Machine Learning applied – updates on the SmartShip project.

What is the aim of SmartShip Project?

Smartship focuses on key environmental challenges in the maritime industry by building a multi-layered optimisation platform. The solution will manage and optimise fuel consumption and energy efficiency, delivering emissions control management functionality as well. Smartship takes into account the requirements of maritime sector regulations, applying the industry’s circular economy concepts.

Together with experienced researchers from Harakopio University in Athens, BlueSoft’s R&D Team applies Machine Learning to predict vessel sea routes optimally.

Each ship course depends on many external factors such as wind speed, sea flows, course speed, and traffic – especially nearby ports, channels, and straits. Times series predictions and Deep Learning algorithms applied to ICT software allow optimising the process of managing the vessels. The solution helps to significantly shorten distance and course time. It also reduces fuel consumption and fossil fuels emissions.

Our team received a large dataset from AIS vessel tracking and used it in predictions consisting of almost 370.000 position data points. We chose the most accurate prediction models such as ARIMA, PROPET, and LTSM – and then compared real-life positions with the ones trained by the algorithms.

Here is short a description of the time series methods mentioned above:

• ARIMA (Auto Regressive Integrated Moving Average) – the model provides a complementary approach to time series forecasting and is one of the most widely used methodologies. Its key purpose is to describe autocorrelations in the data.
• PROPHET – this algorithm detects the trends and seasonality patterns from the dataset first, then combines them together to arrive at the forecasted values.
• LSTM (Long Short Term Memories) – the model refers to Recurrent Neural Networks and is used to track the dependencies of new observations with historical ones. It requires a large volume of data.

The main task was comparing the methods listed above and choosing the most accurate prediction model for replicating real-life conditions.

Our work is the continuation of advanced studies carried out by our Harakopio University colleagues on extracting maritime traffic patterns described in the following scientific working paper: https://www.tandfonline.com/doi/full/10.1080/13658816.2020.1792914.

When employed in the industry, advanced Machine Learning predictions based on historical AIS data observations of vessel routes become an important competitive advantage. They optimize ICT hardware dedicated to managing vessels on board as well as remotely.

Here is the most recent update, together with a description of chosen methodologies and results.

Our developers carried out advanced tests of the chosen time series methodologies as follows:

PROPHET – the algorithm was developed by Facebook in 2017 – Prophet implements what is referred to as additive time series forecasting model. The implementation supports trends, seasonality, and holidays. The advanced test made on the vessel route dataset showed the method to be insufficient since data is not seasonal.

The orange line on the graph shows the trend, but it doesn’t predict the exact position of the vessel. We may assume that this methodology doesn’t give the expected results in terms of position prediction.

ARIMA – the result of our research departed from our expectations. It showed that this method

The last researched method applied to predictions was LSTM that mostly imitates human reasoning.

The blue line showing real-life data follows the orange line showing the predicted outcomes. This means that LSTM could perfectly predict the future position of the vessels. It was the most accurate methodology applied in the research. But the fact of changing weather conditions (for example, sea traffic) indicates that this method may only be considered as a partial solution.

The dataset was taken form AIS (vessel tracking responders) and used in predictions consisting of almost 370.000 position data points. A number of the most accurate prediction models were tested, such as ARIMA, PROPET, and LTSM and compared to real life positions with the ones trained by the algorithms.

Here is short a description of methods mentioned above:

ARIMA –  it provides a complementary approach to time series forecasting and is one of the most widely used methodologies. Its aim is to describe autocorrelations in the data.

PROPHET – it detects the following trend and seasonality from the data first, then combines them together to get the forecasted values.

Overall Trend

Yearly, Weekly, Daily Seasonality

Holiday Effect

By looking at the trends and seasonality detected by the Prophet, we can gain quite a lot of useful insights from the model.

It’s a procedure for forecasting time series data based on an additive model where non-linear trends fit yearly, weekly, and daily seasonality, plus holiday effects. It works best with time series that have strong seasonal effects and several seasons of historical data.

Advanced Machine Learning predictions based on historical AIS dataset observations of the  vessel routs employed in will become important  competitive advantage that optimize ICT hardware dedicated to manage vessels on board as well remotely.