With the rise of Industry 4.0 and ESG management, energy efficiency is vital for the shipbuilding industry. In Korea, industrial electricity costs heavily depend on "Peak Power" demand, where the highest daily peak recorded during specific seasons det...
With the rise of Industry 4.0 and ESG management, energy efficiency is vital for the shipbuilding industry. In Korea, industrial electricity costs heavily depend on "Peak Power" demand, where the highest daily peak recorded during specific seasons determines the base rate for the entire year. Consequently, accurate daily forecasting is crucial for cost reduction, yet existing empirical methods fail to capture the non-linear fluctuations of shipyard power consumption.
This study identifies the optimal forecasting model and proposes an ensemble approach using daily shipyard operating data from 2016 to 2023. The dataset integrates power consumption, production processes, and meteorological factors. Twelve models across five categories—RNN, Transformer, MLP, Gradient Boosting, and Koopa—were evaluated using MAE, RMSE, and R2 scores.
Experimental results revealed that the N-HiTS (Neural Hierarchical Interpolation for Time Series) model achieved superior performance among single models (R2 0.8775, MAE 1,173kW). N-HiTS successfully captured multi-scale seasonality through hierarchical interpolation, outperforming complex Transformer-based models like TFT (R2 0.765) and traditional RNNs.
To enhance stability, ensemble models were constructed using N-HiTS as the base learner. Statistical verification (t-test) indicated that the LSTM Meta Learner provided the most statistically significant improvement (p-value < 0.05), achieving an MAE of 1,243kW and R2 of 0.8653. This research validates the N-HiTS architecture and the LSTM ensemble as robust solutions for precise electricity cost management in smart shipyards.