Informative Missingness in Grid Load Forecasting: Rethinking Evaluation and Missingness-Aware Forecasting

In electricity distribution grids, observations are not lost at random, since feeder outages and spatially coordinated disturbances can hide measurements precisely when the underlying system is under stress, increasingly driven by climate-related events. We study short-term load forecasting under informative missingness using a controlled benchmark in which four missingness mechanisms, missing completely at random, missing at random, node-level missing not at random, and grid-level missing not at random, are applied to identical synthetic load series at the same missing rate. We first identify and correct an evaluation pitfall in which scoring forecasts only on visible positions confounds the mechanism with the evaluation set, inverting the intended difficulty ranking. Under the corrected protocol, informative missingness concentrates error on hidden positions rather than uniformly increasing aggregate error, with the observed-hidden gap growing from $0.6\%$ under missing completely at random to $42.5\%$ under grid-level missing not at random. An explicit dual-stream architecture modeling the observation process is significantly outperformed by a forecaster with a single additional mask channel ($p<10^{-4}$ in every comparison). These findings show that aggregate accuracy can obscure informative missingness and motivate evaluating forecasts by where error occurs, not only how much.