Day-ahead industrial peak demand forecasting: A shipyard benchmark of deep learning forecasters and stacking ensembles

• Multi-source data improves daily industrial peak-demand forecasting for a shipyard. • Twelve ML/DL models are benchmarked under a unified train-test protocol. • N-HiTS is the best single model, outperforming Transformer and boosting baselines. • Stacking with an LSTM meta -learner reduces MAE by 15.7% vs. N-HiTS. • Paired t-tests confirm the ensemble improvement is significant (p < 0.05). Large industrial customers often face demand charges or ratchet mechanisms that make a single high peak day materially influence annual electricity costs. Shipyards are particularly challenging because their daily peak demand is driven by heterogeneous production processes, strong calendar effects, and weather-dependent auxiliary loads. This paper studies daily peak power demand forecasting for a large, anonymized shipyard using a multi-source operational dataset (April 2016–December 2023) that integrates calendar variables, process indicators, and weather covariates (2,018 daily observations; 20 variables). The full chronological split contains 1,758 pre-2023 days and 260 days in 2023; the working-day benchmark used for model comparison contains 1,665 training days and 241 test days. We benchmark 12 forecasting models spanning RNNs, Transformer-based forecasters, MLP-based time-series models, tree-boosting methods, and a Koopman-predictor model. Among single models, N-HiTS achieves the best accuracy on the 2023 working-day test set (MAE 1,474.8 kW, RMSE 1,847.4 kW, and R 2 0.834). To improve upon this baseline, we build stacking ensembles from archived N-HiTS, Autoformer, and TFT forecasts and evaluate four meta learners (linear regression, MLP, LSTM, and XGBoost). XGBoost meta learning achieves the lowest overall MAE (1,241.4 kW), while the LSTM meta learner attains the best RMSE (1,664.8 kW), the highest R 2 (0.865), and the strongest recall on top-decile peak-risk days. Raw-kW paired t-tests confirm significant gains for the linear, XGBoost, and LSTM meta learners relative to the N-HiTS baseline. We further add SHAP-based interpretability for the tree benchmark, rolling residual-based prediction intervals, and a demand-charge sensitivity illustration for the annual maximum day. These results show that benchmark accuracy, interpretability, uncertainty, and operational usefulness should be evaluated jointly when deploying industrial peak-demand forecasting systems.