Project, 2024.05–2024.12

Development of simulator-based torque pattern prediction model and speed ripple reduction logic

This LG Electronics industry-academia project (May to Dec 2024) addressed a trade-off in dryers: raising the compressor operating frequency improves energy efficiency, but at high speed the speed ripple from feedback-based torque control increases noise and vibration. Using data collected from an Ansys TwinBuilder dryer simulator, the team built a Stacked Bi-LSTM model that predicts the next cycle's torque pattern, then developed and compared reinforcement learning controllers (DDPG, TD3, LSTM-TD3) that turn the predicted torque into a compensation current. Joonho Seo was responsible for designing and developing the torque pattern prediction model and running its experiments. The final report lists torque prediction MAPE of 1.42% for TORQUE and 0.46% for OMEGA (target: under 10%), a 24.0% reduction in speed ripple (MASD) with LSTM-TD3, and two software deliverables (a torque pattern prediction system and a speed ripple reduction system).

Where

LG Electronics

My role

Torque prediction model development

Status

completed

Torque prediction MAPE: TORQUE 1.42%, OMEGA 0.46%; 24.0% speed ripple reduction with LSTM-TD3; 2 software deliverables