2024-08-06 - Updates

Modelling

• Adaptive downsampling with RDP
  • Single tunable parameter $\epsilon$
  • Keep between 2% and 10% of original points
  • Needs explicit time features in inputs; either relative ($\Delta t$) or absolute, which need to be normalized
  • Reduces number effective number of samples processed, while retaining accuracy where there are fast changes.
  • Sensitive to noise. Noise needs to be removed
    • Removed using lowpass filter and low amplitude filter on MBI data
• Curriculum Learning with pJiles-Atherton model for hysteresis simulated data on TFT
  • In progress, stages need to be defined
  • For now, 24h simulated JA data
  • Transfer learning to 1h MBI data, validated on SFTPRO+LHC+ZERO
    • Results?
• Neural ODEs
  • Parametrize a function with a neural network and solve with a numerical integrator with the help of autograd, let NN find a the solution to the hysteresis ODE.
  • Needs explicit (and normalized time features)
  • Implemented in JAX
  • Questionable performance, but useful for PINN course project
• SA-PINN
  • Self-adaptive loss weights for Bouc-Wen LSTM
  • Works well, difficult to tune.
  • Optimizes better than vanilla BWLSTM