Mobile QR Code QR CODE : Korean Journal of Air-Conditioning and Refrigeration Engineering

  1. ์ธํ•˜๋Œ€ํ•™๊ต ๊ฑด์ถ•ํ•™๋ถ€ ๋Œ€ํ•™์› ๋ฐ•์‚ฌ๊ณผ์ • (Ph.D. Candidate, Department of Architectural Engineering, Graduate school, Inha University, Incheon, 22212, Korea)
  2. ์ธํ•˜๋Œ€ํ•™๊ต ๊ฑด์ถ•ํ•™๋ถ€ ์กฐ๊ต์ˆ˜ (Assistant Professor, Department of Architectural Engineering, Inha University, Incheon, 22212, Korea)
  3. ํ•œ๊ตญ์—๋„ˆ์ง€๊ธฐ์ˆ ์—ฐ๊ตฌ์› ์ฑ…์ž„์—ฐ๊ตฌ์› (Principal Researcher, Department of Solar Thermal Convergence Lab, Korea Institute of Energy Research, Daejeon, 34129, Korea)
  4. ํ•œ๊ตญ์—๋„ˆ์ง€๊ธฐ์ˆ ์—ฐ๊ตฌ์› ์„ ์ž„์—ฐ๊ตฌ์› (Senior Researcher, Department of Solar Thermal Convergence Lab, Korea Institute of Energy Research, Daejeon, 34129, Korea)
  5. ์„ธ์ข…๋Œ€ํ•™๊ต ๊ธฐ๊ณ„๊ณตํ•™๊ณผ ๊ต์ˆ˜ (Professor, Department of Mechanical Engineering, Sejong University, Seoul, 05006, Korea)



์‹ ๊ฒฝ๋ง(Neural network), ๋”ฅ๋Ÿฌ๋‹(Deep learning), ์ „๊ธฐ ์ˆ˜์š”(Electricity consumption), ์žฌ์‹ค ์ •๋ณด(Occupancy information)

๊ธฐํ˜ธ์„ค๋ช…

EP,i๏ผš์˜ˆ์ธก ๋ถ€ํ•˜ [W]
EM,i๏ผš์ธก์ • ๋ถ€ํ•˜ [W]
Ev๏ผš์ธก์ • ๋ถ€ํ•˜ ํ‰๊ท  [W]
n๏ผš๋ฐ์ดํ„ฐ ์ˆ˜

1. ์„œ๋ก 

์‹ ์žฌ์ƒ์—๋„ˆ์ง€ ์„ค๋น„์˜ ์šด์ „๊ณ„ํš ๋“ฑ ์˜ˆ์ธก์ œ์–ด(MPC-model predictive control)๋ฅผ ์œ„ํ•ด์„œ๋Š” ๋ฏธ๋ž˜์˜ ์—๋„ˆ์ง€ ์‚ฌ์šฉ๋Ÿ‰์„ ์ •ํ™•ํžˆ ์˜ˆ์ธกํ•˜๋Š” ๊ฒƒ์ด ๋งค์šฐ ์ค‘์š”ํ•˜๋‹ค.(1) ํŠนํžˆ, ๋ฏธ๊ตญ๊ณผ ์œ ๋Ÿฝ์—์„œ๋Š” ์‹œ๊ฐ„๋ณ„๋กœ ๋ณ€๋™ํ•˜๋Š” ์ „๊ธฐ๊ฐ€๊ฒฉ(TOU-Time of Use)์— ๋Œ€์‘ํ•˜๊ธฐ ์œ„ํ•˜์—ฌ ์—๋„ˆ์ง€ ์†Œ๋น„๋Ÿ‰์„ ์˜ˆ์ธกํ•ด ์‹ ์žฌ์ƒ์—๋„ˆ์ง€ ์ƒ์‚ฐ์„ ์กฐ์ ˆํ•˜๊ฑฐ๋‚˜ ์ €์žฅํ•˜๋Š” ๋“ฑ ์ตœ์ ์šด์ „์ „๋žต์„ ์ˆ˜๋ฆฝํ•˜๋Š” ์—ฐ๊ตฌ๋ฅผ ์ง„ํ–‰ํ•˜๊ณ  ์žˆ๋‹ค.(2-4) ํ•œ ์˜ˆ๋กœ, Pipattanasomporn et al.(5)์˜ ์—ฐ๊ตฌ์—์„œ๋Š” ๊ฐ€์ •์šฉ ์ „๊ธฐ ์ˆ˜์š” ์˜ˆ์ธก์€ ๋…น์ƒ‰๊ฑด๋ฌผ ๋ฐ ์ฃผํƒ์„ค๊ณ„ ์‘์šฉ ๋ถ„์•ผ์—์„œ ํฐ ์ž ์žฌ๋ ฅ์„ ๊ฐ–๊ณ  ์žˆ์Œ์„ ์„ค๋ช…ํ•˜์˜€๋‹ค.

๊ฑด๋ฌผ์—์„œ ์—๋„ˆ์ง€ ์†Œ๋น„๋Ÿ‰ ์˜ˆ์ธก ์ค‘ HVAC์˜ ์—๋„ˆ์ง€ ์†Œ๋น„๋Ÿ‰์„ ์˜ˆ์ธกํ•˜๋Š” ์—ฐ๊ตฌ๋Š” ๋‹ค์–‘ํ•œ ํ˜•ํƒœ๋กœ ๊ฐœ๋ฐœ๋˜์–ด ์™”์œผ๋ฉฐ, ์˜ˆ์ธก ์„ฑ๋Šฅ์ด ๋›ฐ์–ด๋‚œ ๋ชจ๋ธ์€ ์ œ์–ด์˜ ๋ชฉ์ ์œผ๋กœ๋„ ํ™œ์šฉ๋˜๊ณ  ์žˆ๋‹ค. ํ•˜์ง€๋งŒ ์„ค๋น„๋ฅผ ์ œ์™ธํ•œ ์‹ค๋‚ด์—์„œ ๋ฐœ์ƒํ•˜๋Š” ์ „๊ธฐ๋ถ€ํ•˜๋Š” ๋ถˆํ™•์‹ค์„ฑ์ด ํฐ ์ธ์ž์ธ ์‚ฌ์šฉ์ž์˜ ํ–‰๋™์— ๋”ฐ๋ผ ๊ฒฐ์ •๋˜๊ธฐ ๋•Œ๋ฌธ์— ๊ฑด๋ฌผ์—์„œ ์ „๊ธฐ ์ˆ˜์š”๋ฅผ ์˜ˆ์ธกํ•˜๋Š” ๊ฒƒ์€ ๋งค์šฐ ์–ด๋ ต๋‹ค.(6) ๊ฑด๋ฌผ ์—๋„ˆ์ง€ ์‚ฌ์šฉ๋Ÿ‰ ๊ณ„์‚ฐ ํ”„๋กœ๊ทธ๋žจ(EnergyPlus, ESP-r, TRNSYS) ์—ญ์‹œ ๊ณ ์ •๋œ ์žฌ์‹ค ์Šค์ผ€์ค„์„ ํ†ตํ•ด ์—๋„ˆ์ง€ ์‚ฌ์šฉ๋Ÿ‰์„ ๊ฒฐ์ •ํ•˜๋ฉฐ ์ด๋Š” ์ž„์˜๋กœ ๋ณ€ํ•˜๋Š” ์žฌ์‹ค ํŒจํ„ด์— ๋Œ€ํ•œ ๊ณ ๋ ค๊ฐ€ ์–ด๋ ค์›Œ ์‹ค์ œ ์—๋„ˆ์ง€ ์‚ฌ์šฉ๋Ÿ‰๊ณผ ์ฐจ์ด๋ฅผ ๋ฐœ์ƒ์‹œํ‚จ๋‹ค.(7) Weron(2)์€ ๋‹ค์–‘ํ•œ ๋ฐฉ๋ฒ•๋ก ์˜ ์ „๊ธฐ์ˆ˜์š” ์˜ˆ์ธก ๋ชจ๋ธ์„ ์„ค๊ณ„ํ•ด ์„ฑ๋Šฅ์„ ๋น„๊ต ํ•œ ํ›„ ๊ธฐ๊ณ„ํ•™์Šต ๊ธฐ๋ฐ˜์˜ ์˜ˆ์ธก ๋ชจ๋ธ์ด ๊ฐ€์žฅ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์˜€๋‹ค๊ณ  ๋ฐœํ‘œํ•˜์˜€๋‹ค. Lago et al.(8)์€ ์ „๊ธฐ ์‚ฌ์šฉ๋Ÿ‰๊ณผ ๊ฐ™์ด ์ž„์˜์„ฑ์ด ํฐ ์‹œ๊ณ„์—ด ๋ฐ์ดํ„ฐ๋Š” ์‹ ๊ฒฝ๋งํ•™์Šต๊ตฌ์กฐ๋กœ ์˜ˆ์ธกํ•˜๊ธฐ ์œ ๋ฆฌํ•˜๋‹ค๊ณ  ์–ธ๊ธ‰ํ•˜์˜€๋‹ค. ๊ด€๋ จํ•ด์„œ ์ธ๊ณต์‹ ๊ฒฝ๋ง๋ชจ๋ธ(ANN, Aritificial Neural Network)์˜ ์ „๊ธฐ์ˆ˜์š” ์˜ˆ์ธก๋ชจ๋ธ์„ ์ œ์•ˆํ•˜๋Š” ์—ฐ๊ตฌ๊ฐ€ ๊พธ์ค€ํžˆ ์ง€์†๋˜์–ด ์™”์œผ๋ฉฐ,(9-11) ์ตœ๊ทผ์—๋Š” ๋‹จ์ผ ์€๋‹‰์ธต ๊ตฌ์กฐ์˜ ํ•œ๊ณ„๋ฅผ ๊ฐ€์ง„ ์ธ๊ณต์‹ ๊ฒฝ๋ง์˜ ๋‹จ์ ์„ ๊ฐœ์„ ํ•ด ๋‹ค์ค‘ ์€๋‹‰์ธต ๊ตฌ์กฐ๋ฅผ ๊ตฌํ˜„ํ•  ์ˆ˜ ์žˆ๋Š” ๋”ฅ๋Ÿฌ๋‹ ๋ชจ๋ธ์„ ํ†ตํ•ด ๋ณต์žกํ•œ ๋ฐ์ดํ„ฐ๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ํ•™์Šตํ•˜๋Š”(8) ์˜ˆ์ธก๋ชจ๋ธ์„ ์ œ์•ˆ ๋˜๊ณ  ์žˆ๋‹ค.

๊ตญ๋‚ด์—์„œ๋Š” ๋”ฅ๋Ÿฌ๋‹์„ ํ™œ์šฉํ•œ ์ „๊ธฐ๋ถ€ํ•˜ ์˜ˆ์ธก ์—ฐ๊ตฌ๋Š” ๋งค์šฐ ๋“œ๋ฌผ๋ฉฐ Kim and Hong(12)์€ ๊ณ„์ ˆ๊ณผ ๊ธฐ์ƒ ๋“ฑ ์™ธ๋ถ€์š”์ธ์„ ํ™œ์šฉํ•ด ๋„์‹œ์˜ ์ „๋ ฅ์ˆ˜์š”๋ฅผ ์˜ˆ์ธกํ•˜๋Š” ์ธ๊ณต์‹ ๊ฒฝ๋ง ๋ชจ๋ธ์„ ์ œ์•ˆํ•˜์˜€์œผ๋‚˜, ๋„์‹œ์˜ ์ „๋ ฅ ์ˆ˜์š”๋Š” ๊ฑฐ์ฃผ์ž์˜ ํ–‰๋™ ํŒจํ„ด์— ํฐ ์˜ํ–ฅ์„ ๋ฏธ์น˜์ง€ ์•Š์•„ ์—๋„ˆ์ง€ ์‚ฌ์šฉ์— ์žฌ์‹ค์ž์˜ ํŒจํ„ด์ด ์ง€๋ฐฐ์ ์ธ ์˜ํ–ฅ์„ ์ฃผ๋Š” ์ฃผ๊ฑฐ๊ฑด๋ฌผ์—์„œ๋Š” ์ ์šฉ์ด ์–ด๋ ต๋‹ค. ๊ตญ์™ธ์—์„œ๋„ ๋„์‹œ๋‹จ์œ„์˜ ์œ ์‚ฌ ์—ฐ๊ตฌ๋Š” ๋‹ค์ˆ˜ ์ง„ํ–‰์ค‘์ด๋‹ค. Din and marnerides(13)๋Š” ๊ตญ์ œํ‘œ์ค€ํ™”๊ธฐ๊ตฌ์—์„œ ์ œ๊ณตํ•˜๋Š” ์˜๊ตญ๊ณผ ๋ฏธ๊ตญ๋“ฑ ์ฃผ์š” 6๊ฐœ ๋„์‹œ์˜ ์ „๊ธฐ ์‚ฌ์šฉ๋Ÿ‰์„ ํ™•๋ณดํ•˜๊ณ  ์‹œ๊ฐ, ๋‚ ์งœ, ์ „๊ธฐ๊ฐ€๊ฒฉ, ์˜จ๋„ ๋“ฑ ์ „๊ธฐ ์‚ฌ์šฉ๋Ÿ‰์— ์˜ํ–ฅ์„ ์ฃผ๋Š” ์ธ์ž๋ฅผ ์ž…๋ ฅ๋ฐ์ดํ„ฐ๋กœ ์ „๊ธฐ ์ˆ˜์š”๋ฅผ ์˜ˆ์ธกํ•˜๋Š” ์—ฐ๊ตฌ๋ฅผ ์ง„ํ–‰ํ•˜์˜€๋‹ค. Wanhe(1) ์—ญ์‹œ, ์™ธ๊ธฐ์˜จ๋„์™€ ์Šต๋„ ๋“ฑ ๊ธฐํ›„ ์ •๋ณด๋งŒ์œผ๋กœ ํ•™์Šต๋œ ๋”ฅ๋Ÿฌ๋‹ ๋ชจ๋ธ๋กœ ์ค‘๊ตญ ๋ถ๋ถ€์ง€๋ฐฉ์˜ ์ „๊ธฐ์—๋„ˆ์ง€ ์‚ฌ์šฉ๋Ÿ‰์„ ์˜ˆ์ธกํ•˜์˜€๋‹ค.

ํ•œํŽธ, Marino et al.(14)์˜ ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด ๋„์‹œ, ๊ตญ๊ฐ€ ๋“ฑ ํฐ ๊ทœ๋ชจ์˜ ์ „๋ ฅ์˜ˆ์ธก์€ ๋‹จ์ผ ๊ฑด๋ฌผ๋‹จ์œ„์˜ ์˜ˆ์ธก๋ณด๋‹ค ์ƒ๋Œ€์ ์œผ๋กœ ์‰ฝ๋‹ค๊ณ  ์–ธ๊ธ‰ํ•˜์˜€๋‹ค. ํ•ด๋‹น ์—ฐ๊ตฌ์—์„œ๋Š” ๊ณผ๊ฑฐ 5๋…„ ๋™์•ˆ์˜ ๊ฑด๋ฌผ ์ „๊ธฐ์—๋„ˆ์ง€ ์‚ฌ์šฉ๋Ÿ‰์„ ํ™•๋ณดํ•œ ํ›„ ๋‹ค์–‘ํ•œ ์กฐ๊ฑด์—์„œ ๋”ฅ๋Ÿฌ๋‹ ํ•™์Šต์„ ํ†ตํ•ด ์„ฑ๋Šฅ์„ ๋น„๊ตํ•˜์˜€๋‹ค. ๋‹ค๋งŒ ์—ฐ๊ตฌ์˜ ๋ชฉํ‘œ๋ฅผ ๋ชจ๋ธ์˜ ํ•™์Šต ์„ฑ๋Šฅ์— ๋‘์–ด ๋‹ค์–‘ํ•œ ์กฐ๊ฑด์—์„œ ํ•™์Šต๊ธฐ๊ฐ„์˜ ๋”ฅ๋Ÿฌ๋‹ ์„ฑ๋Šฅ์€ ํ™•์ธ ํ•  ์ˆ˜ ์žˆ์—ˆ์œผ๋‚˜ ์˜ˆ์ธก ์‹œ ์žฌ์‹ค์ƒํ™ฉ ๋ณ€ํ™” ๋“ฑ ์˜ˆ์ธก๋ถˆ๊ฐ€๋Šฅํ•œ ์—๋„ˆ์ง€ ์‚ฌ์šฉ ํŒจํ„ด์ด ๋ฐœ์ƒ๋  ๊ฒฝ์šฐ ํฐ ์˜ค์ฐจ๋ฅผ ๋ณด์˜€๋‹ค. Hosein(15)์€ ๊ธฐ์กด์˜ ์—ฐ๊ตฌ์™€ ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ ์‹œ๊ฐ„, ์™ธ๊ธฐ์˜จ๋„, ์Šต๋„ ๋“ฑ ํ™˜๊ฒฝ์  ์กฐ๊ฑด๊ณผ ๋”๋ถˆ์–ด ์ฃผ๋ง, ์ฃผ์ค‘ ๋“ฑ ์—๋„ˆ์ง€ ์‚ฌ์šฉ ํŒจํ„ด๊ณผ ์˜ํ–ฅ์ด ์žˆ๋Š” ์ž…๋ ฅ์กฐ๊ฑด์„ ์ถ”๊ฐ€ํ•˜์˜€๊ณ  ์ด๋ฅผ ํ†ตํ•ด ์—๋„ˆ์ง€ ์‚ฌ์šฉ๋Ÿ‰์ด ํ•™์Šต ํŒจํ„ด๋ณด๋‹ค ๊ธ‰๊ฒฉํžˆ ์ƒ์Šนํ–ˆ๋˜ ์ฃผ๋ง์„ ์ œ์™ธํ•˜๋ฉด ์˜๋ฏธ์žˆ๋Š” ์˜ˆ์ธก๊ฒฐ๊ณผ๋ฅผ ์ œ๊ณตํ•˜์˜€๋‹ค.

์„ ํ–‰์—ฐ๊ตฌ์—์„œ๋„ ์•Œ ์ˆ˜ ์žˆ๋“ฏ์ด ๋”ฅ๋Ÿฌ๋‹์„ ํ†ตํ•œ ์ „๊ธฐ์ˆ˜์š” ์˜ˆ์ธก์€ ํ•™์Šต ์„ฑ๋Šฅ์€ ์šฐ์ˆ˜ํ•˜์ง€๋งŒ ์žฌ์‹คํ–‰๋™๊ณผ ๊ฐ™์€ ๋น„๋ฌผ๋ฆฌ์ ์ธ ์†์„ฑ์— ๋Œ€ํ•œ ์˜ˆ์ธก์„ฑ๋Šฅ์€ ํ˜„์ €ํžˆ ๋–จ์–ด์กŒ๋‹ค. ํ•œํŽธ, ์ผ๋ฐ˜์ ์ธ ๊ฒฝ์šฐ ์ „๊ธฐ ์‚ฌ์šฉ๋Ÿ‰ ํŒจํ„ด์€ ํ•ด๋‹น ์‹œ๊ฐ„๋Œ€์— ์–ผ๋งˆ๋‚˜ ๋งŽ์€ ๊ฑฐ์ฃผ์ž๊ฐ€ ์žˆ๋Š”์ง€๋ฅผ ๊ฐ€๋ฆฌํ‚ค๋Š” ์žฌ์‹ค๋ฅ (0-1)๋ณด๋‹ค ์žฌ์‹ค์œ ๋ฌด(1/0)๊ฐ€ ํฌ๊ฒŒ ์˜ํ–ฅ์„ ๋ฏธ์นœ๋‹ค.(16) Kim et al.(17)์€ ์žฌ์‹ค์ž์˜ ๊ฑฐ์ฃผ ํŒจํ„ด์€ ๊ฑด๋ฌผ์˜ ์—๋„ˆ์ง€ ์‹œ๋ฎฌ๋ ˆ์ด์…˜์—์„œ ๋งค์šฐ ์ค‘์š”ํ•œ ์ธ์ž์ด๋‚˜ ์žฌ์‹ค์ž์˜ ์Šค์ผ€์ค„์„ ์™„๋ฒฝํžˆ ์˜ˆ์ธกํ•˜๋Š” ๊ฒƒ์€ ๋ถˆ๊ฐ€๋Šฅํ•˜์—ฌ ๋Œ€๋ถ€๋ถ„์˜ ์‹œ๋ฎฌ๋ ˆ์ด์…˜์—์„œ ๊ณ ๋ คํ•˜์ง€ ์•Š๊ฑฐ๋‚˜ ๊ณ ์ • ๊ฐ’์„ ์‚ฌ์šฉํ•œ๋‹ค๊ณ  ์–ธ๊ธ‰ํ•˜์˜€๋‹ค. ๋‹ค๋งŒ, ์ตœ๊ทผ์—ฐ๊ตฌ์—์„œ๋Š” ๊ฐ๊ฐ, ์ง€๊ฐ ์‹ฌ๋ฆฌ์šด๋™ ๋“ฑ ์ธ์ง€์ ์ธ ์ ‘๊ทผ๋ฐฉ๋ฒ•์„ ํ†ตํ•ด ์žฌ์‹ค์ž์˜ ํ–‰๋™ํŒจํ„ด์„ ์˜ˆ์ธกํ•˜๋Š” ๋ชจ๋ธ๋ง ๊ธฐ๋ฒ•์ด ์ œ์•ˆ๋˜๊ณ  ์žˆ๋‹ค.(18-20) ์ด์™€ ํ•จ๊ป˜ ์˜ˆ์ธก์ œ์–ด์— ํ•„์š”ํ•œ ๋‹ค์Œ๋‚ ์˜ ์žฌ์‹ค์œ ๋ฌด์™€ ๊ฐ™์€ ์ •๋ณด์˜ ์ •ํ™•์„ฑ์€ ๊ฑฐ์ฃผ์ž๊ฐ€ ์˜ˆ์ธก๋‹จ๊ณ„์—์„œ ์ ๊ทน์ ์ธ ๊ฐœ์ž…์„ ํ†ตํ•ด ๋”์šฑ ๊ฐœ์„ ๋  ์ˆ˜ ์žˆ๋‹ค. ๋‹ค์–‘ํ•œ ์Šค๋งˆํŠธ ๊ธฐ๊ธฐ์˜ ์ถœํ˜„์€ ์ด๋ฅผ ๋”์šฑ ์†์‰ฝ๊ฒŒ ํ•  ๊ฒƒ์ด๋‹ค. ๋”ฐ๋ผ์„œ ๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” ์žฌ์‹ค์œ ๋ฌด์— ๋Œ€ํ•œ ์ •๋ณด๊ฐ€ ํ™•๋ณด ๋˜์—ˆ์„ ๋•Œ ๋‹ค์Œ๋‚ ์˜ ์ „๊ธฐ ์ˆ˜์š”๋ฅผ ์˜ˆ์ธกํ•˜๋Š” ๋ชจ๋ธ์„ ๋”ฅ๋Ÿฌ๋‹์„ ํ†ตํ•ด ์ œ์•ˆํ•˜๊ณ ์ž ํ•œ๋‹ค.

2. ์—ฐ๊ตฌ ๋ฐฉ๋ฒ• ๋ฐ ๋ฒ”์œ„

์•ž์„œ ์–ธ๊ธ‰ํ–ˆ๋“ฏ์ด ๋ณธ ์—ฐ๊ตฌ์˜ ๋ชฉ์ ์€ ์žฌ์‹ค์œ ๋ฌด ์ •๋ณด๊ฐ€ ์˜ˆ์ธก๋˜์—ˆ์„ ๋•Œ, ๊ณผ๊ฑฐ ์ „๊ธฐ ์‚ฌ์šฉ ๋ฐ์ดํ„ฐ์™€ ์‹œ๊ฐ„์ •๋ณด๋งŒ์œผ๋กœ ์ „๊ธฐ ์‚ฌ์šฉ๋Ÿ‰์„ ์˜ˆ์ธกํ•˜๋Š” ๋ชจ๋ธ์„ ์ œ์•ˆํ•˜๋Š”๋ฐ ์žˆ๋‹ค. ์ด๋ฅผ ์œ„ํ•ด ๋”ฅ๋Ÿฌ๋‹ ์‹ ๊ฒฝ๋ง์„ ์ด์šฉํ•˜์˜€๊ณ  ๋ชจ๋ธ ์ž…๋ ฅ ๊ฐ’์€ ์‹œ๊ฐ„ (time of the day), ๊ธฐ์ €๋ถ€ํ•˜, ์žฌ์‹ค์ •๋ณด ๋“ฑ ์† ์‰ฝ๊ฒŒ ์ž…๋ ฅ์ด ๊ฐ€๋Šฅํ•œ ๋ฐ์ดํ„ฐ๋งŒ์„ ์‚ฌ์šฉํ•ด ์‹ค์šฉ์ ์ธ ๋ชจ๋ธ ์‚ฌ์šฉ์„ฑ์„ ๊ณ ๋ คํ•˜๊ณ ์žํ•œ๋‹ค.

์ „๊ธฐ ์—๋„ˆ์ง€ ์‚ฌ์šฉ๋Ÿ‰์€ ์ธก์ •๊ฐ’์„ ์‚ฌ์šฉํ•˜์˜€์œผ๋ฉฐ, ์‹œ๊ฐ„๋ณ„ ์ „๊ธฐ์—๋„ˆ์ง€ ์‚ฌ์šฉ๋Ÿ‰์„ ์ธก์ •ํ•˜๋Š” ๊ธฐ๊ธฐ๋ฅผ ์„ธ๋Œ€ ๋ฐฐ์ „๋ฐ˜์— ์„ค์น˜ํ•˜์—ฌ ํ™•๋ณดํ•˜์˜€๋‹ค. ๋Œ€์ƒ๊ฑด๋ฌผ์€ ์ธ์ฒœ์ง€์—ญ ๊ฑฐ์ฃผ๋ฉด์  65 m2์˜ 1์ธ ์ฃผ๊ฑฐ๊ฑด๋ฌผ ํ•œ ์„ธ๋Œ€๋กœ ํ•˜์˜€์œผ๋ฉฐ, ์ธก์ •๊ธฐ๊ฐ„์€ ๋ƒ‰๋‚œ๋ฐฉ ์ „๊ธฐ์—๋„ˆ์ง€๊ฐ€ ๋ฐœ์ƒํ•˜์ง€ ์•Š๋Š” ์ค‘๊ฐ„๊ธฐ๋กœ ํ•˜์˜€๋‹ค.

ํ•™์Šต์€ RNN(Reccurent Neural Network) ๋”ฅ๋Ÿฌ๋‹ ๊ตฌ์กฐ ์ค‘ LSTM(Long Short Term Memory) ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ํ†ตํ•ด ์ง„ํ–‰ํ•˜์˜€์œผ๋ฉฐ LSTM์„ ๊ตฌ์„ฑํ•˜๋Š” Hidden layer๋Š” ๋‹จ์ผ Hidden layer ๊ตฌ์กฐ์ธ Single layer ๋ชจ๋ธ๊ณผ ์ข€ ๋” ์‹ฌ์ธตํ™” ๋œ ํ˜•ํƒœ์˜ Deeper layer ์ผ€์ด์Šค๋กœ ๊ตฌ์„ฑํ•ด ํ•™์Šต, ์˜ˆ์ธก ์„ฑ๋Šฅ์„ ๋ถ„์„ ํ•˜์˜€๋‹ค. ๋ณธ ์—ฐ๊ตฌ์—์„œ ์ œ์•ˆ๋œ LSTM ๋ชจ๋ธ์€ MATLAB ์—์„œ ์ œ๊ณตํ•˜๋Š” ํ•จ์ˆ˜๋ฅผ ํ†ตํ•ด ๊ตฌํ˜„๋˜์—ˆ๋‹ค.

3. ๋”ฅ๋Ÿฌ๋‹ ๊ธฐ๋ฐ˜ ์ „๊ธฐ ์ˆ˜์š” ์˜ˆ์ธก ๋ชจ๋ธ

๋”ฅ๋Ÿฌ๋‹ ๋ชจ๋ธ์˜ ๋Œ€ํ‘œ์ ์ธ ๊ตฌ์กฐ๋Š” CNN(Convolutional Neural Network)๊ณผ RNN ๊ตฌ์กฐ๋กœ ๋ถ„๋ฅ˜๋˜๋ฉฐ, Lee(21)์— ๋”ฐ๋ฅด๋ฉด CNN ๊ตฌ์กฐ๋Š” ์ˆœ์„œ๊ฐ€ ์ค‘์š”ํ•˜์ง€ ์•Š์€ ์ •๋ณด์— ๋Œ€ํ•œ ํ•™์Šต์— ๋›ฐ์–ด๋‚˜๊ณ , RNN์€ ์‹œ๊ณ„์—ด์  ํŠน์„ฑ์„ ๋ฐ˜์˜ํ•˜๊ธฐ ์ข‹์€ ๋”ฅ๋Ÿฌ๋‹ ๋ชจ๋ธ์ž„์„ ์–ธ๊ธ‰ํ•˜์˜€๋‹ค. ๋ณธ ์—ฐ๊ตฌ์˜ ์˜ˆ์ธก ๋Œ€์ƒ์€ ์‹œ๊ฐ„ ํ๋ฆ„์— ๋”ฐ๋ฅธ ํ•˜๋ฃจ ์—๋„ˆ์ง€ ์‚ฌ์šฉ๋Ÿ‰์œผ๋กœ ์ด๋Ÿฌํ•œ ๋ฐ์ดํ„ฐ๋Š” ์ด์ „ ์‹œ๊ฐ„๋Œ€์˜ ์‚ฌ์šฉ ํŒจํ„ด์ด ์ผ์ •๋ถ€๋ถ„ ์ง€์†๋˜๋Š” ์‹œ๊ณ„์—ด์  ํŠน์„ฑ์„ ๋„๊ธฐ์— RNN ๊ตฌ์กฐ๋ฅผ ํ†ตํ•ด ๋ชจ๋ธ์„ ๊ฐœ๋ฐœํ•˜๊ณ ์žํ•œ๋‹ค. ๋‹ค๋งŒ, RNN์˜ ๊ฒฝ์šฐ ๋งŽ์€ ์ˆ˜์˜ ์‹œ๊ณ„์—ด์  ์ž…๋ ฅ ๋ฐ์ดํ„ฐ๋ฅผ ์ฒ˜๋ฆฌ ํ•  ๋•Œ ์—๋Ÿฌ๊ฐ€ ๊ธ‰๊ฒฉํžˆ ์ƒ์Šนํ•˜๋Š” ๋ฌธ์ œ๊ฐ€ ๋ฐœ์ƒํ•˜๊ฒŒ ๋œ๋‹ค.(22) ๋”ฐ๋ผ์„œ ์ด๋Ÿฌํ•œ ์˜ค๋ฅ˜๋ฅผ ์ˆ˜์ •ํ•œ RNN ๊ตฌ์กฐ์ธ LSTM ๋ชจ๋ธ์„ ํ†ตํ•ด ํ•™์Šต์„ ์ง„ํ–‰ํ•˜์˜€๋‹ค.(23)

LSTM์€ ํ•™์Šต ์„ฑ๋Šฅ์„ ๊ฒฐ์ •ํ•˜๋Š” ๋‹ค์–‘ํ•œ ์„ค์ •๋ณ€์ˆ˜๋ฅผ ์ง€์ •ํ•ด์•ผ ํ•˜๋ฉฐ, ํ•™์Šต์„ ์ง„ํ–‰ํ•˜๋Š” ์ตœ์ ํ™” ๊ธฐ๋ฒ•์„ ์šฐ์„ ์ ์œผ๋กœ ์„ ์ •ํ•ด์•ผ ํ•œ๋‹ค. MATLAB์—์„œ๋Š” LSTM์— ํ•™์Šต ์ตœ์ ํ™” ๊ธฐ๋ฒ•์œผ๋กœ ํ™•๋ฅ ์  ๊ฒฝ์‚ฌ ํ•˜๊ฐ•๋ฒ•์œผ๋กœ ์•Œ๋ ค์ง„ SGD (Stochastic Gradient Descent)์™€ Adam(Adaptive moment estimation) ์•Œ๊ณ ๋ฆฌ์ฆ˜์ด ์ œ๊ณต๋˜๊ณ  ์žˆ๋‹ค. SGD ๊ธฐ๋ฒ•์€ ๊ธฐ์šธ๊ธฐ๊ฐ€ ๊ฐ€์žฅ ํฐ ๋ฐฉํ–ฅ์œผ๋กœ ํƒ์ƒ‰ ์œ„์น˜๋ฅผ ์„ ์ •ํ•˜๋Š” ๊ณ„์‚ฐ์„ ๋ฐ˜๋ณตํ•ด ์ตœ์ ํ•ด๋กœ ์ด๋™ํ•˜๋Š” ๋‹จ์ˆœํ•œ ๊ตฌ์กฐ๋ฅผ ๊ฐ–๋Š” ์ตœ์ ํ•ด ๊ธฐ๋ฒ•์ด๋‹ค. ํ•ด๋‹น ๊ธฐ๋ฒ•์€ ๋ฐ˜๋ณต๊ณ„์‚ฐ์„ ์œ„ํ•œ ์ถฉ๋ถ„ํ•œ ์ž‘์—…ํ™˜๊ฒฝ์ด ๊ฐ–์ถ”์–ด์ง€์ง€ ์•Š์„ ๊ฒฝ์šฐ ์ตœ์ ํ•ด๋ฅผ ์ฐพ๋Š”๋ฐ ๋งŽ์€ ์‹œ๊ฐ„์„ ์†Œ๋น„ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ๋น„ํšจ์œจ์ ์ด๋‹ค.(24) Adam ์•Œ๊ณ ๋ฆฌ์ฆ˜์€ ์œ ๋™์ ์œผ๋กœ ํ•™์Šต๋ฅ ์„ ์กฐ์ ˆํ•ด ํšจ์œจ์ ์œผ๋กœ ์ตœ์ ํ•ด๋ฅผ ์ฐพ๋Š” ๊ธฐ๋ฒ•์œผ๋กœ ์•Œ๋ ค์ ธ ์žˆ๋‹ค.(25) ํ˜„์žฌ๊นŒ์ง€ ์•Œ๋ ค์ง„ ํ•™์Šต ์•Œ๊ณ ๋ฆฌ์ฆ˜ ์ค‘ ๋ชจ๋“  ์ƒํ™ฉ์—์„œ ํ•ญ์ƒ ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋‚˜ํƒ€๋‚ด๋Š” ์•Œ๊ณ ๋ฆฌ์ฆ˜์€ ์—†์œผ๋‚˜, ๋ณธ ์—ฐ๊ตฌ์˜ ๋ชฉ์ ์ด ๋‹ค์Œ๋‚ ์˜ ์ „๊ธฐ์—๋„ˆ์ง€ ์‚ฌ์šฉ๋Ÿ‰์„ ๋ฐ˜๋ณต์ ์œผ๋กœ ์˜ˆ์ธกํ•ด์•ผํ•˜๊ธฐ์— ํ•™์Šต์„ ํ†ตํ•œ ๋ชจ๋ธ ๊ตฌ์ถ•์‹œ๊ฐ„์„ ์ค„์ผ ์ˆ˜ ์žˆ๋Š” Adam ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ์„ ํƒํ•˜์˜€๋‹ค. ์–ธ๊ธ‰๋œ ํ•™์Šต ์•Œ๊ณ ๋ฆฌ์ฆ˜์€ ๋ชจ๋ธ์—์„œ ์ง€์ •๋œ Hidden Layer์™€ Hidden unit์˜ ๊ฐ€์ค‘์น˜๋ฅผ ์กฐ์ ˆํ•ด ์˜ค์ฐจ๋ฅผ ์ตœ์†Œํ™” ํ•˜๋Š”๋ฐ ์ด๋•Œ, ์ธ๊ณต์‹ ๊ฒฝ๋ง์˜ Hidden Layer์™€ Unit์„ ๊ตฌ์„ฑํ•˜๋Š”๋ฐ ์ •ํ™•ํ•œ ์ง€์นจ์ด๋‚˜ ๊ทœ์น™์€ ์ •ํ•ด์ ธ ์žˆ์ง€ ์•Š๊ณ  ์‚ฌ์šฉ์ž์˜ ๊ฒฝํ—˜์— ์˜์กดํ•ด์•ผ ํ•œ๋‹ค.(24) ๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” ํ•˜๋‚˜์˜ Hidden layer๋งŒ ๊ฐ–๋Š” Single LSTM ๋ชจ๋ธ๊ณผ ์—ฌ๋Ÿฌ Hidden layer๊ฐ€ ๋ฐฐ์—ด๋œ ํ˜•ํƒœ์ธ Deeper LSTM ๋ชจ๋ธ์„ ์„ค๊ณ„ํ•˜์—ฌ ์„ฑ๋Šฅ์„ ๋น„๊ตํ•˜์˜€์œผ๋ฉฐ, Deeper LSTM์€ 3๊ฐœ์˜ Layer๋กœ ๊ตฌ์„ฑํ•˜์˜€๋‹ค. Hidden unit์˜ ๊ฒฝ์šฐ ๋ณธ ์—ฐ๊ตฌ์—์„œ ์‚ฌ์šฉ๋œ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ํˆด์ธ Matlab์—์„œ ๊ธฐ๋ณธ๊ฐ’(26)์œผ๋กœ ์ œ๊ณตํ•˜๋Š” 250๊ฐœ๋ณด๋‹ค ์กฐ๊ธˆ ๋งŽ์€ Layer ๋‹น 300๊ฐœ์˜ Hidden unit์„ ์ ์šฉํ•˜์˜€๋‹ค.

๋‹ค์Œ์œผ๋กœ ๋”ฅ๋Ÿฌ๋‹์—์„œ ์„ฑ๊ณต์ ์ธ ํ•™์Šต์„ ์œ„ํ•ด์„œ๋Š” ๊ณผ๋„ํ•™์Šต ํ˜„์ƒ์ธ Over fitting์„ ์ฃผ์˜ํ•ด์•ผํ•˜๋Š”๋ฐ, ๊ฐ Unit์ด ๊ฐ–๋Š” ๊ฐ€์ค‘์น˜๊ฐ€ ๋‚ฎ์„์ˆ˜๋ก Over fitting์„ ํ”ผํ•  ์ˆ˜ ์žˆ๋‹ค๊ณ  ์•Œ๋ ค์ ธ ์žˆ๋‹ค. ๊ฐ€์ค‘์น˜๋ฅผ ์ž‘๊ฒŒ ํ•˜๋Š” ์ผ๋ฐ˜์ ์ธ ๋ฐฉ๋ฒ•์€ ์ตœ๋Œ€ํ•œ ์ž‘์€ Initial Learn Rate ๊ฐ’์œผ๋กœ๋ถ€ํ„ฐ ํ•™์Šต์„ ์‹œ์ž‘ํ•˜๋Š” ๊ฒƒ์ด๋‹ค. Matlab Documentation์—์„œ๋Š” 0.01์„ ๊ธฐ๋ณธ ์„ค์ • ๊ฐ’์œผ๋กœ ํ•˜๊ณ  ์žˆ์œผ๋ฉฐ, Adam ์•Œ๊ณ ๋ฆฌ์ฆ˜์—์„œ ๊ถŒ์žฅํ•˜๋Š” ์ดˆ๊ธฐ๊ฐ’์€ 0.001์ด๋‹ค.(26) ๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” ๊ถŒ์žฅ ๊ฐ’์ธ 0.001์„ ์‚ฌ์šฉํ•˜์—ฌ ํ•™์Šต์„ ์ง„ํ–‰ํ•˜์˜€๋‹ค. ์ฐธ๊ณ ๋กœ, ๋‚ฎ์€ ๊ฐ€์ค‘์น˜๋ฅผ ๋ชฉํ‘œ๋กœ ๊ณผํ•˜๊ฒŒ ์ž‘์€ ๊ฐ’์ด๋‚˜ 0์„ ์ดˆ๊ธฐ ๊ฐ’์œผ๋กœ ์„ค์ •ํ•˜๋ฉด ํ•™์Šต์ด ์ง„ํ–‰๋˜๋ฉด์„œ ๊ฐ€์ค‘์น˜๊ฐ€ ๊ฐฑ์‹ ๋˜์ง€ ์•Š๋Š” ์˜ค๋ฅ˜๊ฐ€ ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ๋‹ค. ํ•™์Šต์— ์‚ฌ์šฉ๋œ ๋ฐ์ดํ„ฐ๋Š” ์ •๊ทœํ™”(normalization) ๊ณผ์ •์„ ๊ฑฐ์ณค์œผ๋ฉฐ, ์ด๋Š” ํ•™์Šต๊ฒฐ๊ณผ๊ฐ€ ์™œ๊ณก๋˜๊ฑฐ๋‚˜ ๋ฐœ์‚ฐํ•˜๋Š” ๋ฌธ์ œ๋ฅผ ๋ฐฉ์ง€ํ•˜๋Š”๋ฐ ๋„์›€์ด ๋œ๋‹ค. ์—ฐ์‚ฐ์€ GPU(Graphics processing unit-GTX 1060 6GB) ๋ณ‘๋ ฌ ์ฒ˜๋ฆฌ ๊ธฐ์ˆ ์„ ํ†ตํ•ด ์ง„ํ–‰๋˜์—ˆ์œผ๋ฉฐ, GPU๋Š” ๋งŽ์€ ์ž…๋ ฅ๋ฐ์ดํ„ฐ๋ฅผ ๊ธฐ์กด์˜ ์—ฐ์‚ฐ์‹œ์Šคํ…œ๋ณด๋‹ค ๋น ๋ฅด๊ฒŒ ์ฒ˜๋ฆฌ ํ•  ์ˆ˜ ์žˆ๋Š” ์žฅ์ ์ด ์žˆ๋‹ค. ๊ธฐํƒ€ ์„ค์ • ๊ฐ’์€ Matlab์—์„œ ์ œ๊ณตํ•˜๋Š” ์ดˆ๊ธฐ ์„ค์ • ๊ฐ’์„ ์ ์šฉํ•˜์˜€๋‹ค.

4. ์ „๊ธฐ ์‚ฌ์šฉ๋Ÿ‰ ์ธก์ • ๋ฐ ์žฌ์‹ค์ •๋ณด

๋”ฅ๋Ÿฌ๋‹ ๋ชจ๋ธ์„ ํ•™์Šตํ•˜๊ณ  ํ…Œ์ŠคํŠธํ•˜๊ธฐ ์œ„ํ•œ ์ธก์ • ๋Œ€์ƒ์€ ์ผ์ธ ๊ฐ€๊ตฌ ๊ฑด๋ฌผ์ด๋‹ค. ๋”ฐ๋ผ์„œ ์žฌ์‹ค์ž ์ •๋ณด๋Š” ์žฌ์‹ค์œ ๋ฌด๋กœ ํ•œ์ •๋œ๋‹ค. ์„œ๋ก ์—์„œ๋„ ์–ธ๊ธ‰ํ–ˆ๋“ฏ์ด ๊ฐ€์ •์—์„œ ์ „๊ธฐ์‚ฌ์šฉ๋Ÿ‰์˜ ์‹œ๊ฐ„๋ณ„ ๋ณ€ํ™”๋Š” ์žฌ์‹ค์ž์ˆ˜์˜ ๋ณ€ํ™”๋ณด๋‹ค๋Š” ์žฌ์‹ค์œ ๋ฌด์— ๋”ฐ๋ฅธ ์˜ํ–ฅ์ด ํฌ๊ณ , ์ผ์ธ๊ฐ€๊ตฌ์˜ ๊ฒฝ์šฐ ์žฌ์‹ค์ƒํ™ฉ์ด ๋”์šฑ๋” ๊ฐ€๋ณ€์ ์ด๊ธฐ ๋•Œ๋ฌธ์— ์„ ์ •๊ฑด๋ฌผ์˜ ์˜ˆ์ธก์ด ์ข€ ๋” ์–ด๋ ค์šธ ์ˆ˜ ์žˆ๋‹ค.

๋ฐ์ดํ„ฐ ์ˆ˜์ง‘์€ ๋Œ€์ƒ ์„ธ๋Œ€์˜ ๋ฐฐ์ „๋ฐ˜์— ์ธ์ฝ”์–ด๋“œ์‚ฌ์—์„œ ์ œ๊ณตํ•˜๋Š” ์—๋„ˆํ†ก์ œํ’ˆ์„ ์„ค์น˜ํ•˜์—ฌ ์‹œ๊ฐ„๋ณ„ ์ „๊ธฐ์—๋„ˆ์ง€ ์‚ฌ์šฉ๋Ÿ‰์„ ํ™•๋ณดํ•˜์˜€๋‹ค(Fig. 1 ์ฐธ์กฐ).(27) ๋„คํŠธ์›Œํฌ ์ƒํƒœ ๋“ฑ์— ๋”ฐ๋ผ ํŠน์ •์‹œ์ ์—์„œ ์ธก์ •๋ฐ์ดํ„ฐ๊ฐ€ ๋ˆ„๋ฝ๋˜๋Š” ๋˜๋Š” ๋ฌธ์ œ๊ฐ€ ๋ฐœ์ƒํ•˜๋Š” ๊ฒฝ์šฐ, ํ•ด๋‹น์ผ์˜ ๋ฐ์ดํ„ฐ๋Š” ํ•™์Šต๋ฐ์ดํ„ฐ์—์„œ ์ œ์™ธํ•˜์˜€๋‹ค. ์ธก์ • ๊ธฐ๊ฐ„์€ ๋‚œ๋ฐฉ๋ถ€ํ•˜๊ฐ€ ๋ฐœ์ƒํ•˜์ง€ ์•Š๋Š” 4์›” 8์ผ๋ถ€ํ„ฐ 5์›” 18์ผ๊นŒ์ง€ 40์ผ ๋™์•ˆ ์ „๊ธฐ ์‚ฌ์šฉ๋Ÿ‰์„ ์ธก์ •ํ•˜์˜€๋‹ค. ์ธก์ •์€ Fig. 1์˜ ์˜ค๋ฅธ์ชฝ์ฒ˜๋Ÿผ ์ดˆ๋‹จ์œ„๋กœ ์ง„ํ–‰๋˜๋‚˜, ์„œ๋ฒ„์— ๊ธฐ๋ก๋˜๋Š” ๋ฐ์ดํ„ฐ๋Š” 15๋ถ„ ๊ฐ„๊ฒฉ์œผ๋กœ ์—๋„ˆ์ง€ ์‚ฌ์šฉ๋Ÿ‰์„ ๊ธฐ๋กํ•œ๋‹ค. ๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” 15๋ถ„ ๊ฐ„๊ฒฉ์˜ ์—๋„ˆ์ง€ ์‚ฌ์šฉ๋Ÿ‰ ๋ฐ์ดํ„ฐ๋ฅผ ํ•ฉ์‚ฐํ•ด 1์‹œ๊ฐ„ ๋™์•ˆ ๋ˆ„์ ๋œ ์—๋„ˆ์ง€ ์‚ฌ์šฉ๋Ÿ‰ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ํ˜•์„ฑํ•˜์˜€๋‹ค. ์ด๋Š” MPC์˜ ๊ฒฝ์šฐ ์ผ๋ฐ˜์ ์œผ๋กœ ํ•œ ์‹œ๊ฐ„ ๋‹จ์œ„์˜ ์šด์ „๊ณ„ํš์ด ์ด๋ฃจ์–ด์ง€๊ธฐ ๋•Œ๋ฌธ์ด๋‹ค.

Fig. 1. Installed enertalk device(left) and electricity consumption display(right).
../../Resources/sarek/KJACR.2019.31.1.022/fig1.png

LSTM ๋ชจ๋ธ์˜ ์ž…๋ ฅ ๊ฐ’์€ ์‹œ๊ฐ„(1~24h), ๊ธฐ์ €๋ถ€ํ•˜, ์žฌ์‹ค์ •๋ณด(0 or 1)๋กœ ํ•˜๋ฉฐ ์ถœ๋ ฅ ๊ฐ’์€ ์ „๊ธฐ ์—๋„ˆ์ง€ ์‚ฌ์šฉ๋Ÿ‰์œผ๋กœ ํ•œ๋‹ค. ์—ฌ๊ธฐ์„œ ๊ธฐ์ €๋ถ€ํ•˜๋Š” ํ•™์Šต๊ธฐ๊ฐ„ ๋ฐ์ดํ„ฐ ์ค‘ ์žฅ๊ธฐ๊ฐ„ ์žฌ์‹ค์ด ์—†์—ˆ๋˜ ๋‚ ์˜ ์ „๊ธฐ ์‚ฌ์šฉ๋Ÿ‰(W)์„ ์‚ฌ์šฉํ•˜์˜€์œผ๋ฉฐ, ์žฌ์‹ค์ •๋ณด๋Š” ์‹œ๊ฐ„๋ณ„ ์žฌ์‹ค ์œ ๋ฌด๋ฅผ ํŒ๋‹จํ•ด ๊ฑฐ์ฃผ์ƒํƒœ๋ฅผ 1, ๋น„๊ฑฐ์ฃผ ์ƒํƒœ๋ฅผ 0์œผ๋กœ ๋ณ€ํ™˜ํ•˜์—ฌ ์ž…๋ ฅ๊ฐ’์œผ๋กœ ์‚ฌ์šฉํ•˜์˜€๋‹ค. ์žฌ์‹ค ์œ ๋ฌด๋ฅผ ํŒ๋‹จํ•œ ๊ธฐ์ค€์€ ๊ฐ€์žฅ ๋†’์€ ๊ธฐ์ € ์—๋„ˆ์ง€ ์‚ฌ์šฉ๋Ÿ‰์„ ๋ณด์˜€๋˜ ์ˆœ๊ฐ„์ด 126 W์ž„์„ ๊ณ ๋ คํ•ด ์—๋„ˆ์ง€ ์‚ฌ์šฉ๋Ÿ‰์ด 130 W ์ดํ•˜ ์ผ ๊ฒฝ์šฐ ๋น„๊ฑฐ์ฃผ ์ƒํƒœ๋กœ ๋ถ„๋ฅ˜ ํ•˜์˜€๋‹ค. ๋‹ค๋งŒ, ์ด ๊ฒฝ์šฐ ์ทจ์นจ ์‹œ ์—๋„ˆ์ง€ ์‚ฌ์šฉ๋Ÿ‰์ด ๊ฐ์†Œ๋˜์–ด ๋น„ ๊ฑฐ์ฃผ๋กœ ๋ถ„๋ฅ˜๋  ์†Œ์ง€๊ฐ€ ์žˆ์–ด ํ–ฅํ›„ ์žฌ์‹ค ์„ผ์„œ ๊ฐ’์„ ์ž…๋ ฅ๊ฐ’์œผ๋กœ ์‚ฌ์šฉํ•˜๋Š” ๋“ฑ ๋ช…ํ™•ํ•œ ๋ถ„๋ฅ˜ ๊ธฐ์ค€์ด ํ•„์š”ํ•˜๋‹ค. ๊ณผ๊ฑฐ 20์ผ ๋™์•ˆ์˜ ์ธก์ •๋œ ๋ฐ์ดํ„ฐ๋ฅผ ํ†ตํ•ด ๋ชจ๋ธ์„ ํ•™์Šตํ•˜๊ณ  ๋‹ค์Œ๋‚  ์‹œ๊ฐ„๋ณ„ ์ „๊ธฐ ์ˆ˜์š”๋ฅผ ์˜ˆ์ธกํ•˜์˜€๋‹ค. ํ•˜๋ฃจ๋‹จ์œ„ ๋ฐ์ดํ„ฐ ํ†ต์‹ ์„ ๊ฐ€์ •ํ•˜์—ฌ ์ธก์ •๋ฐ์ดํ„ฐ๋ฅผ 24์‹œ๊ฐ„ ๋งˆ๋‹ค ์—…๋ฐ์ดํŠธํ•˜๋Š” ๊ณผ์ •์„ 2์ฃผ๊ฐ„ ๋ฐ˜๋ณตํ•˜์˜€๋‹ค.

์˜ˆ์ธก๊ฒฐ๊ณผ๋ฅผ MBE(Mean Bias Error), RMSE(CV), ์ƒ๋Œ€ ์˜ค์ฐจ(Error)๋ฅผ ํ†ตํ•ด ํ‰๊ฐ€ํ•˜์˜€์œผ๋ฉฐ ์‹(1)~์‹(3)์„ ํ†ตํ•ด์„œ ๊ฐ๊ฐ ๊ณ„์‚ฐ๋œ๋‹ค. MBE๋Š” ์˜ˆ์ธก ๊ฐ’์˜ ์น˜์šฐ์นœ ์ •๋„๋ฅผ ์˜๋ฏธํ•˜๋ฉฐ 0์— ๊ฐ€๊นŒ์šธ์ˆ˜๋ก ํ‰๊ท ๊ฐ’์ด ์ธก์ •๊ฐ’์— ๊ทผ์ ‘ํ•˜๋‹ค. CVRMSE๋Š” ๋ถ„์‚ฐ ์ •๋„๋ฅผ ๊ณ ๋ คํ•ด ๋ชจ๋ธ์˜ ์˜ค์ฐจ๋ฅผ ํŒŒ์•…ํ•˜๋Š” ์˜ค์ฐจ ๋ถ„์„๋ฐฉ๋ฒ•์œผ๋กœ ์˜ค์ฐจ์œจ(%)๋กœ ํ‘œํ˜„๋˜๋ฉฐ, ๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” ์ „์ฒด ๊ตฌ๊ฐ„์˜ ํ‰๊ท  ์ „๊ธฐ์‚ฌ์šฉ ๊ฐ’(Ev)์œผ๋กœ ๋‚˜๋ˆ„์–ด ๊ณ„์‚ฐํ•˜์˜€๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ ์‹(3)์€ ๋งค์‹œ๊ฐ„ ์ธก์ •๊ฐ’ ๋Œ€๋น„ ์ƒ๋Œ€์˜ค์ฐจ๋ฅผ ๋‚˜ํƒ€๋‚ธ๋‹ค.

(1)
M B E = โˆ‘ i = 1 n E P , i - E M , i n E v [ % ]

(2)
C V R M S E = โˆ‘ i = n E P , i - E M , i 2 n / E v [ % ]

(3)
E r r o r = E P , i - E M , i E M , i [ % ]

5. ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ๊ฒฐ๊ณผ ๋ฐ ๋ถ„์„

5.1 ์ „๊ธฐ ์ˆ˜์š” ์˜ˆ์ธก ๋ชจ๋ธ ํ•™์Šต๊ฒฐ๊ณผ

5.1.1 ๋‹จ์ธต LSTM ๋ชจ๋ธ

์•ž์„œ ์–ธ๊ธ‰ํ–ˆ๋“ฏ์ด ์ œ์•ˆ๋œ ํ•™์Šต๋ชจ๋ธ์€ ๊ณผ๊ฑฐ 20์ผ ๋™์•ˆ์˜ ์ธก์ •๋œ ์ž…๋ ฅ๋ฐ์ดํ„ฐ์™€ ์ถœ๋ ฅ๋ฐ์ดํ„ฐ์˜ ๊ด€๊ณ„๋ฅผ ๊ฐ€์žฅ ์ž˜ ๋ฌ˜์‚ฌํ•˜๋Š” ๋‰ด๋Ÿฐ๊ณผ ๋ ˆ์ด์–ด ์‚ฌ์ด์˜ ์—ฐ๊ฒฐ๊ฐ•๋„๋ฅผ ํ†ตํ•ด ๋ชจ๋ธ์„ ๊ตฌ์ถ•ํ•œ๋‹ค. ์—ฌ๊ธฐ์„œ ํ•™์Šต๊ธฐ๊ฐ„์˜ ๊ณผ๊ฑฐ ์‹œ๊ฐ„๋ณ„ ์—๋„ˆ์ง€ ์‚ฌ์šฉ๋Ÿ‰ ๋ณ€๋™ ํŒจํ„ด์„ ๋น„์Šทํ•˜๊ฒŒ ์œ ์ถ”ํ•˜๋Š” ๋ชจ๋ธ์„ ๊ตฌ์ถ•ํ•˜๊ฒŒ ๋  ๊ฒฝ์šฐ ๋น„ ํ•™์Šต ๊ธฐ๊ฐ„์ธ ๋‹ค์Œ๋‚ ์˜ ์ „๊ธฐ์—๋„ˆ์ง€ ์‚ฌ์šฉ๋Ÿ‰์„ ๋” ์ž˜ ๋ฌ˜์‚ฌ ํ•  ๊ฒƒ์œผ๋กœ ๊ธฐ๋Œ€ ํ•  ์ˆ˜ ์žˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” ์˜ˆ์ธก์— ์•ž์„œ ๋ชจ๋ธ์˜ ํ•™์Šต ์„ฑ๋Šฅ์„ ๋จผ์ € ๋ถ„์„ํ•˜์˜€๋‹ค. ๋น„๊ต๋ฅผ ์œ„ํ•ด ๋‹จ์ธต LSTM๊ณผ ๋™์ผํ•œ Layer์™€ Hidden untis์œผ๋กœ ๊ตฌ์„ฑ๋œ ANN ๋ชจ๋ธ์„ ์ด์šฉํ•˜์˜€์œผ๋ฉฐ ANN ํ•™์Šต ์•Œ๊ณ ๋ฆฌ์ฆ˜์€ ์˜ค์ฐจ๋ฅผ ์—ญ๋ฐฉํ–ฅ์œผ๋กœ ์ „๋‹ฌํ•ด ๋‰ด๋Ÿฐ์‚ฌ์ด ์—ฐ๊ฒฐ๊ฐ•๋„๋ฅผ ์กฐ์ ˆํ•˜๋Š” Feed-Forward Back Prop ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ์‚ฌ์šฉํ•˜์˜€๊ณ , ์ด๋Š” ์‹œ๊ณ„์—ด ๋ฐ์ดํ„ฐ๋ฅผ ๋ถ„์„ํ•ด ๋‹จ๊ธฐ ๋ถ€ํ•˜๋ฅผ ์˜ˆ์ธกํ•˜๋Š” ์—ฐ๊ตฌ์— ์ฃผ๋กœ ์‚ฌ์šฉ๋˜์—ˆ๋˜ ๋ชจ๋ธ์ด๋‹ค.

Fig. 2๋Š” 20์ผ ๋™์•ˆ์˜ ๋‹จ์ธต LSTM ๋ชจ๋ธ์˜ ํ•™์Šต ์„ฑ๋Šฅ์„ ๋ถ„์„ํ•œ ๊ทธ๋ž˜ํ”„๋กœ ๋ชจ๋ธ์€ ํ•™์Šต ๊ธฐ๊ฐ„ ๋™์•ˆ ์ „๊ธฐ์—๋„ˆ์ง€ ๋ฐœ์ƒ ๊ฑฐ๋™์„ ๊ฑฐ์˜ ์ •ํ™•ํžˆ ๋ฌ˜์‚ฌํ•˜์˜€๋‹ค. ๋ฐ˜๋ฉด ANN ๋ชจ๋ธ์€ ์—๋„ˆ์ง€์‚ฌ์šฉ ๋ฐœ์ƒ ํŒจํ„ด์„ ๋ฌ˜์‚ฌํ•˜๊ธด ํ•˜์˜€์œผ๋‚˜ ๊ธฐ์ €๋ถ€ํ•˜๊ฐ€ ๋ฐœ์ƒํ•˜๋Š” ์‹œ์ ์—์„œ ํฐ ์˜ค์ฐจ๋ฅผ ๋ณด์˜€๋‹ค. ๊ทธ๋ฆผ์—์„œ ANN ๋ชจ๋ธ์˜ Hidden Layer ๊ตฌ์„ฑ์€ LSTM ๋ชจ๋ธ๊ณผ ๋™์ผํ•˜์˜€์œผ๋‚˜ ์ตœ์ ํ™” ๊ธฐ๋ฒ•์˜ ์ฐจ์ด์—์„œ ์˜ค์ฐจ๊ฐ€ ๋ฐœ์ƒํ•œ ๊ฒƒ์œผ๋กœ ํŒ๋‹จ๋œ๋‹ค. ๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” ๋น„๊ต๋ฅผ ์œ„ํ•ด ๋™์ผํ•œ ์ˆ˜์˜ ํžˆ๋“  Unit์„ ์‚ฌ์šฉํ•˜์˜€์ง€๋งŒ ANN์„ ์œ„ํ•ด ๊ทธ ์ˆ˜๋ฅผ ํฌ๊ฒŒ ์ฆ๊ฐ€ ์‹œํ‚ค๋ฉด LSTM๊ณผ ๋™์ผํ•œ ์ˆ˜์ค€์˜ ํ•™์Šต ํšจ๊ณผ๋ฅผ ๋ณด์˜€๋‹ค.

Fig. 2. Model learning performance : ANN vs. Single LSTM.
../../Resources/sarek/KJACR.2019.31.1.022/fig2.png

Fig. 3์˜ ์‚ฐ์ ๋„๋Š” ์ ๋“ค์ด ๋Œ€๊ฐ์„ ์— ๊ฐ€๊นŒ์šธ์ˆ˜๋ก ๋ชจ๋ธ์ด ์ •ํ™•ํ•˜๋‹ค๊ณ  ํ•  ์ˆ˜ ์žˆ๋Š”๋ฐ ๋‹จ์ธต LSTM์˜ ๊ฒฝ์šฐ ๋Œ€๋ถ€๋ถ„์˜ ์ ๋“ค์ด ๋Œ€๊ฐ์„ ๊ณผ ๊ฑฐ์˜ ์ผ์น˜ํ•˜๋Š” ๊ฒƒ์„ ํ™•์ธ ํ•  ์ˆ˜ ์žˆ๋‹ค. ๋‹จ์ธต LSTM ๋ชจ๋ธ์€ ์†Œ์ˆ˜์  ๋‘˜์งธ ์ž๋ฆฌ๊นŒ์ง€ MBE ์˜ค์ฐจ๋ฅผ ๋ณด์ด์ง€ ์•Š์•˜์œผ๋ฉฐ CVRMSE๋Š” ๊ฐ๊ฐ 0.14% ์ •๋„์˜ ์˜ค์ฐจ๋งŒ ๋ณด์˜€๋‹ค. ANN ๋ชจ๋ธ์˜ ๊ฒฝ์šฐ ์ฃผ์–ด์ง„ ํ•™์Šต์กฐ๊ฑด์—์„œ 3.7%์˜ MBE๋ฅผ ๋‚˜ํƒ€๋ƒˆ์œผ๋‚˜ ๋ถ„์‚ฐ์„ ๋‚˜ํƒ€๋‚ด๋Š” CVRMSE๋Š” Fig. 3๊ณผ ๊ฐ™์ด 50%๊ฐ€ ๋„˜๋Š” ์—๋Ÿฌ๋ฅผ ๋‚˜ํƒ€๋‚ด์—ˆ๋‹ค.

Fig. 3. Scatter plot of Fig. 2 : ANN vs. Single LSTM.
../../Resources/sarek/KJACR.2019.31.1.022/fig3.png

5.1.2 ๋‹ค์ธต LSTM ๋ชจ๋ธ

๋‹ค์ธต LSTM ๋ชจ๋ธ ์—ญ์‹œ ๋‹จ์ธต LSTM ๋ชจ๋ธ๊ณผ ๊ฐ™์€ ๊ธฐ๊ฐ„ ๋™์•ˆ์˜ ํ•™์Šต ์„ฑ๋Šฅ์„ ๋น„๊ตํ•˜์˜€๋‹ค. ๋‹ค์ธต LSTM ๋ชจ๋ธ์˜ Hidden Layer๋งˆ๋‹ค ๊ตฌ์„ฑ๋˜๋Š” Hidden Unit์˜ ๊ฐœ์ˆ˜๋Š” Single LSTM ๋ชจ๋ธ๊ณผ ๋™์ผํ•˜๋‚˜, Hidden Layer์˜ ๋ฐฐ์—ด์„ 3์—ฐ์†์œผ๋กœ ํ•˜์—ฌ ์ธต๋งˆ๋‹ค ๊ณ ๋ ค๋˜๋Š” ๋ณ€์ˆ˜๋ฅผ ๋Š˜๋ ค ์—ฐ์‚ฐ์˜ ๊ฒฝ์šฐ์˜ ์ˆ˜๋ฅผ ๋Š˜๋ฆฐ ์‹ฌ์ธตํ™”๋œ ๋ชจ๋ธ์ด๋‹ค. ๋”ฐ๋ผ์„œ ๋ณธ ์—ฐ๊ตฌ์—์„œ ์ œ์•ˆ๋œ ๋‹ค์ธต LSTM ๋ชจ๋ธ์˜ Hidden Unit ๊ฐœ์ˆ˜๋Š” 900๊ฐœ(300ร—3Layer)์ด๋‹ค. Layer์™€ Unit์ด ๋Š˜์–ด๋‚˜๋ฉด, ์ฆ‰ ์‹ ๊ฒฝ๋ง์ด ๊นŠ์–ด์งˆ์ˆ˜๋ก ํ•™์Šต์— ์†Œ์š”๋˜๋Š” ์‹œ๊ฐ„์€ ์ฆ๊ฐ€ํ•˜๋‚˜ Hidden Layer์˜ ์ฆ๊ฐ€๋Š” ๋ณต์žกํ•œ ๋ฌธ์ œ๋ฅผ ํ‘ธ๋Š”๋ฐ ์œ ๋ฆฌํ•˜๋‹ค๊ณ  ์•Œ๋ ค์ ธ ์žˆ๋‹ค.(14) ANN ๋ชจ๋ธ ์—ญ์‹œ ๋™์ผํ•œ ์‹ฌ์ธตํ™”๋œ ๋ชจ๋ธ์„ ๊ตฌ์ถ•ํ•ด ๋‹จ์ธต layer์™€ ๋‹ค์ธต layer์˜ ์„ฑ๋Šฅ์„ ๋น„๊ตํ•˜๋ ค ํ•˜์˜€์œผ๋‚˜, LSTM ๋ชจ๋ธ๊ณผ ๋™์ผํ•œ layer ๊ฐœ์ˆ˜๋ฅผ ์ ์šฉํ•  ์‹œ Matlab์ด ANN ํ•จ์ˆ˜์— ํ• ๋‹นํ•˜๋Š” ๋ฉ”๋ชจ๋ฆฌ๋ฅผ ์ดˆ๊ณผํ•˜๋Š” ์˜ค๋ฅ˜๊ฐ€ ๋ฐœ์ƒํ•˜์˜€๋‹ค.

Fig. 4๋Š” ๋‹ค์ธต LSTM์˜ ํ•™์Šต์„ฑ๋Šฅ์„ ์˜๋ฏธํ•˜๋ฉฐ ๊ทธ๋ž˜ํ”„์—์„œ ์•Œ ์ˆ˜ ์žˆ๋“ฏ์ด ๋ชจ๋“  ์ ๋“ค์ด ๋Œ€๊ฐ์„  ์ฃผ๋ณ€์— ๋ถ„ํฌํ•˜๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ๋‹ค. ๋‹ค์ธต LSTM ๋ชจ๋ธ ์—ญ์‹œ MBE๋Š” ์†Œ์ˆ˜ ๋‘˜์งธ์ž๋ฆฌ์—์„œ 0% ์ˆ˜์ค€์ด์—ˆ์œผ๋ฉฐ CVRMSE๋Š” ๋‹จ์ธต LSTM ๋ชจ๋ธ๋ณด๋‹ค ๊ฐœ์„ ๋œ 0.04% ์ˆ˜์ค€์˜ ์˜ค์ฐจ๋งŒ ๋ณด์˜€๋‹ค. ๊ฐ ๋ชจ๋ธ๋ณ„ ์˜ค์ฐจ๋Š” Table 1์— ์ •๋ฆฌํ•˜์˜€๋‹ค.

Fig. 4. Learning performance : Deeper LSTM.
../../Resources/sarek/KJACR.2019.31.1.022/fig4.png

Table 1. Learning error comparison among models

Hidden Unit

MBE(%)

RMSE(W)

CVRMSE(%)

Single LSTM

300

0.00

0.14

0.14

Deeper LSTM

300ร—3

0.00

0.04

0.04

Back prof ANN model

300

-3.73

57.37

54.60

5.2 ๋ชจ๋ธ์„ ์ด์šฉํ•œ ์ „๊ธฐ ์ˆ˜์š” ์˜ˆ์ธก๊ฒฐ๊ณผ

ํ•™์Šต๋œ LSTM ๋ชจ๋ธ์„ ํ†ตํ•ด ๋‹ค์Œ๋‚ ์˜ ์—๋„ˆ์ง€ ์‚ฌ์šฉ๋Ÿ‰์„ ์˜ˆ์ธกํ•˜๊ณ  ์‹ค์ œ ์ธก์ •๋œ ๋ฐ์ดํ„ฐ์™€ ๋น„๊ตํ•ด ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ•˜์˜€๋‹ค. Fig. 5์—๋Š” ํ•˜๋ฃจ๋‹จ์œ„ ์˜ˆ์ธก์„ฑ๋Šฅ๊ณผ ์˜ˆ์ธก์„ ์ง„ํ–‰ํ•˜๊ธฐ ์œ„ํ•ด ์„ ํ–‰๋˜์—ˆ๋˜ ํ•™์Šต ์„ฑ๋Šฅ ์ผ๋ถ€๋ฅผ ํ‘œํ˜„ํ•˜์˜€๋‹ค. ์ด์ „ ์ ˆ์˜ ๊ฒฐ๊ณผ ๊ทธ๋ž˜ํ”„์—์„œ ์•Œ ์ˆ˜ ์žˆ๋“ฏ์ด ํ•™์Šต ๊ธฐ๊ฐ„์—๋Š” ๋‘ ๋ชจ๋ธ ๋ชจ๋‘ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์˜€๋‹ค. ๋‹ค๋งŒ, ์˜ˆ์ธก์„ฑ๋Šฅ์—์„œ๋Š” ์—๋„ˆ์ง€ ์‚ฌ์šฉ๋Ÿ‰์˜ ์ž„์˜์„ฑ์„ ๋ฐ˜์˜ํ•˜์ง€ ๋ชปํ•ด ์‚ฌ์šฉํŒจํ„ด์„ ํ•™์Šต ์„ฑ๋Šฅ๋งŒํผ ์™„๋ฒฝํžˆ ๋ฌ˜์‚ฌํ•˜์ง€ ๋ชปํ•˜์˜€์œผ๋ฉฐ ํ•ด๋‹น ์ผ์—๋Š” ๋‹จ์ธต ๋ฐ ๋‹ค์ธต ๋ชจ๋ธ ๊ฐ๊ฐ 29%, 25% ์ˆ˜์ค€์˜ CVRMSE ์˜ค์ฐจ๋ฅผ ๋ณด์˜€๋‹ค. ๊ทธ๋ฆผ์—์„œ ์˜ˆ์ธก์ผ ์ƒˆ๋ฒฝ 1์‹œ๊ฒฝ ๋งค์šฐ ํฐ ์ „๊ธฐ์‚ฌ์šฉ๋Ÿ‰์„ ๊ธฐ๋กํ•˜์˜€๋Š”๋ฐ ํ•™์Šต๊ธฐ๊ฐ„๋™์•ˆ ํ•ด๋‹น ์‹œ๊ฐ„์—์„œ ์œ ์‚ฌํ•œ ํŒจํ„ด์ด ์—†์—ˆ๊ณ  ๋”ฅ๋Ÿฌ๋‹ ๋ชจ๋ธ์ด ๋งค์šฐ ํ•œ์ •์ ์ธ ์ž…๋ ฅ ๊ฐ’์„ ์‚ฌ์šฉํ•˜์˜€๊ธฐ ๋•Œ๋ฌธ์— ๋ฐœ์ƒํ•œ ๊ฒƒ์œผ๋กœ ๋ณด์ธ๋‹ค.

Fig. 5. Model performance comparison during learning and prediction periods: Single LSTM vs. Deeper LSTM.
../../Resources/sarek/KJACR.2019.31.1.022/fig5.png

Fig. 6์€ 20์ผ ํ•™์Šต ํ›„ ํ•˜๋ฃจ ์˜ˆ์ธก์„ 4๋ฒˆ ๋ฐ˜๋ณตํ•˜์—ฌ(sequencing) ์˜ˆ์ธก ๊ฒฐ๊ณผ๋งŒ์„ ์žฌ์‹ค ์Šค์ผ€์ค„๊ณผ ํ•จ๊ป˜ ํ‘œ์‹œํ•˜์˜€์œผ๋ฉฐ, ๋ณด๋ผ์ƒ‰ ๋ง‰๋Œ€๊ทธ๋ž˜ํ”„๋Š” ํ•ด๋‹น์‹œ๊ฐ„์— ์‚ฌ์šฉ์ž๊ฐ€ ๊ฑฐ์ฃผํ•˜์˜€์Œ์„ ์˜๋ฏธํ•œ๋‹ค. ๋‘ ๋ชจ๋ธ ๋ชจ๋‘ ์ „๊ธฐ์ˆ˜์š”๋ฅผ ์œ ์‚ฌํ•˜๊ฒŒ ๋ฌ˜์‚ฌํ•˜์˜€์œผ๋‚˜ ์žฌ์‹ค์ด ์—†๋Š” ๋น„๊ฑฐ์ฃผ ๊ตฌ๊ฐ„์—์„œ ๋‹จ์ธต LSTM ๋ชจ๋ธ์ด ๊ธฐ์ €๋ถ€ํ•˜๋ฅผ ๊ณผ์†Œ ์˜ˆ์ธกํ•˜๋Š” ํ˜„์ƒ์ด ๋ฐœ์ƒํ•˜์˜€์œผ๋‚˜ ์‹ฌ์ธตํ™”๋œ ๋‹ค์ธต LSTM ๋ชจ๋ธ์€ ์ข€ ๋” ์•ˆ์ •์ ์ธ ์˜ˆ์ธก ์„ฑ๋Šฅ์„ ๋ณด์˜€๋‹ค. ๋˜ํ•œ ๋‹ค์ธต LSTM ๋ชจ๋ธ์€ ๊ธฐ์ €๋ถ€ํ•˜๊ฐ€ ์ง€์†์ ์œผ๋กœ ๋ฐœ์ƒํ•˜๋‹ค ๊ฐ‘์ž๊ธฐ ์ „๊ธฐ์‚ฌ์šฉ๋Ÿ‰์ด ์ƒ์Šนํ•˜๋Š” ์‹œ์ (80 h) ๋“ฑ ์ผ๋ฐ˜์ ์ธ ๋ฌผ๋ฆฌ ๋ชจ๋ธ์ด ๊ตฌํ˜„ํ•˜๊ธฐ ์–ด๋ ค์šด ํŒจํ„ด๊นŒ์ง€ ๋ฌ˜์‚ฌํ•˜๋Š” ์„ฑ๋Šฅ์„ ๋‚˜ํƒ€๋‚ด์—ˆ๋‹ค.

Fig. 6. Sequencing and consumption prediction(first 4 days).
../../Resources/sarek/KJACR.2019.31.1.022/fig6.png

ํŠน์ • ์ƒํ™ฉ์—์„œ ๋งค์šฐ ๋›ฐ์–ด๋‚œ ์˜ˆ์ธก์„ฑ๋Šฅ์„ ๋ณด์ด๋Š” ๋”ฅ๋Ÿฌ๋‹์— ๋Œ€ํ•œ ์˜ˆ๋Š” ๋ฌด์ˆ˜ํžˆ ๋งŽ์ง€๋งŒ, ์•„์ง ์ด์— ๋Œ€ํ•œ ๋…ผ๋ฆฌ์  ์„ค๋ช…์€ ๋งŽ์ด ๋ถ€์กฑํ•œ ๊ฒƒ์ด ํ˜„์‹ค์ด๋‹ค.

๋งˆ์ง€๋ง‰์œผ๋กœ ํ•˜๋ฃจ๋‹จ์œ„ ์—๋„ˆ์ง€์‚ฌ์šฉ๋Ÿ‰ ์˜ˆ์ธก์„ 2์ฃผ ๋™์•ˆ ์ง„ํ–‰ํ•œ ๊ฒฐ๊ณผ๋ฅผ Fig. 7์— ๋‚˜ํƒ€๋‚ด์—ˆ๋‹ค. ๋จผ์ €, ๋‹จ์ธต LSTM ๋ชจ๋ธ์€ MBE-6%, CVRMSE 26% ์ˆ˜์ค€์˜ ์˜ˆ์ธก ์„ฑ๋Šฅ์„ ๋ณด์˜€๋‹ค. ์˜ˆ์ธก ๊ธฐ๊ฐ„ ์ „๋ฐ˜์— ๊ฑธ์ณ ์—๋„ˆ์ง€ ์‚ฌ์šฉ ํŒจํ„ด์€ ์œ ์‚ฌํ•˜๊ฒŒ ๋ฌ˜์‚ฌํ•˜์˜€์œผ๋‚˜ ๊ทธ๋ž˜ํ”„์˜ 48~72h๊ตฌ๊ฐ„์ฒ˜๋Ÿผ ๊ธฐ์ €๋ถ€ํ•˜ ๊ตฌ๊ฐ„์—์„œ ํŒจํ„ด์„ ๋ฒ—์–ด๋‚˜๊ฑฐ๋‚˜ 168~240h์ฒ˜๋Ÿผ ํ•™์Šต๋ฐ์ดํ„ฐ์—๋Š” ์กด์žฌํ•˜์ง€ ์•Š์•˜๋˜ ์žฅ๊ธฐ๊ฐ„ ๋น„๊ฑฐ์ฃผ ์ƒํƒœ์—์„œ ๋‘๋“œ๋Ÿฌ์ง„ ์˜ค์ฐจ๋ฅผ ๋ณด์˜€๋‹ค. ๋‹ค์ธต LSTM ๋ชจ๋ธ์€ ์–ธ๊ธ‰ํ•œ ๊ตฌ๊ฐ„์—์„œ ๋ณด๋‹ค ๊ฐœ์„ ๋œ ๊ฒฐ๊ณผ๋ฅผ ๋ณด์˜€์œผ๋‚˜ ์ „์ฒด์ ์ธ ์˜ค์ฐจ ์ˆ˜์ค€์˜ ๋‘ ๋ชจ๋ธ์—์„œ ๋น„์Šทํ•˜๊ฒŒ ๋‚˜ํƒ€๋‚ฌ๋‹ค. ๋‘ ๋ชจ๋ธ์˜ ํ•™์Šต์„ฑ๋Šฅ์€ Table 2์— ์ •๋ฆฌํ•˜์˜€๋‹ค.

Fig. 7. Sequencing and consumption prediction for 2 weeks : Single LSTM vs Deeper LSTM.
../../Resources/sarek/KJACR.2019.31.1.022/fig7.png

Table 2. Prediction performance by error estimation of the single and deeper LSTM models

Hidden Unit

MBE(%)

RMSE(W)

CVRMSE(%)

Single LSTM

300

-6.38

22.02

26.21

Deeper LSTM

300ร—3

-9.1

19.62

23.1

Table 3์€ ์˜ˆ์ธก๊ธฐ๊ฐ„(336์‹œ๊ฐ„)๋™์•ˆ ๋ฐœ์ƒํ•œ ์ƒ๋Œ€ ์˜ค์ฐจ์˜ ๋ถ„ํฌ ๊ตฌ๊ฐ„์„ ์˜๋ฏธํ•˜๋ฉฐ ๋Œ€๋ถ€๋ถ„์˜ ์˜ค์ฐจ๋Š” 20% ์ดํ•˜์—์„œ ๋ฐœ์ƒํ•˜์˜€๋‹ค. ๋‹ค๋งŒ, 60% ์ด์ƒ ์˜ค์ฐจ๋ฅผ ๋ณด์ธ ํšŒ์ˆ˜๋Š” ๋‹จ์ธต LSTM ๋ชจ๋ธ์ด 40์‹œ๊ฐ„ ๋‹ค์ธต LSTM์ด 28์‹œ๊ฐ„์œผ๋กœ ์‹ฌ์ธต๋ชจ๋ธ์ธ ๋‹ค์ธต LSTM ๋ชจ๋ธ์ด ๋‘๋“œ๋Ÿฌ์ง„ ์˜ค์ฐจ๊ฐ€ ์ ๊ฒŒ ๋‚˜ํƒ€๋‚ฌ์œผ๋ฉฐ, ์ด๋Š” ์•ž์„œ ์–ธ๊ธ‰ํ•œ ๊ธฐ์ €๋ถ€ํ•˜ ๊ตฌ๊ฐ„์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ๋‹ค. ์ œ์•ˆํ•œ ๋‘ LSTM ๋ชจ๋ธ ๋ชจ๋‘ ํ•˜๋ฃจ๋‹จ์œ„ ์ „๊ธฐ์‚ฌ์šฉ๋Ÿ‰ ์˜ˆ์ธก์— ์žˆ์–ด ์ž…๋ ฅ๋ฐ์ดํ„ฐ ์„ธํŠธ๊ฐ€ 20์ผ ์ •๋„๋กœ ๋น„๊ต์  ์ ์ง€๋งŒ ์˜๋ฏธ์žˆ๋Š” ์˜ˆ์ธก์„ฑ๋Šฅ์„ ๋ณด์˜€๋‹ค. ๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” ๋‹ค์–‘ํ•œ ์ข…๋ฅ˜์˜ ๋ฐ์ดํ„ฐ ์ˆ˜์ง‘์ด ์–ด๋ ค์šด ์ฃผ๊ฑฐ๊ฑด๋ฌผ์„ ๋Œ€์ƒ์œผ๋กœ ํ•˜์—ฌ ๊ฐ€์šฉํ•œ ์ž…๋ ฅ๋ฐ์ดํ„ฐ๋งŒ์„ ์‚ฌ์šฉํ•˜์—ฌ ๋ชจ๋ธ์„ ๊ตฌ์ถ•ํ•˜์˜€์œผ๋‚˜, ํ–ฅํ›„ ๊ฐ ๊ฐ€์ „๊ธฐ๊ธฐ์˜ ์†Œ๋น„์ „๋ ฅ ๋ฐ ์‚ฌ์šฉ ์Šค์ผ€์ค„ ๋“ฑ ํ•™์Šต์„ ์œ„ํ•ด ๋” ๋งŽ์€ ์ข…๋ฅ˜์˜ ๋ฐ์ดํ„ฐ ์„ธํŠธ๊ฐ€ ํ˜•์„ฑ๋œ๋‹ค๋ฉด ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์€ ๋”์šฑ ๊ฐœ์„ ๋  ๊ฒƒ์œผ๋กœ ํŒ๋‹จ๋œ๋‹ค.

Table 3. Frequencies of error distribution

0 < Error < 20

20 โ‰ค Error < 40

40 โ‰ค Error < 60

60 โ‰ค Error < 80

80 < Error(%)

Total

Single LSTM

201

69

26

16

24

336

Deeper LSTM

224

63

21

9

19

336

6. ๊ฒฐ ๋ก 

๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” ์˜ˆ์ธก์ œ์–ด๋ฅผ ์œ„ํ•œ ์šด์ „๊ณ„ํš ์ˆ˜๋ฆฝ์„ ์œ„ํ•ด ์ „๊ธฐ ์ˆ˜์š” ์˜ˆ์ธก์„ ๋”ฅ๋Ÿฌ๋‹ ๋ชจ๋ธ์„ ์ด์šฉํ•˜์—ฌ ์ˆ˜ํ–‰ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์ œ์•ˆํ•˜์˜€๋‹ค. ํ•™์Šต์„ ์œ„ํ•ด ์‚ฌ์šฉ๋œ ์ž…๋ ฅ๋ฐ์ดํ„ฐ๋Š” ์‹œ๊ฐ, ๊ธฐ์ €๋ถ€ํ•˜, ์žฌ์‹ค์ •๋ณด๋กœ ๋น„๊ต์  ๋‹จ์ˆœํ•œ ๋ฐ์ดํ„ฐ๋งŒ์„ ์‚ฌ์šฉํ•˜์˜€๊ณ , ์—ฐ๊ตฌ์˜ ์‹คํšจ์„ฑ์„ ๋†’์ด๊ธฐ ์œ„ํ•ด ์‹ค์ œ ๊ฑด๋ฌผ์—์„œ ์ธก์ •ํ•œ ์ „๋ ฅ ๋ฐ์ดํ„ฐ๋ฅผ ์ด์šฉํ•˜์˜€๋‹ค.

ํ•™์Šต์€ ๋”ฅ๋Ÿฌ๋‹ LSTM ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ์‚ฌ์šฉํ•ด ์ง„ํ–‰๋˜์—ˆ์œผ๋ฉฐ, LSTM์€ ๊ตฌ์„ฑ๋˜๋Š” ๋ ˆ์ด์–ด ๊ฐœ์ˆ˜๋ฅผ ๋‹จ์ธต๊ณผ ๋‹ค์ธต์œผ๋กœ ์„ค๊ณ„ํ•ด ๋‘ ์ผ€์ด์Šค์˜ ํ•™์Šต๊ณผ ์˜ˆ์ธก ์ •ํ™•๋„๋ฅผ ์‹ค์ œ ๋ฐ์ดํ„ฐ์™€ ๋น„๊ตํ•˜์—ฌ ๋ชจ๋ธ์„ ํ‰๊ฐ€ํ•˜์˜€๋‹ค. ๊ณผ๊ฑฐ 20์ผ ๋™์•ˆ ํ•™์Šตํ•œ ๋ฐ์ดํ„ฐ๋ฅผ ํ†ตํ•ด ๋‹ค์Œ๋‚ ์˜ ์ „๊ธฐ์—๋„ˆ์ง€ ์‚ฌ์šฉ๋Ÿ‰์„ ์˜ˆ์ธกํ•˜๋Š” ๋ฐฉ๋ฒ•์œผ๋กœ 14์ผ ๋™์•ˆ ์ง„ํ–‰๋˜์—ˆ์œผ๋ฉฐ, ๋ณธ ์—ฐ๊ตฌ์—์„œ ์ œ์•ˆ๋œ LSTM ๋ชจ๋ธ์€ ๋ชจ๋‘ MBE๊ฐ€ 0%์— ๊ฐ€๊นŒ์šด ๋†’์€ ํ•™์Šต์„ฑ๋Šฅ์„ ๋ณด์˜€๋‹ค. ์˜ˆ์ธก์„ฑ๋Šฅ์—์„œ๋Š” ๊ฐ๊ฐ CVRMSE 23%(๋‹ค์ธต LSTM), 26%(๋‹จ์ธต LSTM) ์ˆ˜์ค€์„ ๋ณด์˜€๋‹ค. ์ƒ๋Œ€์˜ค์ฐจ๊ฐ€ 60% ์ด์ƒ ๋ฐœ์ƒํ•˜๋Š” ํšŒ์ˆ˜๋„ ๋‹จ์ธต LSTM์ด 40์‹œ๊ฐ„, ๋‹ค์ธต LSTM ๋ชจ๋ธ์ด 28์‹œ๊ฐ„์œผ๋กœ ์‹ฌ์ธต LSTM ๋ชจ๋ธ์ด ์˜ˆ์ธก๋ชจ๋ธ๋กœ ๋” ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์˜€๋‹ค. ๋Œ€๋ถ€๋ถ„์˜ ์˜ค์ฐจ๋Š” 20% ์ดํ•˜์—์„œ ๋ฐœ์ƒํ•˜์—ฌ ์šด์ „๊ณ„ํš ์ˆ˜๋ฆฝ์„ ์œ„ํ•œ ์ˆ˜์š” ์˜ˆ์ธก๋ชจ๋ธ๋กœ์จ ์˜๋ฏธ ์žˆ๋Š” ๊ฒฐ๊ณผ๋ฅผ ์ œ๊ณตํ•˜์˜€๋‹ค. ๋‹ค๋งŒ ๋ณธ ์—ฐ๊ตฌ์˜ ๊ฒฐ๊ณผ๋Š” ์žฌ์‹ค์ •๋ณด ๊ธฐ๋ฐ˜ ์ „๊ธฐ ์—๋„ˆ์ง€ ์‚ฌ์šฉ๋Ÿ‰ ์˜ˆ์ธก ๋ชจ๋ธ์„ ๊ฐœ๋ฐœํ•˜๋Š” ๊ธฐ์ดˆ๋‹จ๊ณ„ ์—ฐ๊ตฌ๋กœ ์‚ฌ์šฉ๋œ ์žฌ์‹ค ์ •๋ณด๊ฐ€ ์ „๊ธฐ ์‚ฌ์šฉ๋Ÿ‰์„ ๊ธฐ์ค€์œผ๋กœ ์ถ”์ •๋˜์—ˆ๋Š”๋ฐ, ์‹ค์ฆ์„ ๊ณ ๋ คํ•œ ํ–ฅํ›„ ์—ฐ๊ตฌ์—์„œ๋Š” ์„ผ์„œ ๋ฐ ์‚ฌ์šฉ์ž ๊ฐœ์ž…์„ ํ†ตํ•ด ์ธก์ •๋ฐ์ดํ„ฐ์˜ ์ •ํ™•์„ฑ์ด ํ™•๋ณด๋˜์–ด์•ผ ํ•œ๋‹ค.

๋ณธ ์—ฐ๊ตฌ์˜ ๊ฒฐ๊ณผ๋Š” ์‹ค์ œ ์ธก์ •๋œ ์ „๊ธฐ ์‚ฌ์šฉ ๋ฐ์ดํ„ฐ๋ฅผ ํ™œ์šฉํ•ด ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๋Š” ์ˆ˜์ค€์˜ ๋‹ค์Œ๋‚  ์ „๊ธฐ์—๋„ˆ์ง€ ์ˆ˜์š”๋ฅผ ์˜ˆ์ธก์„ ํ•  ์ˆ˜ ์žˆ์—ˆ๋‹ค. ํ–ฅํ›„ ๋‹ค์–‘ํ•œ ์ข…๋ฅ˜์˜ ๋ฐ์ดํ„ฐ ์ˆ˜์ง‘์ด ์–ด๋ ค์šด ์ฃผ๊ฑฐ๊ฑด๋ฌผ ๋‹จ์ผ์„ธ๋Œ€์˜ ์‹œ์Šคํ…œ ์šด์ „๊ณ„ํš์šฉ์œผ๋กœ ์ด์šฉ๋  ์ˆ˜ ์žˆ์„ ๊ฑฐ๋ผ ๊ธฐ๋Œ€ํ•œ๋‹ค.

ํ›„ ๊ธฐ

๋ณธ ์—ฐ๊ตฌ๋Š” ํ•œ๊ตญ์—๋„ˆ์ง€๊ธฐ์ˆ ์—ฐ๊ตฌ์›์˜ ์ฃผ์š”์‚ฌ์—…์œผ๋กœ ์ˆ˜ํ–‰ํ•œ ๊ฒฐ๊ณผ์ž…๋‹ˆ๋‹ค(B8-2424-01).

References

1 
He W., 2017, Load forecasting via deep neural networks, Procedia Computer Science, Vol. 122, pp. 308-314DOI
2 
Weron R., 2014, Electricity price forecasting : A review of the state-of-the-art with a look into the future, International Journal of Forecasting, Vol. 30, No. 4, pp. 1030-1081DOI
3 
Martinez-Anido C. B., Brinkman G., Hodge B. M., 2016, The impact of wind power on electricity prices, Renewable Energy, Vol. 94, pp. 474-487Google Search
4 
Baldick R., 2012, Wind and energy markets : A case study of Texas, IEEE Systems Journal, Vol. 6, No. 1, pp. 27-34DOI
5 
Pipattanasomporn M., Kuzlu M., Rahman S., 2012, An algorithm for intelligent home energy management and demand response analysis, IEEE Transactions on Smart Grid, Vol. 3, No. 4, pp. 2166-2173DOI
6 
Hyun S. H., Park C. S., Augenbroe G. L. M., 2008, Analysis of uncertainty in natural ventilation predictions of high-rise apartment buildings, Building Services Engineering Research and Technology, Vol. 29, No. 4, pp. 311-326DOI
7 
Wouters P., Heijmans N., Loncour X., 2004, Outline for a general framework for the assessment of innovative ventilation systems, RESHYVENT working report nWP4 D, Vol. 4Google Search
8 
Lago J., De Ridder F., De Schutter B., 2018, Forecasting spot electricity prices : Deep learning approaches and empirical comparison of traditional algorithms, Applied Energy, Vol. 221, pp. 386-405DOI
9 
Roldn-Blay C., Escriv-Escriv G., lvarez-Bel C., Roldn-Porta C., Rodrguez-Garca J., 2013, Upgrade of an artificial neural network prediction method for electrical consumption forecasting using an hourly temperature curve model, Energy and Buildings, Vol. 60, pp. 38-46DOI
10 
Jetcheva J. G., Majidpour M., Chen W. P., 2014, Neural network model ensembles for building-level electricity load forecasts, Energy and Buildings, Vol. 84, pp. 214-223DOI
11 
Hippert H. S., Pedreira C. E., Souza R. C., 2001, Neural networks for short-term load forecasting : A review and evaluation, IEEE Transactions on power systems, Vol. 16, No. 1, pp. 44-55DOI
12 
Kim M., Hong C., 2016, The Artificial Neural Network based Electric Power Demand Forecast using a Season and Weather Informations, Journal of the Institute of Electronics and Information Engineers, Vol. 53, No. 1, pp. 71-78DOI
13 
Din G. M. U., Marnerides A. K., 2017, Short term power load forecasting using deep neural networks, In Computing, Networking and Communications (ICNC), 2017 International Conference on IEEE, pp. 594-598DOI
14 
Marino D. L., Amarasinghe K., Manic M., 2016, Building energy load forecasting using deep neural networks, In Industrial Electronics Society, IECON 2016-42nd Annual Conference of the IEEE, pp. 7046-7051DOI
15 
Hosein S., Hosein P., 2017, Load forecasting using deep neural networks, In Power & Energy Society Innovative Smart Grid Technologies Conference (ISGT), 2017 IEEE, pp. 1-5DOI
16 
Asare-Bediako B., 2014, SMART Energy Homes and the Smart Grid, University of Technology, EindhovenGoogle Search
17 
Kim J. H., Park S. L., Kim D. W., 2011, Agent-based Building Energy Simulation, Architectural Institute of Korea, Vol. 27, No. 12, pp. 315-324Google Search
18 
Fujii H., Tanimoto J., 2004, Integration of building simulation and agent simulation for exploration to environmentally symbiotic architecture, Building and Environment, Vol. 39, No. 8, pp. 885-893DOI
19 
Kashif A., Le X. H. B., Dugdale J., Ploix S., 2011, Agent based Framework to Simulate Inhabitantsโ€™ Behaviour in Domestic Settings for Energy Management, In ICAART, Vol. 2, pp. 190-199DOI
20 
Bonabeau E., 2002, Agent-based modeling: Methods and techniques for simulating human systems, Proceedings of the National Academy of Sciences, Vol. 99, pp. 7280-7287DOI
21 
Lee L. J., 2017, A study on fundamental and application of CNN and RNN, Broadcasting and Media Magazine, Vol. 22, No. 1, pp. 87-95Google Search
22 
Saito G., ๋…„๋„, Deep learning from scratch, Hanbit Media Inc.Google Search
23 
Baccouche M., Mamalet F., Wolf C., Garcia C., Baskurt A., 2011, Sequential deep learning for human action recognition, In International Workshop on Human Behavior Understanding, Springer, Berlin, Heidelberg, pp. 29-39DOI
24 
Kinga D., Adam J. B., 2015, A method for stochastic optimization, In International Conference on Learning Representations (ICLR), Vol. 5Google Search
25 
Kong D. S., Kwak Y. H., Huh J. H., 2010, Artificial Neural Network Based Energy Demand Prediction for the Urban District Energy Planning, Journal of the Architectural Institute of Korea, Vol. 26, No. 2, pp. 221-230Google Search
26 
Documentation M., 2005, The MathWorks Inc.
27 
, https://www.enertalk.com