From Routing to Weighting: Mask-Only Policies for Repairing Incomplete Photovoltaic Monitoring Records

Repairing an incomplete photovoltaic monitoring record requires choosing an imputation operator, and the right choice depends on which evidence the fault left intact. The decision can use only what a monitoring system holds at repair time. We study mask-only policies, which read the observation mask (which cells are missing) and the clock but never the values being repaired, and ask how fine such a decision must be. Two policies are compared: a fault-aware router that diagnoses each missing cell into one of three regimes and dispatches it to one operator, and MAWC, a mask-aware weighted combination that predicts a convex weight over the whole operator pool per cell. Both are fixed on two desert development sites and evaluated once on four held-out sites spanning two further climates and cadences from 1 to 15 min, with mixed fault masks and time-contiguous evaluation segments, two protocol corrections without which a per-cell policy cannot be measured. The router improves on the operator that a deployer could have committed to at every held-out site and lies within 7% of the best fixed assignment of operators to its regimes at three of them. MAWC improves on the router at all four sites, significantly at three, improves on per-regime convex weights at all four, and falls below the regime oracle at two, which no hard router using those regimes can do. Its gain comes from hedging rather than finer selection: a selector trained to pick one operator per cell with the same features loses 15% to 34%. The advantage holds under retrained deep members, doubled replication, changed fault composition and a higher missing rate.