Vessel-Type Electrical Load Forecasting for Cold Ironing Berths: A Data-Driven Approach

1. Introduction

Cold ironing has become a practical pathway for reducing emissions from ships during berthing by allowing vessels to connect to shore-side electricity instead of relying on auxiliary engines [1]. Adding to that, EU regulation requires ships at berth to connect to onshore power for their electrical needs instead of using their own engines, with the objective of reducing greenhouse gas emissions and noise pollution in ports [2]. It is therefore obvious that its successful implementation depends not only on the availability of shore power infrastructure, but also on the ability to estimate the electrical demand created by different vessel types, port calls, and berthing activities.

1.1. Existing Literature

Following an extensive review of the literature, several studies were identified that employ data-driven, bottom-up, and forecasting methodologies to estimate ship power requirements. These works demonstrate the relevance of vessel characteristics, operational profiles, and maritime data in supporting energy planning for port electrification [3] [4] [5].

In [6], Probabilistic bottom-up models have been used to create synthetic electrical load profiles for cruise ship cabins by combining passenger behaviour, key-card data, electrical consumer characteristics, route information, and climate effects on air-conditioning demand. The functionality of this model directly follows the definition of this research where power demand can be inferred from structured maritime datasets.

Similarly, other bottom-up approaches, such as the MariTEAM model [7], estimate ship resistance and power consumption using operational profiles, weather conditions, and fleet-level ship data, demonstrating the value of well-designed modelling for energy and emissions assessment. These studies show that synthetic or estimated load information can support decision-making where measured data are incomplete.

Data-driven techniques have also been applied in maritime energy studies to capture nonlinear relationships between ship operating conditions and power-related outputs. For example, authors in [8] employed machine learning models such as Deep Neural Networks to predict generator load in a shipboard microgrid, using fuel propulsion demand. Another study [9] implemented machine learning techniques to extract representative templates for ship-propulsion. This highlights that data-driven applications can prove to be a robust framework for load demand forecasting.

Accurate berthing duration is also important for improving cold ironing planning and management. Although this research does not directly address load forecasting [10], it remains beneficial for the needs of this study, as it demonstrates the use of forecasting techniques to provide valuable insights into ship berthing schedules, which are closely related to the operational requirements of shore-side electricity systems.

Similar to previous studies employing data-driven approaches to estimate vessel load demand where direct measurements are unavailable, this research utilises vessel characteristics and route information to predict power requirements. While existing approaches provide valuable insights into individual vessel energy demand, the proposed method extends this capability by estimating the accumulated power demand of vessels intending to berth at a port. This enables port operators to anticipate demand fluctuations, optimise shore power allocation, reduce peak grid stress, support renewable energy and storage integration, and inform strategic infrastructure planning.

2. Methodology

2.1. Problem Definition

This study develops a methodology that responds to the volatile nature of energy management in ports during cold ironing planning. Onshore power supply allows berthed ships to switch off their auxiliary engines and receive electricity from the port network; however, this transfers a highly variable electrical load from the vessel to the shore-side energy system. This variability is influenced by uncertain berthing duration, changes in arrival and departure schedules, loading and unloading activities, and congestion caused by overlapping ship calls. The challenge is further increased by the limited availability of measured ship-level consumption data, as ports often have access only to general vessel information, arrival records, schedules, and berth allocation data rather than detailed onboard power measurements. Therefore, forecasting ship power and energy demand provides a necessary decision-support tool for estimating the expected electrical load of individual vessels and the accumulated demand at port level, enabling more reliable cold ironing planning, improved energy management, and more efficient port operation.

2.2. Data Acquisition

The effectiveness of a data-driven forecasting approach is largely dependent on the availability of well-structured and representative datasets. Consequently, accurate acquisition of the input dataset will directly affect completeness of the study, hence, to ensure reliability, data collection was divided into two sections. The first section focused on gathering and organising relevant vessel information from available sources, including insights provided by ship operators and data obtained from subscription-based maritime platforms [11] [12]. This stage was essential for identifying key vessel characteristics, such as vessel type, gross tonnage, and available power demand indicators, which formed the basis of the modelling process. The second section used this structured information to generate synthetic power demand data for vessels where direct measurements were unavailable. These synthetic data were then used to support ship load demand forecasting, enabling the estimation of individual vessel power consumption and the aggregated cold ironing demand at port level.

2.3. Data Processing

The gathered information including vessel type, gross tonnage, available power consumption values, and peak power demand indicators, were arranged into a structured database for subsequent modelling. Since the raw dataset contained different vessel classes, inconsistent entries, and varying levels of completeness, a preprocessing stage was required to improve consistency and prepare the data for the learning procedure. A sorting algorithm was therefore applied to classify each vessel according to its type, category, and relevant operational characteristics. This process enabled vessels with similar technical features and expected operational profiles to be grouped together, ensuring consistency in comparison and reducing the computational burden during the estimation process. The resulting structured dataset forms the basis for the synthetic power demand generation process and supports the estimation of individual ship power consumption and aggregated cold ironing demand at port level. Table 1 illustrates Cargo and Tanker ship characteristics provided by port operators showing real power demand during loading and rest operation.

Table 1. Build and Power characteristics of ships served in Limassol Port.

Flag	Length	Max Amps	Power During Loading kW	Power During Rest kW	On-shore connection
PORTUGAL	129.62	400	1000	350	YES
LIBERIA	147.87	400	330	300	YES
MARSHA ISLANDS	147.83	115	660	550	YES
LIBERIA	170.15	125	500	450	YES
CYPRUS	121.75	400	150	120	YES
CYPRUS	133.34	650	250	220	YES
LIBERIA	134.61	320	240	140	YES
LIBERIA	192.92	380	650	500	YES
PANAMA	333.96	380	2000	1700	YES
SINGAPORE	366.107	1200	2000	1500	YES
SINGAPORE	334.98	1250	2000	1200	YES
PANAMA	333.96	380	2000	1700	YES

Table 1. Build and Power characteristics of ships served in Limassol Port.

Flag	Length	Max Amps	Power During Loading kW	Power During Rest kW	On-shore connection
PORTUGAL	129.62	400	1000	350	YES
LIBERIA	147.87	400	330	300	YES
MARSHA ISLANDS	147.83	115	660	550	YES
LIBERIA	170.15	125	500	450	YES
CYPRUS	121.75	400	150	120	YES
CYPRUS	133.34	650	250	220	YES
LIBERIA	134.61	320	240	140	YES
LIBERIA	192.92	380	650	500	YES
PANAMA	333.96	380	2000	1700	YES
SINGAPORE	366.107	1200	2000	1500	YES
SINGAPORE	334.98	1250	2000	1200	YES
PANAMA	333.96	380	2000	1700	YES

Table 2. Real recordings of ship arrivals in Limassol port.

Type	GT	Date in Port	Duration in Port
Passenger	167704	15/09/2025	10h15m
Passenger	82820	06/10/2025	10h26m
Passenger	42,363	10/10/2025	10h55m
Passenger	69472	05/11/2025	16h11m
Passenger	44697	27/11/2025	1d5h22m
Cargo	95,514	07/12/2025	16h29m
Cargo	35,574	21/07/2025	15h
Cargo	35,878	27/07/2025	1d11h11m
Cargo	32427	03/11/2025	1d8h34m
Cargo	21,092	23/08/2025	19h27m
Tanker	1896	01/07/2025	1d2h44m
Tanker	30069	12/11/2025	15h
Tanker	3828	22/10/205	12h53m

Table 2. Real recordings of ship arrivals in Limassol port.

Type	GT	Date in Port	Duration in Port
Passenger	167704	15/09/2025	10h15m
Passenger	82820	06/10/2025	10h26m
Passenger	42,363	10/10/2025	10h55m
Passenger	69472	05/11/2025	16h11m
Passenger	44697	27/11/2025	1d5h22m
Cargo	95,514	07/12/2025	16h29m
Cargo	35,574	21/07/2025	15h
Cargo	35,878	27/07/2025	1d11h11m
Cargo	32427	03/11/2025	1d8h34m
Cargo	21,092	23/08/2025	19h27m
Tanker	1896	01/07/2025	1d2h44m
Tanker	30069	12/11/2025	15h
Tanker	3828	22/10/205	12h53m

2.4. Empirical Utilisation of Data

As established in the preceding literature review, ship electrical demand can be estimated by examining the relationship between a vessel’s physical and operational characteristics and its reported power consumption. Previous data-driven studies have demonstrated that variables such as vessel type, size, operational role, and usage profile can significantly influence onboard power demand, particularly during berthing when shore-side electricity is required for cold ironing operation. Building on this principle, the present study develops a data-driven regression-based methodology to generate synthetic power demand values for vessels where direct consumption measurements are unavailable.

The approach uses the three main data classes collected during the initial data-gathering stage: vessel type, gross tonnage, and reported power demand or peak power demand. These variables were selected because they provide a practical representation of the vessel’s category, scale, and electrical load requirement. Therefore, it is important to learn the underlying relationship between these input characteristics and vessel power demand, and to use this association to produce realistic synthetic estimates for ships with incomplete power data.

In this study, Random Forest Regression is employed due to its ability to capture nonlinear relationships between mixed data types and to reduce the limitations associated with simple linear assumptions [13]. In summary, the three types of data can be described as:

Vessel Type-Gross Tonnage

$$X={(Type}_i,{GT}_i)$$

(1)

Peak Demand-Gross Tonnage

$$({Peak}_{peak,i},{GT}_i)$$

(2)

Vessel Type-Peak Demand

$$({Type}_i,{Peak}_{peak,i})$$

(3)

Utilising the linear regression approach, the model is trained to identify the relationship between vessel characteristics and power demand. Associating vessel types with GT and known peak power demand creates a power density coefficient. This ratio expressed in kW/GT describes the relationship between vessel tonnage and power demand and is later used to calculate peak power for unknown data. Eq. 4 displays the relationship between Peak power and Gross Tonnage with αtype being the power density coefficient for each type of vessel.

$${\alpha }_{type}={\displaystyle \frac{P_{peak}}{GT}}$$

(4)

For that, the information fed into the model comprised of ship type, gross tonnage and peak power demand. Specifically, GT was used as the main size-related predictor, while vessel type was included as a categorical variable to account for operational differences between cruise ships, ferries, container ships, and tankers. During training, the Random Forest model constructed a set of (B) decision trees, where each tree was trained on a randomly selected bootstrap sample of the training data. At each node, the algorithm attempts to select the best Power value that reduces the Mean Squared Error. The final Random Forest prediction was obtained by averaging the outputs of all trees where B is the total number of trees and Tb is the prediction by each tree using the X input characteristics.

$$MSE={\displaystyle \frac{1}{N}}{\displaystyle \sum _{i=1}^N}{{(P}_{real,i}-P_{predicted.i})}^2$$

(5)

$$P_{predicted,i}={\displaystyle \frac{1}{\mathit{B}}}{\displaystyle \sum _{b=1}^B}{T_b\left(X\right)}^2$$

(6)

The predicted power demand was then used to calculate new vessel-type power density coefficients, expressed as

$${\alpha }_{type}={\displaystyle \frac{P_{predicted,i}}{GT}}$$

(7)

Continuing with the learning process, the raw information gathered from port operators and subscription databases was employed to classify each ship with the appropriate coefficient. Based on these characteristics, ships with similar characteristics and operational profiles were grouped together to ensure consistency in comparison and computational burden. Figure 1 illustrates the sorting algorithm flow diagram that categorises the ships into broader classes as mentioned in the previous section.

To evaluate the performance of the model over the large databases, a combined 80:20 train-test split with the Root Mean Squared Error is employed. This ensures that power density coefficients are not guessed nor based only on the proportional relationship between GT and Vessel type whereas, large prediction errors are penalised. This is illustrated in Eq8 where the difference between real and predicted values are squared which means large errors contribute more to the error. Essentially the dataset is first divided so that 80% of the vessel records are used for training and 20% are kept aside for testing. The model learns the relationship between vessel type, gross tonnage, and power demand using only the training data. After training, the model is applied to the unseen 20% testing data, and its predicted power demand values are compared with the known reported values. This approach is suitable for a large dataset because the model has enough data to learn from, while the testing subset remains large enough to give a reliable indication of how well the model performs on new vessel records.

$$RMSE=\sqrt{{\displaystyle \frac{1}{N}}{\displaystyle \sum _{i=1}^N}{{(P}_{real,i}-P_{predicted.i})}^2}$$

(8)

Eventually, these coefficients provide a standardised way to quantify differences in power use across ship classes and form the basis for later estimating the overall power demand of each vessel. Table 3 shows the examples of the power density coefficient allocations. Power density coefficient is shown in Column 4.

3. Results

3.1. Ship Daily Load Profile

Since a power demand coefficient was established for each vessel type, it was possible to predict a load demand daily curve for ships berthing in Limassol port. Initially, a load demand weight assumption was made depending on the vessel’s assigned power density coefficient. For instance, the typical daily demand profiles for ships follow a characteristic pattern. Passenger vessels experience higher loads during daytime and evening periods due to passenger activities, food preparation, and entertainment systems. Night-time loads remain significant due to HVAC and hotel services. Cargo vessels, in contrast, show relatively flatter load profiles because their electrical demand is primarily associated with auxiliary machinery and cargo handling systems operating only momentarily.

Hence, using the berthing records acquired from the Limassol port databases, 5 different ships were simulated to generate 24-hour operation cycle of vessels arriving in Limassol port. Table 4 illustrates the characteristics of the 5 ships in question.

For each ship, the trained model estimated the expected electrical demand profile throughout a 24-hour period by considering operational characteristics such as vessel category, gross tonnage, passenger capacity, and berth duration. At the same point, to ensure accuracy, the load curve was cross-checked with the real readings available. The generated load curves reflect variations between peak and off-peak operating states, reproducing realistic fluctuations in onboard energy consumption during real life conditions. By integrating real vessel schedules and operational constraints, the simulations were able to emulate dynamic port-side demand with improved accuracy compared to static estimation methods.

Figure 2. Output load demand cycles of 5 different vessels.

Figure 2. Output load demand cycles of 5 different vessels.

3.2. Accumulated Power Demand in Port

Ultimately, the previously defined datasets and computational procedures were utilised to derive an aggregated daily power demand profile for the port of Limassol. The modelling framework integrated the vessel classification algorithm and the power density coefficients assigned per category as well as temporal operational parameters extracted from berthing datasets such as arrival times and total berthed durations. These inputs were combined within a unified simulation environment to represent the time-dependent allocation of energy demand across port activities. The resulting formulation produces a continuous aggregate load curve, in 30-minute intervals, capturing the temporal progress of total electrical demand at the port of Limassol over a 24-hour operational cycle. Eq. 9 displays the number of intervals when calculating the accumulated power demand in the 24-hour operation cycle.

The simulation begins with the reading of the database containing vessel type and gross tonnage. Then, the model searches for the corresponding power density coefficient generated in the previous simulation. Once the correct coefficient is identified, the model calculates the peak power demand for the vessel in question. This carries on for all the ships arriving in port that single day. Subsequently, for each time interval if there is a ship berthed in port, its peak power demand is added to the loop. If there are multiple ships overlapping, their peak power demand is summed and added to the demand for that time instance. In essence, a nested loop is created that generates the accumulated power needs of the port for that day.

$$P_{port}\left(t\right)=\left[t_1,t_2,t_3,\dots t_N\right],N=48$$

(9)

Figure 3. Accumulated forecasted power demand of all ships berthed in Limassol port for 03/08/2025.

Figure 3. Accumulated forecasted power demand of all ships berthed in Limassol port for 03/08/2025.

4. Conclusion

Cold ironing, also referred to as shore-to-ship power or onshore power supply, enables vessels to receive electrical energy while berthed, allowing auxiliary engines to be shut down during port stays. In isolated electrical networks, cold ironing can also provide an opportunity to utilise surplus renewable energy that might otherwise be curtailed. However, the successful implementation of cold ironing infrastructure requires a reliable assessment of the electrical demand imposed by different vessel categories, making demand estimation essential for port planning, grid integration, and energy management.

This study developed a data-driven methodology for estimating ship power demand using vessel type, gross tonnage, and reported or peak power demand as the primary inputs. The application of Random Forest Regression enabled the identification of relationships between vessel characteristics and electrical demand, allowing the generation of synthetic demand values for vessels lacking direct measurements. A vessel-type power density coefficient was derived and used to estimate individual ship demand and construct representative daily power profiles for cruise ships, cargo vessels, and tankers during berthing. By aggregating these individual demand estimates, the study provided an assessment of the accumulated port-level power requirement, supporting the evaluation of future capacity needs and grid reinforcement requirements. The main contribution of this work to both research and industry lie in the development of a practical and transferable framework for estimating cold ironing demand from limited vessel data, offering a valuable tool for port operators, infrastructure developers, and energy planners in the design and deployment of sustainable port electrification systems. Furthermore, the connection of large maritime loads may introduce power quality challenges, including voltage fluctuations, harmonic distortion, and sudden load variations during connection or disconnection events. Therefore, future cold ironing deployment should integrate demand forecasting with detailed assessments of grid stability, infrastructure capacity, and power quality mitigation measures.