Enhancing Long-Term Load Forecasting with NLinear Model via Post-Hoc Calibration

This study presents a long-term load forecasting method that integrates the NLinear model with post-hoc calibration. Four calibration techniques—scale, shift, affine, and hour-of-day bias—are applied to correct bias and scale errors without retraining. Using Gangeo feeder load data (2015–2019), results show that scale calibration improved NLinear performance: MAE decreased from 0.5922 to 0.5687 (−4.0%), RMSE from 0.8166 to 0.7800 (−4.5%), and MAPE from 18.79% to 17.72% (−5.7%). Compared to LSTM (MAE 0.7507, MAPE 26.46%) and Autoformer (MAE 0.9712, MAPE 33.86%), the NLinear with scale calibration achieved superior accuracy. Visualization and hourly error analysis confirmed reduced bias and variability, with error medians closer to zero and narrower variance. These results demonstrate that post-hoc calibration enhances the accuracy and reliability of NLinear-based long-term load forecasting.