How amidouetude/ml-mcp-wt works
This MATLAB research pipeline studies real-time ML-surrogate MPC for wind turbines, diagnoses MATLAB predict() overhead, and supplies manual-forward bypasses plus reproducibility and sensitivity checks. The main workflow runs stages 1–6, builds/trains surrogate models, designs controllers, and evaluates closed-loop behavior. Track B adds residual-learning experiments and alternate plant integration; the repository also documents campaign provenance checks and paper outputs. Relationships below follow README statements; sampled source was limited to the Python digest reference, so finer call-level wiring is not asserted.
Main entry points: sf_gp.m Surrogate functions · stage3_run_simulation.m Closed-loop simulation · manual_forward_generic.m Manual forward passes · run_all.m Stage runner
Components 20
Core pipeline
- Stage runner
run_all.m - Turbine plant
common/wt_step.m - Training data
common/stage1_generate_data.m - V1/V2 training
common/stage2_main.m - Surrogate functions
common/sf_gp.m - Controller design
common/stage3_design_controllers.m - Closed-loop simulation
common/stage3_run_simulation.m
Bypass and residuals
- Unmodeled plant
piste_B_nominal_correction/wt_step_true.m - Residual data
piste_B_nominal_correction/stage1_generate_residual_data.m - Residual training
piste_B_nominal_correction/stage2_train_residual.m - Manual forward passes
piste_B_nominal_correction/manual_forward_generic.m - Residual controllers
piste_B_nominal_correction/stage3_design_controllers_residual.m
Verification
- Domain checks
piste_B_nominal_correction/check_surrogate_domains.m - Sensitivity studies
piste_B_nominal_correction/run_sensitivity_noise_delay.m - Campaign verification
common/verifier_lot_campagne.m - Manifest digest
common/digest_manifeste.m - Digest reference
common/digest_manifeste_reference.py
Research outputs
- Monte Carlo
piste_B_nominal_correction/run_extended_monte_carlo.m - Paper results
paper/main.tex
Other components
- Researcher
Connections 18
- Researcher → Stage runner: runs stages
README.md - Stage runner → Training data: runs pipeline
README.md - Training data → V1/V2 training: trains from
- V1/V2 training → Surrogate functions: produces models
README.md - Surrogate functions → Controller design: supplies models
- Controller design → Closed-loop simulation: evaluates closed loop
- Turbine plant → Closed-loop simulation: simulates turbine
- Unmodeled plant → Residual data: generates residual data
README.md - Residual data → Residual training: trains residual model
README.md - Residual training → Residual controllers: provides model
README.md - Manual forward passes → Surrogate functions: bypasses predict
README.md - Manual forward passes → Residual controllers: supports inference
- Surrogate functions → Domain checks: checks operating domain
README.md - Closed-loop simulation → Sensitivity studies: tests robustness
README.md - Researcher → Campaign verification: verifies campaign
- Manifest digest → Digest reference: cross-checks convention
common/digest_manifeste_reference.py - Closed-loop simulation → Monte Carlo: extends evaluation
README.md - Monte Carlo → Paper results: reports results
README.md