Industrial load forecasting: A novel methodology based on a peak-aware perspective

Industrial facilities increasingly operate as behind-the-meter (BtM) energy hubs combining on-site generation, storage, and flexible loads behind a single metering point, where day-ahead hourly demand forecasts directly feed energy management: battery dispatch, production shifting, demand-charge optimization, and contracted-power compliance. Yet conventional average-error metrics treat all errors equally, even though an error during a production peak is far more consequential than one at night. This work addresses day-ahead hourly load forecasting for an industrial agri-food plant and makes two contributions. First, we propose an operationally aligned evaluation methodology with four metrics: a Peak-weighted Mean Absolute Error (PWMAE), a peak-specific error ( MAE peak ), an off-peak-specific error ( MAE off ), and a Peak Sensitivity Index (PSI), enabling model selection by operational risk exposure rather than average accuracy alone. Second, under a causal, reproducible framework issued daily at 00:00 using only calendar and local weather, we show for this facility that the best model family depends on data granularity and on the operational criterion. With aggregate-only data, Temporal Convolutional Networks (TCNs) achieve stable accuracy and the lowest peak-weighted error. With per-machine submetering the picture simplifies: gradient boosting leads under both criteria simultaneously (LightGBM without lags, selected on validation under either objective, reaching 3.46 kWh MAE and 3.91 kWh peak-weighted error), so the trade-off that characterizes the aggregate-only regime does not arise. Data granularity therefore governs not only which model family to deploy but whether the operational criterion alters the choice at all. In both scenarios, the proposed approach outperforms a representative literature baseline and a commercial solution deployed at the site, especially around production peaks.