Measurement Uncertainty and Deep Learning: A Metrology-Aware Framework for Load Forecasting

Accurate load forecasting is essential for power system operation, yet most deep learning (DL) pipelines neglect the measurement uncertainty affecting sensor-acquired input data. This article proposes a metrology-aware, data-centric, and architecture-agnostic framework that incorporates instrumental uncertainty into DL workflows without modifying the predictive model. Input uncertainty is modeled during training through guide to the expression of uncertainty in measurement (GUM)-consistent data augmentation (DA) and during inference through repeated perturbation of the test samples. The framework is validated on public Ausgrid substation data using a long short-term memory (LSTM) network for one-day-ahead and one-week-ahead forecasting. Representative results show that uncertainty-aware DA reduces the one-week-ahead mean absolute percentage error (MAPE) from 0.12 to 0.09 in the single-substation forecasting scenario and the one-week-ahead mean absolute error (MAE) from 1.80 to 1.59 MW in the cross-substation generalization scenario, corresponding to relative reductions of approximately 25% and 12%, respectively. The analysis also considers robustness under perturbed inputs, cross-substation generalization, computational cost, statistical significance, and comparisons with model-centric baselines based on deep ensembles and Monte Carlo Dropout. Repeated uncertainty-aware testing shows very small variability under realistic input perturbations, with standard deviations lower than 10−3 MW for MAE and root mean square error (RMSE) and lower than 10−4 for MAPE. The corresponding empirical interval widths remain limited, with mean prediction interval width (MPIW) values below 0.06 MW in all the considered cases. Conversely, model-centric baselines exhibit larger variability, up to 0.18 MW for MAE, 0.22 MW for RMSE, and 0.02 for MAPE in the ensemble configuration, confirming that model-related uncertainty dominates measurement-induced uncertainty in the considered case study. Overall, the proposed framework bridges metrological principles and DL-based forecasting, enabling more robust and traceable predictions for measurement-intensive power applications.