Controller Parameter Sensitivity¶

Research Question¶

How do OFO tuning parameters (step sizes, dual weights, throughput/switching costs) affect voltage regulation performance, and what are the optimal operating points?

Overview¶

The OFO controller has several tuning parameters that trade off between voltage regulation aggressiveness, throughput preservation, and batch-size stability. This analysis sweeps each parameter one-at-a-time (shared) or across sites independently (per-site) to characterize the sensitivity landscape.

The sweep mode auto-selects based on the number of DC sites:

1 DC site (e.g., IEEE 13): shared sweep, which varies each parameter while keeping others at baseline.
2+ DC sites (e.g., IEEE 34): per-site sweep, which sweeps all per-site parameter combinations independently, producing heatmap visualizations.

Scripts¶

Script	Purpose
`sweep_ofo_parameters.py`	Automated parameter sweep with comparison plots

Usage¶

IEEE 13-Bus: Shared Sweep¶

python examples/offline/sweep_ofo_parameters.py \
    --system ieee13

IEEE 34-Bus: Per-Site Sweep¶

# per-site sweep (independent parameters per site, auto-selected)
python examples/offline/sweep_ofo_parameters.py \
    --system ieee34

# Force shared sweep (shared parameters across sites)
python examples/offline/sweep_ofo_parameters.py \
    --system ieee34 --sweep-mode shared

Overriding Time Resolution¶

For faster sweeps during exploration, use a coarser time step:

python examples/offline/sweep_ofo_parameters.py \
    --system ieee34 --dt 60

Parameters Swept¶

Parameter	`OFOConfig` field	Effect
`voltage_dual_step_size`	`voltage_dual_step_size`	Aggressiveness of voltage constraint enforcement. Most impactful: values 5-10 can cut violation time 3x vs baseline (1.0).
`primal_step_size`	`primal_step_size`	Learning rate for batch-size updates. Sweet spot ~0.05; default 0.1 is slightly suboptimal.
`latency_dual_step_size`	`latency_dual_step_size`	Aggressiveness of latency constraint enforcement. Nearly no effect under standard load (latency dual not binding).
`w_throughput`	`w_throughput`	Weight on throughput preservation in objective. Values > 0.01 cause severe degradation; keep <= 1e-3.
`w_switch`	`w_switch`	Regularization for batch-size changes. Value of 5.0 gives mild improvement over baseline 1.0.

Outputs¶

Saved to outputs/<system>/sweep_ofo_parameters/:

results_<system>_sweep_ofo_parameters.csv: one row per run with voltage metrics (violation time, integral, worst vmin/vmax), per-site power, and per-model throughput (avg_throughput_tokens_per_s__<model>) and ITL-miss fraction (itl_violation_frac__<model>)
Per-parameter subdirectories with voltage and batch-size plots
Heatmaps (per-site mode) showing cross-site parameter interactions

Configuration¶

Baseline OFO parameters: OFOConfig(...), the center of the sweep grid, defined inline in each script's experiment function
Simulation length: TOTAL_DURATION_S constant in each script (default 3600s)

See Building Simulators for the full component API.