- Article
An energy manager may receive an issued photovoltaic forecast without plant metadata or forecast weather. This study evaluates a global residual correction of the complete issued trajectory using EMSx vendor forecasts and measurements from anonymized industrial systems. With calendar cutoffs common to all systems, residual LightGBM reduces the vendor mean absolute error on 1,392,864 test deliveries from 47 systems from 100.57 to 83.26 kWh, or 17.2%, and improves every system. An ablation attributes most of this gain to the shape of the issued profile. A model fitted without the evaluated system, without system identity, and with a scale derived from issued forecasts reduces the error by 18.8%, so the correction does not require a training history of the plant. Linear models for each delivery step and compact Transformer and BiLSTM residual models give smaller reductions. Without task training, Chronos-2 supplied with the issued trajectory reaches a slightly lower error than LightGBM on a shared subset of the test, whereas without the trajectory it is worse than the vendor. A partition that is chronological only within each system overstates the gain of such a global model: on identical deliveries issued before the end of its fitting data, the error falls by 53.4%, compared with 12.6% after the common cutoff. Rolling calibration of residual quantiles lowers the weighted interval score from 66.80 to 64.44 kWh and raises central 90% coverage from 0.826 to 0.915, with wider intervals.
13 November 2026
