฐ๊ฒฐ์ •๋ณด๋ฅผ ํ…Œ์ด๋ธ”ํ™” ๋œ ๋ฐ์ดํ„ฐ๋กœ ์ฒ˜๋ฆฌํ•œ ๊ฒฐ๊ณผ์ด๋‹ค.

Fig. 4. Connection information in D/L
../../Resources/kiiee/JIEIE.2022.36.1.037/fig4.png

์‹œ๊ณ„์—ด ๋ฐ์ดํ„ฐ๋ฅผ ์ทจ๋“ํ•  ์ˆ˜ ์žˆ๋Š” ์ž๋™ํ™” ๊ฐœํ๊ธฐ๋ฅผ ๊ตฌ๋ถ„ํ•˜์—ฌ, ์ด๋ฅผ ์ƒ์œ„ ๋…ธ๋“œ๋กœ ์„ค์ •ํ•˜๊ณ  ํ•˜์œ„๋…ธ๋“œ์˜ ์—ฐ๊ฒฐ ์กฐํ•ฉ์„ ๊ฒ€์ƒ‰ํ•˜์—ฌ Fig. 5์™€ ๊ฐ™์ด ์ž๋™ํ™” ๊ฐœํ๊ธฐ ๋…ธ๋“œ(ํŒŒ๋ž€์ ) ์‚ฌ์ด์˜ ๋‹จ์ˆœํ™”๋œ ๋ถ€๋ถ„ ๊ทธ๋ž˜ํ”„(subgraph) ํ˜•ํƒœ๋กœ ๋ณ€ํ™˜ํ•œ๋‹ค.

Fig. 5. Subgraph representation of swtichโ€™s connection information of four D/L
../../Resources/kiiee/JIEIE.2022.36.1.037/fig5.png

๋ถ€๋ถ„ ๊ทธ๋ž˜ํ”„๋กœ ์ฒ˜๋ฆฌ๋œ ๋‹จ์„ ๋„ ์—ฐ๊ฒฐ์ •๋ณด์™€ ๊ฐœํ๊ธฐ ์‹œ๊ณ„์—ด ๋ฐ์ดํ„ฐ๋ฅผ ๊ฒฐํ•ฉํ•˜์—ฌ, ์‹œ๊ณต๊ฐ„์ ์œผ๋กœ ์—ฐ๊ด€๋œ ํ˜•ํƒœ๋กœ ๋ณ€ํ™˜ํ•ด ์˜ˆ์ธก ์ด์ „ 12์‹œ๊ฐ„์œผ๋กœ ์ดํ›„ 1์‹œ๊ฐ„์”ฉ 3์‹œ๊ฐ„์˜ ์ƒ ์ „๋ฅ˜๊ฐ’์„ ์˜ˆ์ธกํ•˜๋Š”๋ฐ ์ ์šฉํ•œ๋‹ค. ์•ž์œผ๋กœ (1/ 2/ 3hr)๋กœ ํ‘œ๊ธฐํ•œ๋‹ค. ๋˜ํ•œ, ์ž๋™ํ™” ๊ฐœํ๊ธฐ์˜ ๋ถ€๋ถ„ ๊ทธ๋ž˜ํ”„๋ฅผ ์•„๋ž˜ ์ˆ˜์‹์„ ํ™œ์šฉํ•˜์—ฌ ์ธ์ ‘ํ–‰๋ ฌ(adjacent matrix) ํ˜•ํƒœ๋กœ ์ฒ˜๋ฆฌํ•œ๋‹ค.

$A_{ij}=\begin{cases} 1&{if}\left\{v_{i},\: v_{j}\right\}\in E{and}i\ne j,\: \\ 0& otherwise \end{cases}$

์—ฌ๊ธฐ์„œ $v$ ๋Š” ๋…ธ๋“œ, $E$ ๋Š” ๊ฐ„์„  ์ง‘ํ•ฉ์„ ๋‚˜ํƒ€๋‚ธ๋‹ค. ๊ฐ„๋‹จํ•œ ์˜ˆ์‹œ๋กœ ์ถฉ๋ถ/์˜ค์†ก/๊ณต๋ถ-D/L์˜ ์—ฐ๊ฒฐ์„ฑ์€ ์•„๋ž˜์™€ ๊ฐ™์€ Table 1์˜ ํ˜•ํƒœ๋กœ ์ •๋ฆฌ๋˜๋ฉฐ, ํ–‰๊ณผ ์—ด์˜ ์ฒซ ๋ฒˆ์งธ๋Š” ๊ฐœํ๊ธฐID๋ฅผ ์˜๋ฏธํ•œ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ ๋…ธ๋“œ๋ณ„ ์ƒ๋ณ„ ์ „๋ฅ˜๊ฐ’ ์ค‘ ๋ˆ„๋ฝ ๋ฐ์ดํ„ฐ๋Š” ์„ ํ˜•๋ณด๊ฐ„๋ฒ•์œผ๋กœ ์ฒ˜๋ฆฌํ•˜๊ณ . ์ „์ฒด ์‹œ๊ณ„์—ด ์ž…๋ ฅ๋ฐ์ดํ„ฐ๋Š” z-score๋กœ ์ •๊ทœํ™”ํ•˜์˜€๋‹ค.

Table 1. D/L adjacent matrix

๊ฐœํ๊ธฐ

ID

13088

16200

4688

9528

15344

4696

15358

13088

0

1

1

0

0

0

1

16200

1

0

0

0

0

0

0

4688

1

0

0

1

1

0

0

9528

0

0

1

0

0

0

0

15344

0

0

1

0

0

1

0

4696

0

0

0

0

1

0

0

15358

1

0

0

0

0

0

60

3. D/L ๋ณ„ ๊ฐœํ๊ธฐ ๋ถ€ํ•˜์˜ˆ์ธก๋ชจ๋ธ

์ตœ๊ทผ ๋”ฅ๋Ÿฌ๋‹ ๋ถ„์•ผ์—์„œ ์‘์šฉ๋˜๋Š” ๊ทธ๋ž˜ํ”„ ์‹ ๊ฒฝ๋ง(GNN)์€ ๋‹ค์–‘ํ•œ ์•Œ๊ณ ๋ฆฌ์ฆ˜๊ณผ ๊ฒฐํ•ฉํ•˜์—ฌ ํ™œ์šฉ๋˜๊ณ  ์žˆ๋‹ค. ํŠนํžˆ ์‹œ๊ณ„์—ด ์ถ”๋ก  ๋ถ„์•ผ์—์„œ๋Š” ๊ธฐ์กด LSTM, GRU ๋“ฑ์˜ ์ˆœํ™˜์‹ ๊ฒฝ๋ง(Recurent Neural Network, RNN)๊ณ„์—ด์˜ ๋ชจ๋ธ์„ ํ†ตํ•ด ์ƒˆ๋กœ์šด ์ž…๋ ฅ๊ฐ’๊ณผ ์ด์ „์˜ ์ถœ๋ ฅ๊ฐ’์˜ ๋น„์ค‘์„ ๊ฒŒ์ดํŠธ๋กœ ๊ณ ๋ คํ•จ์œผ๋กœ์จ ๋‹ค์–‘ํ•œ ์‹œ๊ณ„์—ด ๋ฌธ์ œ๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ํ•ด๊ฒฐํ•˜์˜€๋‹ค(12,13). ํ•˜์ง€๋งŒ ์ด๋Š” ์ˆœ์ฐจ์  ์—ฐ์‚ฐ์„ ํ†ตํ•œ ๋ณ‘๋ ฌ์ฒ˜๋ฆฌ์— ํ•œ๊ณ„๋กœ ์—ฐ์‚ฐ๋Ÿ‰์ด ํฌ๋‹ค๋Š” ๋‹จ์ ์„ ์ง€๋‹Œ๋‹ค. ์ด๋Ÿฌํ•œ ๋ฌธ์ œ์ ์„ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ๊ฒŒ์ดํŠธ ์ปจ๋ณผ๋ฃจ์…˜(gated convolution)(14) ๋ฐฉ๋ฒ•๋ก ์ด ๊ฐœ๋ฐœ๋˜์—ˆ๋‹ค.

๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” D/L ๋‚ด ์ž๋™ํ™” ๊ฐœํ๊ธฐ์˜ ๋‹จ๊ธฐ ๋ถ€ํ•˜์˜ˆ์ธก์„ ๋ชฉํ‘œ๋กœ ํ•˜๋ฉฐ, ๊ธฐ์กด์˜ ๊ธฐ๊ธฐ๋ณ„ ๋ชจ๋ธ์ด ์•„๋‹Œ ํ•˜๋‚˜์˜ ์‹œ๊ณต๊ฐ„ ๊ทธ๋ž˜ํ”„ ์‹ ๊ฒฝ๋ง ๊ธฐ๋ฐ˜ ๋ชจ๋ธ์„ ํ™œ์šฉํ•œ๋‹ค.

3.1 ์‹œ๊ณต๊ฐ„ ๊ทธ๋ž˜ํ”„ ์ปจ๋ณผ๋ฃจ์…˜ ์‹ ๊ฒฝ๋ง (ST-GCNs)

๋ณธ ์ ˆ์—์„œ ์„ค๋ช…ํ•˜๋Š” ๋ชจ๋ธ์˜ ์ „๋ฐ˜์ ์ธ ๊ตฌ์กฐ์— ๊ด€ํ•ด์„œ ์„ค๋ช…ํ•œ๋‹ค. ์‹œ๊ณต๊ฐ„ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ํ•™์Šต์„ ์ง„ํ–‰ํ•˜๊ธฐ ์œ„ํ•œ ๊ณต๊ฐ„ ๊ทธ๋ž˜ํ”„ ์ปจ๋ณผ๋ฃจ์…˜(spatial graph convolution)๊ณผ ์‹œ๊ฐ„ ๊ฒŒ์ดํŠธ ์ปจ๋ณผ๋ฃจ์…˜(temporal gated convolution) (14)์ธ ๋‘ ๊ฐ€์ง€ ๊ตฌ์กฐ๋กœ ๊ตฌ์„ฑ๋˜๋ฉฐ, Fig. 6๊ณผ ๊ฐ™๋‹ค.

Fig. 6. ST-GCNs architecture
../../Resources/kiiee/JIEIE.2022.36.1.037/fig6.png

์ฒซ์งธ๋กœ, ๊ณต๊ฐ„ ์ •๋ณด์— ๋Œ€ํ•œ ์—ฐ์‚ฐ $*_{๐’ข}$์€ $K$ ํฌ๊ธฐ ์ปค๋„, ์ฒด๋น„์…ฐํ”„ ๊ณ„์ˆ˜(Chebyshev coefficient)์ธ $\Theta_{i,\: j}\in R^{K}$, ๊ทธ๋ฆฌ๊ณ  ๋ผํ”Œ๋ผ์‹œ์•ˆ ํ–‰๋ ฌ(Laplacian matrix)์ธ $L$๋กœ $y_{j}=\sum_{i = 1}^{C_{i}}\Theta_{i,\: j}(L)x_{i}$ ๋กœ ๊ณต๊ฐ„์  ๊ทธ๋ž˜ํ”„ ์ปจ๋ณผ๋ฃจ์…˜์„ ๋‚˜ํƒ€๋‚ธ๋‹ค. ๊ฐ„๋‹จํžˆ 3์ฐจ์› ํ‘œํ˜„์œผ๋กœ $\Theta *_{๐’ข}๐“ง$, $๐“ง\in R^{M'\times node\times C_{i}}$ ์ •์˜๋œ๋‹ค. ์ด๋Ÿฌํ•œ ์—ฐ์‚ฐ์€ ๋„คํŠธ์›Œํฌ๊ฐ€ ๊ฐ ๋…ธ๋“œ ๊ฐ„์˜ ์—ฐ๊ฒฐ ๊ด€๊ณ„๋ฅผ ๋‚ดํฌํ•˜์—ฌ ํ•™์Šตํ•˜๋„๋ก ํ•œ๋‹ค(11).

๋‘ ๋ฒˆ์งธ๋กœ, ์‹œ๊ฐ„ ์ •๋ณด์— ๋Œ€ํ•œ ํ•™์Šต์€ ์‹œ๊ฐ„ ๊ฒŒ์ดํŠธ ์ปจ๋ณผ๋ฃจ์…˜(temporal gated convolution) ๊ตฌ์กฐ๋ฅผ ํ™œ์šฉํ•œ๋‹ค. CNN๊ธฐ๋ฐ˜ ๊ตฌ์กฐ์— ๊ฒŒ์ดํŠธ(gate)๋ฅผ ๋„์ž…ํ•˜์—ฌ, LSTM์—์„œ ๋ฐœ์ƒํ•˜๋Š” ์ฒด์ธ ๊ตฌ์กฐ์™€ ์ด์ „ ํžˆ๋“ (hidden) ์ƒํƒœ์— ์˜์กดํ•˜๋Š” ๊ณ„์‚ฐ์  ๋ฌธ์ œ์ ์„ ํ•ด๊ฒฐํ•˜๋Š” ๊ตฌ์กฐ์ด๋‹ค. ์ด๋ฅผ ํ†ตํ•ด ๊ณ„์‚ฐ๋Ÿ‰์„ ์ปค๋„์˜ ํฌ๊ธฐ k์— ๋Œ€ํ•ด, $O(N/k)$์˜ ๊ณ„์‚ฐ์  ์ด๋“๊ณผ ์žฅ๊ธฐ๊ฐ„ ์ข…์†์„ฑ์„ ์œ ์ง€ํ•  ์ˆ˜ ์žˆ๋‹ค. ์ด๋ฅผ ๊ฒŒ์ดํŠธ ์„ ํ˜• ์œ ๋‹›(gated linear unit, GLU)๋ผ๊ณ  ํ•˜๋ฉฐ. ์ˆ˜์‹์€ ์•„๋ž˜์™€ ๊ฐ™๋‹ค(11,14).

$T *_{๐’ฏ}Y = C_{out_{1}}(I)\otimes\sigma(C_{out_{2}}(I))\in R^{\left(M-K_{t}+1\right)\times C_{out}}$

์—ฐ์‚ฐ $*_{๐’ฏ}$์€ $T\in R^{K_{t}\times C_{i n}\times 2C_{out}}$ ๋Š” ํฌ๊ธฐ $K_{t}$ ์ธ ์ปค๋„๋กœ ์—ฐ์‚ฐ ์ž…๋ ฅ๊ฐ’($I$ )์— ๋Œ€ํ•œ ์ปจ๋ณผ๋ฃจ์…˜๊ณผ ์‹œ๊ทธ๋ชจ์ด๋“œ(sigmoid) ํ•จ์ˆ˜๋ฅผ ํ†ต๊ณผํ•œ ์ปจ๋ณผ๋ฃจ์…˜ ์ถœ๋ ฅ์— ๋Œ€ํ•œ โŠ— ์—ฐ์‚ฐ์€ ์•„๋‹ค๋งˆ๋ฅด ๊ณฑ(element-wise product), $M$์€ ์‹œ๊ณ„์—ด ๋ฐ์ดํ„ฐ์˜ ๊ธธ์ด, $C$๋Š” ์ปจ๋ณผ๋ฃจ์…˜, ๊ทธ๋ฆฌ๊ณ  $C_{"\in}"$์™€ $C_{out = out_{1}+out_{2}}$์€ ๊ฐ๊ฐ ์ปจ๋ณผ๋ฃจ์…˜์˜ ์ž…๋ ฅ๊ณผ ์ถœ๋ ฅ ์ฑ„๋„์˜ ํฌ๊ธฐ๋ฅผ ์˜๋ฏธํ•œ๋‹ค. ์ด๋Š” LSTM์˜ ๊ฒŒ์ดํŠธ์™€ ๊ฐ™์ด ์ •๋ณด์˜ ์ž…์ถœ๋ ฅ์„ ์ œ์–ดํ•˜๋Š” ์—ญํ• ์„ ํ•œ๋‹ค. ์•ž์„œ ์†Œ๊ฐœํ•œ Fig. 6์—์„œ ์‹œ๊ณต๊ฐ„ ์ปจ๋ณผ๋ฃจ์…˜ ๋ธ”๋ก(spatial-temporal convolutional block)์ด๋ผ๊ณ  ํ•˜๋ฉฐ, ์ˆ˜์‹์€ ์•„๋ž˜์™€ ๊ฐ™์ด ๋‚˜ํƒ€๋‚ผ ์ˆ˜ ์žˆ๋‹ค(11).

$i^{l+1}= T_{1}*_{๐’ฏ}Re LU\left(\theta^{l}*_{๐’ข}\left(T_{0}^{l}*_{๐’ฏ}i^{l}\right)\right)$

๊ฒฐ๊ณผ์ ์œผ๋กœ, $i^{l+1}$์˜ ์˜ˆ์ธก๊ฐ’๊ณผ ์‹ค์ธก๊ฐ’์„ Loss๊ฐ’์„ ์ตœ์†Œํ™”ํ•˜๋ฉฐ, ์‹ ๊ฒฝ๋ง์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ํ•™์Šตํ•œ๋‹ค.

3.2 ๋น„๊ต ๋ชจ๋ธ

ํ•™์Šต๋œ ST-GCN ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ๊ฒ€์ฆํ•˜๊ธฐ ์œ„ํ•ด์„œ ๋‹ค์Œ๊ณผ ๊ฐ™์ด ์‹œ๊ณ„์—ด ์˜ˆ์ธก์— ๋„๋ฆฌ ์‚ฌ์šฉ๋˜๋Š” ARIMA ๋ชจ๋ธ๊ณผ ๋”ฅ๋Ÿฌ๋‹ ๋ชจ๋ธ์ธ LSTM์„ ๋น„๊ต๊ตฐ์œผ๋กœ ์„ ํƒํ•˜์˜€๋‹ค. ๋น„๊ต๊ตฐ์˜ ๊ฒฝ์šฐ ๋ชจ๋ธ ํŠน์„ฑ์ƒ D/L์— ์†ํ•œ ๊ฐ ์ž๋™ํ™” ๊ฐœํ๊ธฐ๋ณ„ 23๊ฐœ์˜ ๋ชจ๋ธํ•™์Šต์„ ์ˆ˜ํ–‰ํ•˜๊ณ  ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ•œ๋‹ค. ๋น„๊ต๊ตฐ์— ๋Œ€ํ•œ ์†Œ๊ฐœ๋Š” ์•„๋ž˜์™€ ๊ฐ™๋‹ค.

โˆ™ARIMA(Auto Regressive Integrated Moving Average(8)): ์‹œ๊ณ„์—ด ์˜ˆ์ธก์— ๋งŽ์ด ์‚ฌ์šฉ๋˜๋Š” ๊ธฐ๋ฒ•์œผ๋กœ์„œ ์ž๊ธฐํšŒ๊ท€(aurto regression)๋ชจํ˜•์˜ ์‹œ๊ฐ„ ์ฐจ์ด์ธ p, ์ด๋™ํ‰๊ท (movig average)๋ชจํ˜•์˜ ์‹œ๊ฐ„ ์ฐจ์ด์ธ q, ์ฐจ๋ถ„(Differencing)์˜ ํšŸ์ˆ˜ d์ธ ์ดˆ๋งค๊ฐœ๋ณ€์ˆ˜(hyperparmeter)์ธ [p, d, q]๋ฅผ ํ†ตํ•ด ์ž๊ธฐํšŒ๊ท€, ์ฐจ๋ถ„, ์˜ค์ฐจ์— ๋Œ€ํ•œ ์ž๊ธฐํšŒ๊ท€, ๊ทธ๋ฆฌ๊ณ  ๊ณ„์ ˆ ์ฃผ๊ธฐ๋ฅผ ๋ชจ๋ธ๋งํ•œ๋‹ค. ๋ชจ๋ธ์€ ARIMA๊ธฐ๋ฐ˜ ๋‹จ๊ธฐ๋ถ€ํ•˜ ์˜ˆ์ธก์— ๊ด€ํ•œ ๋…ผ๋ฌธ(8)์˜ ์‹คํ—˜๊ฒฐ๊ณผ๋ฅผ ์ฐธ๊ณ ํ•˜์—ฌ, ์ดˆ๋งค๊ฐœ๋ณ€์ˆ˜ [1, 0, 0]์„ ์ค‘์‹ฌ์œผ๋กœ ๊ทธ๋ฆฌ๋“œ ์„œ์น˜(grid search)ํ•œ๋‹ค. ๋‹จ๊ธฐ๋ถ€ํ•˜ ์˜ˆ์ธก๋ชจ๋ธ์€ ๋งค์‹œ๊ฐ„ ํ•™์Šต๋ฐ์ดํ„ฐ๋ฅผ ์ ์ฆ์ ์œผ๋กœ ๋ฐ˜์˜ํ•˜์—ฌ ์ตœ์ ํ™”ํ•œ๋‹ค.

โˆ™LSTM(Long Short-Term Memory(15)): ๊ธฐ๊ณ„ํ•™์Šต ๋ชจ๋ธ์ด ๊ฐ€์ง„ ๊ธฐ์šธ๊ธฐ ์†Œ์‹ค ๋ฌธ์ œ(gradient vanishing problem)์™€ ๋ฐ์ดํ„ฐ ์žฅ๊ธฐ์˜์กด์„ฑ(long-term dependency) ๋ชจ๋ธ๋ง์—์„œ ์ข‹์€ ์„ฑ๋Šฅ์„ ๋ณด์ด๋Š” ๋ชจ๋ธ์ด๋‹ค. ๋‚ด๋ถ€์— ๋ง๊ฐ(forget) ๊ฒŒ์ดํŠธ, ์ž…๋ ฅ(input) ๊ฒŒ์ดํŠธ, ๊ทธ๋ฆฌ๊ณ  ์ถœ๋ ฅ(output) ๊ฒŒ์ดํŠธ๋กœ ๊ตฌ์„ฑ๋œ ๊ฒŒ์ดํŠธ๋“ค์„ ํ†ตํ•ด LSTM ๋‚ด ์…€ ์ •๋ณด๋“ค์„ ๊ฐ€์ค‘ํ•œ ์ƒํƒœ์ •๋ณด๋“ค์„ ์ˆœํ™˜ ๊ตฌ์กฐ๋กœ ์ดํ›„ ์…€์— ์ „๋‹ฌํ•˜๋„๋ก ๊ตฌ์„ฑ๋œ๋‹ค. ๋…ผ๋ฌธ(16)์—์„œ ์ดˆ๋งค๊ฐœ๋ณ€์ˆ˜๋ฅผ ์ฐธ๊ณ ํ•˜์˜€์œผ๋ฉฐ, Table 2์™€ ๊ฐ™๋‹ค.

Table 2. Hyperparameers setting of the LSTM model

ํžˆ๋“ 

๋ ˆ์ด์–ด ์ˆ˜

ํžˆ๋“ 

๋…ธ๋“œ ์ˆ˜

ํ•™์Šต์œจ

ํ•™์ŠตํšŸ์ˆ˜

3

150

0.001

1-600

4. ์‹คํ—˜ ๋ฐ ๋ถ„์„

๋ณธ ์žฅ์—์„œ๋Š” ๊ฐœํ๊ธฐ ์—ฐ๊ฒฐ์ •๋ณด๋ฅผ ํ™œ์šฉํ•œ ๊ทธ๋ž˜ํ”„ ์‹ ๊ฒฝ๋ง ๋ชจ๋ธ์ธ ST-GCNs๊ธฐ๋ฐ˜ ๋‹จ๊ธฐ๋ถ€ํ•˜์˜ˆ์ธก ๋ชจ๋ธ์˜ ํ•™์Šต๊ฒฐ๊ณผ์™€ ์˜ˆ์ธก ์ •ํ™•์„ฑ์„ ๊ฒ€์ฆํ•˜๊ธฐ ์œ„ํ•œ ์„ฑ๋Šฅํ‰๊ฐ€ ๊ฒฐ๊ณผ๋ฅผ ์ •๋ฆฌํ•œ๋‹ค.

4.1 ๊ทธ๋ž˜ํ”„ ์‹ ๊ฒฝ๋ง ๋ชจ๋ธํ•™์Šต

๋ณธ ์‹คํ—˜์— ์‚ฌ์šฉ๋œ ๊ฐœํ๊ธฐ ๋…ธ๋“œ ์ •๋ณด์™€ ๋ถ€ํ•˜ ๋ฐ์ดํ„ฐ๋Š” ์ข…ํ•ฉ๋ฐฐ์ „์ž๋™ํ™”์‹œ์Šคํ…œ์—์„œ ์ถ”์ถœํ•˜์˜€๋‹ค. ๋ฐ์ดํ„ฐ๋Š” 2019๋…„ 8์›”-2021๋…„ 8์›” ๊ฐ„ ์ถฉ๋ถ์˜ ๊ณต๋ถ, ์„œํ‰, ๋•์ดŒ, ์—ฐ์ œ 4๊ฐœ์˜ D/L 23๊ฐœ ๊ฐœํ๊ธฐ์˜ ์‹œ๊ฐ„๋‹น ์ƒ์ „๋ฅ˜ ์‹œ๊ณ„์—ด ๋ฐ์ดํ„ฐ ์ •๋ณด๋กœ ๊ตฌ์„ฑ๋˜๋ฉฐ, ์—ฐ๊ฒฐ์ •๋ณด๋Š” ์ž๋™ํ™” ๊ฐœํ๊ธฐ์— ๋Œ€ํ•œ ๋ถ€๋ถ„ ๊ทธ๋ž˜ํ”„๋กœ ํ‘œํ˜„๋œ ์ธ์ ‘ํ–‰๋ ฌ์„ ์‚ฌ์šฉํ•œ๋‹ค. ST-GCNs ๋ชจ๋ธ์€ ์‹œ๊ณต๊ฐ„์  ์—ฐ๊ด€ ๋ฐ์ดํ„ฐ๋ฅผ ํ†ตํ•ด D/L๋ณ„ ์˜ˆ์ธก ์ด์ „ ์‹œ์  12์‹œ๊ฐ„ ๋™์•ˆ์˜ ํŒจํ„ด์ •๋ณด๋กœ (1/ 2/ 3hr) ์ดํ›„ ์ƒ ์ „๋ฅ˜๊ฐ’์„ ์ถ”์ •ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉํ•œ๋‹ค. ํ•™์Šต ๋ฐ ํ…Œ์ŠคํŠธ ๋ฐ์ดํ„ฐ๋Š” 4๋Œ€ 1๋กœ ๋‚˜๋‰˜๋ฉฐ, ๊ธฐ๊ฐ„์œผ๋กœ๋Š” ๊ฐ๊ฐ 20๊ฐœ์›”๊ณผ 5๊ฐœ์›”์˜ ์‹œ๊ฐ„๋ณ„ ์ „๋ฅ˜ ๋ฐ์ดํ„ฐ๋กœ ๊ตฌ์„ฑ๋œ๋‹ค. ์‹คํ—˜ํ™˜๊ฒฝ์€ ๋ฐ์Šคํฌํƒ‘ ํ™˜๊ฒฝ(CPU: Intel(R) 10th i7-10875H NVIDIA GeForce RTX 2080)์—์„œ ๋ชจ๋ธํ•™์Šต์„ ์ง„ํ–‰ํ•˜์˜€๋‹ค.

4.2 ์„ฑ๋Šฅ ์ง€ํ‘œ

ST-GCNs ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ํ™•์ธํ•˜๊ธฐ ์œ„ํ•˜์—ฌ, ํ‰๊ท ์ ˆ๋Œ€ ์˜ค์ฐจ(mean absolute error, MAE), ์ œ๊ณฑ๊ทผ ํ‰๊ท ์ œ๊ณฑ ์˜ค์ฐจ(root mean squared error, RMSE), ๊ทธ๋ฆฌ๊ณ  ํ‰๊ท  ์ ˆ๋Œ€ ๋ฐฑ๋ถ„์œจ ์˜ค์ฐจ(mean absolute percentage error, MAPE)๋ฅผ ์ธก๋„๋กœ ์‚ฌ์šฉํ•œ๋‹ค. ์œ„์˜ ์„ธ ๊ฐ€์ง€ ์ธก๋„๋Š” (1 / 2 / 3hr) ์ดํ›„ ์˜ˆ์ธก๊ฐ’์œผ๋กœ ๊ฐ๊ฐ ๋น„๊ตํ•œ๋‹ค.

(1)
$R MSE=\sqrt{\dfrac{1}{H}\sum_{t=1}^{H}\dfrac{1}{N}\sum_{j=1}^{N}(X_{j}^{t}-\hat X_{j}^{t})^{2}}$

(2)
$MAE=\dfrac{1}{H}\sum_{t=1}^{H}\dfrac{1}{N}\sum_{j=1}^{N}\left | X_{j}^{t}-\hat X_{j}^{t}\right |$

(3)
$MAPE=100\times\dfrac{1}{H}\sum_{t=1}^{H}\dfrac{1}{N}\sum_{j=1}^{N}\dfrac{\left | X_{j}^{t}-\hat X_{j}^{t}\right |}{X_{j}^{t}}$

์—ฌ๊ธฐ์„œ, $H$ ๋Š” ์‹œ๊ฐ„์„, $N$์€ ํ•™์Šต ๊ฐœํ๊ธฐ์˜ ์ˆ˜๋ฅผ ์˜๋ฏธํ•œ๋‹ค. ์„ฑ๋Šฅ ์ง€ํ‘œ๋Š” ์˜ˆ์ธก์„ฑ๋Šฅ์„ ํ™•์ธํ•˜๊ธฐ ์œ„ํ•œ ์ธก๋„๋กœ ๋‚ฎ์„์ˆ˜๋ก ์ข‹์€ ์„ฑ๋Šฅ์„ ์˜๋ฏธํ•œ๋‹ค.

4.3 ์‹คํ—˜๊ฒฐ๊ณผ

๋ณธ ๋…ผ๋ฌธ์—์„œ ์ œ์•ˆ๋œ ๋ชจ๋ธ๊ณผ ๋น„๊ต๋ชจ๋ธ์˜ ์˜ˆ์ธก ๊ฒฐ๊ณผ๋Š” Table 3๊ณผ ๊ฐ™๋‹ค. RMSE์™€ MAE ๊ด€์ ์—์„œ, ST-GCNs๊ธฐ๋ฐ˜ ๋ชจ๋ธ์€ ARIMA์™€ LSTM์˜ ๊ฒฐ๊ณผ์™€ ๋น„๊ตํ•˜์—ฌ ์ค€์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ฃผ์—ˆ๋‹ค.

Table 3. Comparison of forecasting results

Model

๊ณต๋ถ(1/ 2/ 3 hr)

RMSE

MAE

MAPE(%)

ARIMA

8.01/10.84/12.98

4.59/6.57/8.21

22.50/32.07/41.49

LSTM

6.00/7.95/10.27

3.44/4.86/6.68

22.96/27.80/33.63

ST-GCN

5.88/7.01/7.98

3.12/3.71/4.36

35.61/36.91/38.91

์„œํ‰(1/ 2/ 3 hr)

ARIMA

13.18/18.91/23.13

8.87/12.78/14.97

40.65/41.46/42.47

LSTM

9.78/14.71/19.51

4.90/7.71/10.56

8.64/ 13.19/17.81

ST-GCN

11.31/14.38/16.98

6.49/7.94/9.42

11.50/13.65/15.45

๋•์ดŒ(1/ 2/ 3 hr)

ARIMA

11.35/15.41/17.73

8.27/10.95/12.72

8.31/10.39/ 12.36

LSTM

13.03/16.12/21.50

6.16/8.19/11.68

5.88/9.29/13.35

ST-GCN

10.42/13.87/16.75

5.58/7.46/9.33

7.66/10.19/12.67

์—ฐ์ œ(1/ 2/ 3 hr)

ARIMA

13.52/18.75/22.23

9.73/13.06/15.18

11.10/14.42/16.92

LSTM

19.20/20.54/21.64

11.22/11.94/12.75

16.11/18.81/21.73

ST-GCN

17.61/20.25/22.99

10.40/12.05/13.86

19.39/22.29/25.52

์—ฐ์žฌ-D/L์˜ ๊ฒฝ์šฐ๋Š” ํ•™์Šต๋˜์ง€ ์•Š์€ ์ƒˆ๋กœ์šด ํŒจํ„ด์˜ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด์„œ๋Š” ARIMA์ฒ˜๋Ÿผ ํ•™์Šต๋ฐ์ดํ„ฐ๋ฅผ ์ ์ฆ์ ์œผ๋กœ ์ฆ๊ฐ€์‹œํ‚ค๋ฉฐ ์ตœ๊ทผ์˜ ๋ฐ์ดํ„ฐ๋ฅผ ๋ฐ˜์˜ํ•˜๋Š” ๋ฐฉ์‹์ด ์ด์ ์„ ๊ฐ€์กŒ๋‹ค. ํ•˜์ง€๋งŒ ์ด๊ฒƒ์€ ์ˆ˜์‹œ๋กœ ๋ชจ๋ธ์„ ๊ฐฑ์‹ ํ•ด์•ผ๋งŒ ํ•˜๋Š” ๋ฌธ์ œ์ ์„ ์ง€๋‹Œ๋‹ค.

์ œ์•ˆ๋œ ๋ชจ๋ธ์€ MAPE ๊ธฐ์ค€ ๊ณต๋ถ D/L์—์„œ ๊ฐ€์žฅ ๋‚ฎ์€ ์„ฑ๋Šฅ์„ ๋ณด์˜€๋‹ค. D/L ๋‚ด ๋‘ ๊ฐœํ๊ธฐ์˜ ํ•™์Šต๋ฐ์ดํ„ฐ์™€๋Š” ์ „ํ˜€ ๋‹ค๋ฅธ ํ˜•ํƒœ ํŒจํ„ด์˜ ์˜ˆ์ธก ๊ฒฐ๊ณผ๋กœ ๋ณด์ธ๋‹ค. ์ด ๋‘ ๊ฐœํ๊ธฐ๋ณ„ ์˜ค์ฐจ๋Š” 67%์™€ 81%๋กœ ์˜ˆ์ธกํ•˜์˜€์œผ๋ฉฐ, LSTM์˜ ๊ฒฝ์šฐ์—๋Š” 77%์™€ 26%๋ฅผ ๋ณด์˜€๋‹ค. ์ด๋Š” ์•ž์œผ๋กœ ์ œ์•ˆ๋œ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ๋†’์ด๋Š” ๋ฐ ์žˆ์–ด์„œ, ๋‹ค์–‘ํ•œ ํ•™์Šต๋ฐ์ดํ„ฐ๊ฐ€ ํ•„์š”ํ•˜๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค€๋‹ค.

Fig. 7. MAPE-Boxplot result of test models
../../Resources/kiiee/JIEIE.2022.36.1.037/fig7.png

Fig. 7์€ ์‹คํ—˜๋ชจ๋ธ ๊ฐ„ MAPE ์„ฑ๋Šฅ์„ ๋ฐ•์Šค ํ”Œ๋กฏ์œผ๋กœ ๋‚˜ํƒ€๋‚ด๋ฉฐ, ์ด๋Š” ๋”ฅ๋Ÿฌ๋‹ ๊ณ„์—ด ๋ชจ๋ธ์ด ARIMA์™€ ๋น„๊ตํ•˜์—ฌ, ์ „๋ฐ˜์ ์œผ๋กœ ๋†’์€ ์ •ํ™•๋„์™€ ์˜ค๋ฅ˜์— ๋Œ€ํ•œ ๋ถ„์‚ฐ๋„๊ฐ€ ๋‚ฎ์Œ์„ ํ™•์ธํ•  ์ˆ˜ ์žˆ๋‹ค. ์‹คํ—˜ ๊ฒฐ๊ณผ๋ฅผ ์ข…ํ•ฉํ•˜๋ฉด, ๊ฐœํ๊ธฐ๋ณ„ LSTM ๊ธฐ๋ฐ˜ ๋ชจ๋ธ์ด ๊ฐ€์žฅ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ์–ป์„ ์ˆ˜ ์žˆ์—ˆ๋‹ค.

Table 4. Time consumption of training

Model

Time Consumption(s)

LSTM

11,598 x 23 switches

ST-GCNs

1,044 x 4 D/L

ํ•˜์ง€๋งŒ, ์ด๋Ÿฌํ•œ ์„ฑ๋Šฅ์„ ์–ป๊ธฐ ์œ„ํ•ด ํ•˜๋‚˜์˜ ๋ชจ๋ธ์„ ํ•™์Šตํ•˜๊ธฐ ์œ„ํ•ด ์‚ฌ์šฉ๋œ ํ•™์Šต ์‹œ๊ฐ„(Table 4)์€ ST-GCN๊ณผ ๋น„๊ตํ•˜์—ฌ ๋Œ€๋žต 11๋ฐฐ์˜ ์‹œ๊ฐ„์ด ์†Œ์š”๋˜์—ˆ์œผ๋ฉฐ, ์ƒ์„ฑ๋˜๋Š” ๋ชจ๋ธ์˜ ๊ฐœ์ˆ˜๋ฅผ ๋ฐ˜์˜ํ•  ๊ฒฝ์šฐ 63๋ฐฐ๊นŒ์ง€ ๋Š˜์–ด๋‚œ๋‹ค. ์ด๋Š” ์ „๊ตญ ๋‹จ์œ„์˜ D/L์— ๋Œ€ํ•œ ์ ‘๊ทผ์— ์žˆ์–ด์„œ ์—ฐ์‚ฐ์ž์›ํ™•๋ณด์— ๋Œ€ํ•œ ์–ด๋ ค์›€์„ ๊ฐ€์ง„๋‹ค. ๋”ฐ๋ผ์„œ ๋ณธ ๋…ผ๋ฌธ์—์„œ๋Š” ์ œ์•ˆ๋œ ๋ชจ๋ธ์ด ์—ฐ์‚ฐ์‹œ๊ฐ„๊ณผ ์ •ํ™•๋„ ์ธก๋ฉด์„ ๊ณ ๋ คํ•˜์—ฌ, ์‹ค์ œ ์‚ฐ์—…์— ์‘์šฉํ•˜๋Š” ๋ฐ ์žฅ์ ์ด ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ํŒ๋‹จ๋œ๋‹ค.

5. ๊ฒฐ ๋ก 

๋ณธ ์—ฐ๊ตฌ๋Š” ๋ฐฐ์ „์„ ๋กœ์— ์„ค์น˜๋œ FRTU๋กœ ๋ถ€ํ„ฐ ์ทจ๋“๋˜๋Š” ์‹œ๊ฐ„๋ณ„ ์ธก์ •๋ฐ์ดํ„ฐ์™€ ๊ฐœํ๊ธฐ ์—ฐ๊ฒฐ๊ทธ๋ž˜ํ”„ ๊ตฌ์กฐ๋ฅผ ํ™œ์šฉํ•˜์—ฌ ์—ฐ๊ฒฐ์ •๋ณด๋ฅผ ๋ฐ˜์˜ํ•œ ์‹œ๊ฐ„์— ๋”ฐ๋ฅธ ๋ถ€ํ•˜๋ฅผ ํ‘œํ˜„ ๋ฐ ์˜ˆ์ธกํ•˜๋Š” ์—ฐ๊ตฌ๋ฅผ ์ง„ํ–‰ํ•˜์˜€๋‹ค. ์ „๋ ฅ ๋ถ„์•ผ์—์„œ์˜ ์ตœ์‹  ๊ทธ๋ž˜ํ”„ ์‹ ๊ฒฝ๋ง ๊ธฐ๋ฐ˜ ๋ชจ๋ธ์˜ ํ•˜๋‚˜์ธ ST-GCN์— ์ ์šฉํ•œ ์‚ฌ๋ก€์—ฐ๊ตฌ๋ฅผ ํ†ตํ•ด ์ž๋™ํ™” ๊ฐœํ๊ธฐ์˜ ๋ถ€ํ•˜์˜ˆ์ธก์—์„œ์˜ ์œ ์šฉ์„ฑ์„ ๊ฒ€์ฆํ•˜์˜€๋‹ค. ๋˜ํ•œ, D/L ๋‚ด ์—ฐ๊ฒฐ๋œ ๋ชจ๋“  ์ž๋™ํ™” ๊ฐœํ๊ธฐ์˜ ๋ถ€ํ•˜์˜ˆ์ธก์„ ํ•˜๋‚˜์˜ ๋ชจ๋ธ๋กœ ํ•™์Šตํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ํŽธ์˜์„ฑ์„ ํ™•์ธํ•˜์˜€๋‹ค. ์ด๋Ÿฌํ•œ ์žฅ์ ์€ ์ „๋ ฅ ๋ถ„์•ผ์˜ ํŠน์„ฑ์ƒ ๋‹ค์–‘ํ•œ ํ˜•ํƒœ๋กœ ์—ฐ๊ฒฐ๋œ ๋ฐฐ์ „๋ง ๋˜๋Š” ์†ก์ „๋ง์—์„œ ์ƒ์‚ฐ๋˜๋Š” ์‹œ๊ณ„์—ด์˜ ์—ฐ๊ด€์„ฑ์„ ํ†ตํ•ด ์ •์ „๋ณต๊ตฌ, ๋ณดํ˜ธํ˜‘์กฐ, ์ตœ์ ํ™” ๋“ฑ์˜ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๋Š”๋ฐ ํ™œ์šฉ๋  ์ˆ˜ ์žˆ์„ ๊ฒƒ์ด๋‹ค. ์ฐจํ›„ ์—ฐ๊ตฌ์—์„œ๋Š” ๋ชจ๋ธ์˜ ์˜ˆ์ธก์„ฑ๋Šฅ ํ–ฅ์ƒ๊ณผ ํ•˜๋‚˜์˜ D/L์ด ์•„๋‹Œ ํ™•์žฅ๋œ ์ „๋ ฅ๋ง์—์„œ์˜ ์ „๋ฅ˜ ๋ฐ ์ œ์–ด ์ƒํƒœ, GIS ์œ„์น˜์ •๋ณด, ๊ธฐ์ƒ ์‹œ๊ณ„์—ด ๋ฐ์ดํ„ฐ, ๊ทธ๋ฆฌ๊ณ  ๊ณ ๊ฐ์ •๋ณด ๋“ฑ์„ ํ™œ์šฉํ•œ ๋ชจ๋ธ ์—ฐ๊ตฌ๋ฅผ ์ง„ํ–‰ํ•  ์˜ˆ์ •์ด๋‹ค.

Acknowledgements

๋ณธ ๋…ผ๋ฌธ(์ €์„œ)์€ 2021ํ•™๋…„๋„ ๋ชฉํฌ๋Œ€ํ•™๊ต ๊ต๋‚ด์—ฐ๊ตฌ๋น„ ์ง€์›์— ์˜ํ•˜์—ฌ ์—ฐ๊ตฌ๋˜์—ˆ์Œ.

This Research was supported by Research Funds of Mokpo National University in 2021.

References

1 
Kim M. S., Ha B. N., Kwan S. C., Park Y. B., 2016, Research on Conceptual Design of the Next-generation Distribution Management System, KEPRIGoogle Search
2 
Kim D. W., 2019, Design and Verification of Advanced Distribution Management System using Information and Communication Convergence Technology, KSDIM, Vol. 15, No. 4, pp. 19-29DOI
3 
Lim S. I., 2012, A Section Load Management Method using Daily Load Curve in Distribution Systems, JIEIE, Vol. 26, No. 6, pp. 47-52DOI
4 
Kim B. S., Song K. B., 2005, Short-Term Load Froecasting of Transformer using Artificial Neural Networks, JIEIE, Vol. 19, No. 7, pp. 20-25DOI
5 
Jain A., Satish B., 2009, Short Term Load Forecasting by Clustering Technique Based on Daily Average and Peak Loads, IEEE Power & Energy Society General Meeting, pp. 1-7DOI
6 
Dong X., Qian L., Huang L., 2017, Short-term Load Forecasting in Smart Grid: A Combined CNN and K-means Clustering Approach, IEEE International Conference on Big Data and Smart Computing (BigComp), pp. 119-125DOI
7 
Lee H., Lee B., Ahn H., 2020, An LSTM Neural Network Model for Forecasting Daily Peak Electric Load of EV Charging Stations, Journal of Internet Computing and Services, Vol. 21, No. 5, pp. 119-127DOI
8 
Alberg D., Last M., pp 241-249 2018, Short-term Load Forecasting in Smart Meters with Sliding Window-based ARIMA Algorithms, Vietnam Journal of Computer Science, Vol. 5DOI
9 
Scarselli F., Gori M., Tsoi A. C., Hagenbuchner M., Monfardini G., 2008, The Graph Neural Network Model, IEEE Transactions on Neural Networks, Vol. 20, No. 1, pp. 61-80DOI
10 
Liu Z., Zhou J., 2020, Introduction to Graph Neural Networks, Synthesis Lectures on Artificial Intelligence and Machine Learning, Vol. 14, No. 2, pp. 1-127DOI
11 
Yu B., Yin H., Zhu Z., 2018, Spatio-temporal Graph Convolutional Networks: A Deep Learning Framework for Traffic Forecasting, In Proceedings of the 27th International Joint Conference on Artificial Intelligence, pp. 3634-3640DOI
12 
Sak H., Senior A. W., Beaufays F., 2014, Long Short-term Memory Recurrent Neural Network Architectures for Large Scale Acoustic ModelingGoogle Search
13 
Cho K., 2014, Learning Phrase Representations Using RNN Encoder-decoder for Statistical Machine Translation, arXiv preprint arXiv:1406.1078Google Search
14 
Xu Y., 2018, Large-scale Weakly Supervised Audio Classification Using Gated Convolutional Neural Network, In 2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 121-125DOI
15 
Hochreiter S., Schmidhuber J., 1997, Long Short-term Memory, Neural computation, Vol. 9, No. 8, pp. 1735-1780DOI
16 
Lee H., Lee B., Ahn H., 2020, An LSTM Neural Network Model for Forecasting Daily Peak Electric Load of EV Charging Stations, Journal of Internet Computing and Services, Vol. 21, No. 5), pp. 119-127DOI

Biography

Jaein Kim
../../Resources/kiiee/JIEIE.2022.36.1.037/au1.png

He received Ph.D. degree in Applied Mathematics from Korea University in 2020.

He is currently a Researcher in Data Science Lab., KEPCO.

Joo-Young Moon
../../Resources/kiiee/JIEIE.2022.36.1.037/au2.png

He received the M.S. degree from the Department of Data Science, Seoul National University of Science and Technology.

He is currently a Researcher in Data Science Lab., KEPCO.

Jae-Hyun Lee
../../Resources/kiiee/JIEIE.2022.36.1.037/au3.png

He received the M.S. degree in engineering from School of Integrated Tecnology (Energy program), Gwanju Institute of Science and Technology in 2020.

He is currently a Researcher in Data Science Lab., KEPCO.

Sung-Ho Park
../../Resources/kiiee/JIEIE.2022.36.1.037/au4.png

He received M.S. degree in Convergence engineering for future city from Sungkyunkwan University.

He is currently working at Data Science Lab., KEPCO in Researcher.

His work in KEPCO focuses specifically on the deep learning, anomaly detection, energy management system(EMS).

Sung-min Kim
../../Resources/kiiee/JIEIE.2022.36.1.037/au5.png

He received M.S. degree in electrical engineering from Yonsei University in 2000.

He is currently a Assistant Vice President, Head of Distribution Planning Team, KEPCO.

Dong-Sub Kim
../../Resources/kiiee/JIEIE.2022.36.1.037/au6.png

He received Ph.D. degree in Technology Policy from Yonsei University in 2014.

He joined Mokpo National University in 2020, where he is currently a professor at the Department of Electrical and Control Engineering.