Patch-Based Transformer With Mixture of Experts and Conformal Inference for Curtailment Time Series Forecasting

The growth of variable renewable energy sources has increased the occurrence of curtailment in power systems, particularly in regions affected by transmission constraints and limited operational flexibility. Reliable forecasting of curtailment time series, together with well-calibrated uncertainty estimates, is therefore essential to support operational planning, market decisions, and risk management. This paper proposes a curtailment forecasting framework that combines a high-capacity neural time series model with distribution-free conformal prediction to jointly address accuracy and uncertainty quantification. The forecasting backbone is a Transformer-based encoder-decoder architecture enhanced with reversible instance normalization to mitigate distribution shift, patch-based tokenization to reduce sequence length and improve temporal representation, and a Mixture of Experts feed-forward module to enable conditional computation and improved generalization. The proposed method is evaluated against a wide range of state-of-the-art models, including NBEATS, NBEATSx, NHITS, PatchTST, LSTM, GRU, TFT, Informer, and Autoformer, across multiple forecasting horizons. The model achieves an MSE of 0.0059 for a horizon equal to 3h, corresponding to a 13.2% reduction relative to the second-best methods. For a horizon equal to 5h, the MSE is reduced to 0.0119, representing a 29.6% improvement, while for a horizon equal to 10, the model attains an MSE of 0.0461, outperforming NHITS by 22.9%. Ablation analysis confirms the importance of patch embedding, the Mixture of Experts mechanism, and reversible instance normalization. The conformal prediction module achieves an empirical coverage of 91.27% for a nominal 90% level, with adaptive and practically useful interval widths.