fix: level the grid profile below the peak, not only the absolute peak

[andig]

Problem

The attenuate_demand_peaks / attenuate_feedin_peaks / attenuate_grid_peaks strategies penalize only p_{side}_peak, the horizon maximum of the grid power. That caps the absolute peak but says nothing about the profile underneath it.

Once the maximum is pinned by a time step the batteries cannot shave — a battery already at s_max, or a surplus far beyond c_max — the objective becomes completely flat over every profile below the cap, so the solver returns an arbitrary vertex. On the reported 192-step case the feed-in tail came out as

172, 0, 429, 525, 0, 0, 0, 943, 0, 0, 1179, 0, 1252, 0, 1415, 0

Peak capped, curve not leveled.

Change

Penalize the step to step ramp of the grid power in addition to the peak:

p_{side}_ramp[t] >= |p_grid[t] - p_grid[t-1]|

with prc_p_ramp = penalty_base * 1e-5, a hundred times below prc_p_peak = penalty_base * 1e-3, so lowering the peak still wins wherever the two disagree. The peak constraint is rewritten in power terms (p_grid[t] <= p_{side}_peak) so both constraints share the same expression, including the portion beyond p_max_imp / p_max_exp.

Effect on the reported case

	peak	total variation	economics
before	1853 W	20666	6.686112
after	1806 W	8899	6.686112

Cost neutral: identical import cost, export revenue and stored value. The tail is flat 402, 402, 402, … instead of the sawtooth, and the morning ramp climbs monotonically to the cap.

Solving also gets ~20x faster on this case (12.7 s → 0.5 s): the ramp term removes the degeneracy CBC was branching through.

Two small dips remain (t=21, t=105) at ~5e-6 objective difference, i.e. below CBC's termination tolerance rather than a formulation gap.

Tests

Coverage lives in the data driven harness on #101: test_cases/026-attenuate-feedin-peaks-below-cap.json. It pins the export peak in step 0 beyond reach of c_max, so the remaining charge energy can be placed anywhere in the following steps without touching the peak, and the strict schedule comparison pins the leveled result.

• without this change: grid_export [950, 50, 50, 50, 80]
• with it: grid_export [950, 57.5, 57.5, 57.5, 57.5]

That case therefore fails until this PR lands. uv run pytest on this branch → 15 passed, ruff check clean; with #101 merged in as well → 19 passed.

🤖 Generated with Claude Code