DC power flow PTDF relief: |flow| reduction per MW of load shed is -sign(flow)*H[l,b], not +sign*H
When choosing which load to shed (or import to cut) to relieve an overloaded line with a PTDF matrix H (flow = H @ P, P = injections, load = negative injection): shedding 1 MW at bus b is +1 MW injection, so flow_l changes by +H[l,b]. The reduction in |flow_l| is therefore -sign(flow_l)*H[l,b]. Cutting an import/generator at b by 1 MW is -1 MW injection, so the reduction is +sign(flow_l)*H[l,b]. Getting the load sign backwards silently rejects every useful bus only on lines whose flow is negative in the from->to orientation. In our case a radial feeder (H=1.0 to its downstream substations) looked unfixable and the planner fell back to least-bad plans. Sanity check: on a radial line every downstream load should show positive relief. Related design lesson for wildfire/PSPS planners on a real grid: if every line above a fire-risk threshold is a hard violation and plans are capped at a few actions, almost no Santa Ana hour has a feasible plan, because dozens of lines exceed it and most are 220-500 kV import corridors. Utility practice (e.g. SCE Jan 2025 PSPS: 373 distribution circuits) is to de-energize sub-transmission/distribution and patrol the bulk corridors. Encoding that took feasible hours from a handful to 100 of 102.