Predictive cold-ironing infrastructure planning for nearly zero-energy ports: A NARX digital-shadow forecasting framework

Onshore power supply (OPS), also termed cold ironing (CI), is central to port decarbonisation, yet shore-side infrastructure is still dimensioned by deterministic, peak-based rules that ignore the temporal dynamics of vessel traffic. This study couples a modular cold-ironing virtual model with a machine-learning Digital Shadow and validates it at the Port of Ancona using a full year of hourly vessel-traffic data (8784 hours) and 55 exogenous features spanning meteorological, temporal and vessel-traffic variables. The reconstructed load is structurally hostile to conventional planning: an annual demand of 52.2 GWh, a peak-to-average ratio of 4.42, a load factor of 22.6%, and an absolute peak of 26.3 MW reached in fewer than 5% of operating hours. Three learning paradigms are evaluated within a thirteen-model ensemble under two architectures: a static configuration using only exogenous inputs, and a nonlinear autoregressive configuration with exogenous inputs (NARX). The NARX architecture raises the best one-step coefficient of determination from 0.562 to 0.851 and reduces RMSE by 42%, and the multilayer perceptron is the most accurate model, exceeding SARIMA, LSTM and TCN baselines. An ablation attributes most of the skill to autoregressive memory (+0.289 R2) and a smaller but genuine gain to exogenous variables (+0.032). One-step accuracy degrades under fully recursive operation (24-hour R2 of 0.61), which bounds autonomous day-ahead use. Sizing infrastructure at the 95th demand percentile (14.5 MW) rather than the 26.3 MW peak reduces required capacity by 44.9%, with residual exceedance managed through forecast-informed connection scheduling. The framework relies on publicly accessible data and is designed for transfer to other ports after retraining.