Non-Forecasting Self-Supervised Masked-Reconstruction Framework

This study investigates self-supervised reconstruction of hidden values within an observed window for solar and wind energy time series, without forecasting beyond the window. The DualAxis architecture sequentially models temporal and cross-feature dependencies in 24 h windows, providing point reconstructions and a 96-dimensional representation. A chronological protocol separates training, validation, and retrospective test partitions without overlapping source rows and uses shared evaluation masks. DualAxis is compared with a Transformer, a gated recurrent unit (GRU) autoencoder, SAITS, and classical imputers. On the main test set with 30% random masking, the root mean squared error (RMSE) over ten signals, averaged over three training runs, is 0.0908 for DualAxis and 0.1031 for SAITS, whereas linear interpolation reaches 0.0774. Two independent datasets, rolling-origin validation, and structured missingness show that method rankings depend on evaluation conditions. A composite quality-control (QC) score combining reconstruction error and embedding outlierness changes substantially when the density component is removed or the model is retrained. Controlled corruptions do not confirm reliable automatic fault detection. Monte Carlo dropout uncertainty correlates moderately with error for both generation channels, but the intervals do not reach nominal coverage. The results delineate the capabilities and substantial limitations of multi-task non-forecasting analysis of renewable energy data.