Document Type : Research Note
Author
Faculty of Engineering University of Palermo Italy
Abstract
Short-term load forecasting (STLF) studies can report strong accuracy when target construction, data provenance, units, and forecast-time information are not audited. This study presents a methodological audit using 189,888 quarter-hourly records labelled from January 2020 to May 2025. Although complete, the dataset’s provenance cannot be independently verified, and its target follows the near-exact construction: Load = 1500 + 100(Temperature − 28). Load values are therefore reported in source units rather than asserted physical kW.
The diagnostic contribution demonstrates how target reconstruction, chronological splitting, forecast-information checks, interpretable baselines, and explicit claim boundaries alter the interpretation of an apparently high-performing STLF experiment. Ordinary Least Squares is exact because it recovers the construction rule. Random Forest and Gradient Boosting are retained as falsification benchmarks, not evidence of algorithmic competition. Near-zero correlations and feature importance for other variables confirm that the dataset does not represent realistic multivariate or autoregressive demand dynamics.
Because realized temperature appears to be available at prediction time, the nominal metrics represent ex-post conditional fit rather than deployable ex-ante forecasting accuracy. Without fabricating observations, an analytical stress test propagates temperature-forecast error and independent load noise through the recovered equation. Temperature-error standard deviations of 0.5, 1.0, and 2.0 °C imply RMSE values of 50, 100, and 200 source units, respectively, before behavioral noise. The study provides a reproducible audit protocol and negative methodological result: accuracy, uncertainty, renewable-integration relevance, and operational value cannot be inferred from constructed targets without verified measured demand, issue-time weather forecasts, renewable generation, and system constraints.
Keywords
- short-term load forecasting
- constructed dataset
- weather-forecast error
- uncertainty stress test
- model governance
Main Subjects
)