Weather-Regime-Conditioned Evaluation of Photovoltaic Power Forecasts Across Two Independent Datasets
Year: 2026
Type: article
Abstract: Photovoltaic (PV) power forecasts are typically evaluated using aggregate error metrics computed over an entire test period, which may conceal condition-dependent accuracy differences. We evaluate whether forecasting error varies systematically with weather regime, ramp events, and regime transitions, applying an evaluation layer to existing, frozen forecast outputs from two independent PV datasets (SOLETE, Denmark; Alice Springs (DKASC), Australia) across 50 prediction cells and five forecast horizons (5–180 min). Regimes, ramps, and transitions were defined from training-period-only thresholds, with forecast-time context kept separate from retrospective conditioning; differences were assessed using episode-level Mann–Whitney U tests on contiguous condition episodes, with Holm–Bonferroni correction and bootstrap confidence intervals. Significance counts are conditional on the episode definition, and residual dependence between episodes was not modeled. Transition-conditioned differences were statistically significant in 45 of 50 cells, with higher error inside transition windows in 43 of them, and forecast-time ramp context in 18 of 50, almost all with higher error under ramps. Forecast-origin regime differences were significant in most cells (46 and 48 of 50 for variable and highly variable versus stable conditions), but nearly all of these showed lower, not higher, episode-level error under the variable regimes. At the sample level, by contrast, simpler models on Alice Springs (DKASC) showed substantially higher error under variable conditions (relative mean absolute error (MAE) increases up to about 140% across horizons, and up to 227% in individual cells), while deep-learning models showed smaller, horizon-dependent effects; this contrast reflects rank-based episode-level versus mean-dominated aggregate comparison of a right-skewed error distribution. Under cross-dataset transfer, R3-versus-R1 degradation reversed direction relative to the in-domain CNN-BiLSTM pattern (R2-versus-R1 changes were already negative in-domain and differed only in magnitude), an open question that this study does not resolve. This study contributes a systematic, statistically supported evaluation protocol for weather-regime-, ramp-, and transition-conditioned PV forecasting accuracy, showing that aggregate metrics can obscure forecasting behavior relevant to operational reliability, and that the direction of a condition-dependent effect can depend on whether it is measured at the episode or the aggregate level.
Source: Energies
DOI: 10.3390/en19225404
Language: English
Field: Energy
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $3,137
OpenAlex ID: W7225756530
FWCI:
Cites:
Cited by:
Related to: