Procurement of electricity gains more and more focus in enterprises especially with the introduction of electric mobility. At a maritime container terminal the electricity consumption is highly related to the number of container movements of each day. Short-term load forecasting (STLF) methods have not yet been systematically researched when applied to container terminals. Therefore it seems reasonable that the inclusion of knowledge about the number of next day's container movements might improve the forecasting of next day's electricity demand of the container terminal. One way to include this knowledge in the forecasting process is to use Case-Based Reasoning methods. In this thesis a concept for a corresponding system is outlined and implemented. In addition, also further concepts for using established STLF-methods are described and implemented. Besides the system based on Case-Based Reasoning, naive methods, time-series models, arti cial neural networks and simulation are being implemented and compared to each other. The implementations are tested with data from the use-case ContainerTerminal Altenwerder in Hamburg, Germany. Goal of the thesis is to evaluate, which method is best suited in what situation.
Rising energy prices and uctuating power generation from decentralized sources
into grids that were not designed for this purpose are controversially discussed
topics nowadays. With the introduction of E-Mobility to maritime container
terminals new possibilities for the terminal operator arise to optimize the energy
demand which have not been possible before. Not only does the relevance of
energy demand rise in the operational strategies but also for the rst time
exibility in regard to the time of the energy demand can be achieved. This exibility
results from the fact that there is a possibility to control, within a certain range,
the point in time when the battery of the vehicle is recharged. The exibility
can be used in several use-cases to gain economic bene ts. For example,
balancing energy can be o ered to the grid or peak shaving of the load curve can
be applied. The basis for the use of this exibility is the knowledge about the
expected power demand of the container terminal, especially the demand for the
next day as time-series of 15-minute values. This forecast is widely referred to
as short-term load forecasting (STLF). STLF methods have been a topic in
scienti c research for a long time. Several methods have been established by now.
Most of them are applied to complete grids or to larger groups of electricity
consumers. Reviews on established methods can for example be found in [
In the following the single methods are shortly introduced, each with a brief introduction how they can be applied for forecasting the load of a container terminal. After that, the current state of work is described and a short outlook on the further work is presented. 2 2.1
From the class of naive methods the equivalent day approach is commonly used
by utility companies to predict next days load curves. Equivalent day in this
case means to take the data of the same day of the week before as forecast for
the same day (according to the calendar) in the future [
From the class of mathematical models originate di erent time-series analysis
and regression models. ARMA describes a number of linear models that can be
used for stationary, time-discrete stochastic processes. For this case
autoregressive processes (AR) are combined with moving average models (MA). The term
autoregressive describes stochastic models, that explain an output variable yt
by the linear combination of past values and a current error term .
MovingAverage Models explain future values by the error terms of the past values. The
error terms are referenced as the deviation of a past value from the average. An
ARMA model can be written as
yt =
0 +
1yt 1 + ::: +
pyp 1 + t
1 t 1
:::
q t q
describes the weights of the autoregressive part and the weights of the
Moving-Average part. p refers to the order of the model that describes how
far values date back in the past. The challenge of the model is to estimate the
weights. Methods like Least Squares or Maximum-Likelihood can be used for
this purpose. Pre-condition is that the time series is stationary, that means that
trend and seasonality have been removed, and the average of the remaining
series is about zero. This can be reached by component analysis and integrating
the time series (ARIMA-Models). For the use of ARIMA-Models for forecasting
purposes at the container terminal the Hyndman-Khandakar algorithm [
Arti cial Neural Networks (ANN) are inspired by the structure of the human brain. They consist of a large number of parallel processing units. These neurons are quite simple units that are connected to each other. These connections can be activated using given rules. Each connection from neuron i to neuron j has a individual weight wi;j that is adjusted during the training process that is needed to prepare the network for its task. Each network consists of an input-layer, which receives the input data, a number of hidden layers that are responsible for the computation, and an output-layer for the result. Each neuron has an activation function that is responsible for taking the weighted sum of the inputs and calculating the corresponding output for the neuron. Di erent ANN-structures and training algorithms have been proposed over time. For the use of an ANN for forecasting purposes the number of container movements per quarter hour can be used as input. A network with one hidden layer yields promising results when trained using a common back-propagation algorithm. 2.4
Simulation is a problem solving method which replicates a system in an
executable model as exact as possible. The advantage of simulation in comparison
to analytical methods is the modeling of system-speci c dynamic dependencies
and reciprocal e ects over time. Stochastic e ects can be observed [
Case-Based Reasoning (CBR) is a method from the eld of Arti cial Intelligence that is based on the assumption that similar problems have similar solutions. If a similar problem to the current problem has been found, the corresponding solution is adapted to t the current problem situation. In a general approach the problem description as well as the solution build up a case. A number of cases is stored in the case-base from which cases similar to the current problem are retrieved. CBR uses domain knowledge for case modeling and adaptation. In order to use this method for forecasting the electricity demand load curve of a container terminal the case can be composed of attribute-value data, so a structural CBR-method is used. The problem description is based on the list of ship arrivals and departures which also includes the number of containers to be handled per ship. It is part of the operating plan for the terminal. All information about ship handling and in uencing external factors like the weather for exactly one day build up a case together with the load curve of that day. Based on next day's list of ship arrivals and departures a similar day can be found. The corresponding load curve of that similar day can be adapted regarding di erences in the container numbers, weather in uences or seasonal e ects to compute the corresponding forecast. 3
The research on this thesis began in spring 2015. For the implementation and the evaluation ship arrival and departure data as well as electricity consumption data for 4 years from the Container-Terminal Altenwerder is available. Currently, a concept for the load forecasting process for each of the described methods has been described and corresponding prototypes have been implemented. First forecasting results of each method are available. The next step will be to adapt the prototypes to improve the forecasting results based on knowledge gained from the rst results. For example, the CBR-system constantly underestimates the real consumption. This is due to the fact that the average consumption of the terminal increases each year. So a factor to regard this average raise can be introduced and cases from dates closer to the current date might get weighted higher. After this a systematic evaluation of all methods against each other will be performed. This evaluation is supposed to not only regard the forecasting accuracy but also further factors like portability, computational e ort or implementation e ort. The research is supposed to be nished by summer 2017 so that the thesis can be completed within the year 2017.