=Paper=
{{Paper
|id=Vol-1815/paper14
|storemode=property
|title=Case Representation and Adaptation for Short-Term Load Forecasting at a Container Terminal
|pdfUrl=https://ceur-ws.org/Vol-1815/paper14.pdf
|volume=Vol-1815
|authors=Norman Ihle
|dblpUrl=https://dblp.org/rec/conf/iccbr/Ihle16
}}
==Case Representation and Adaptation for Short-Term Load Forecasting at a Container Terminal==
https://ceur-ws.org/Vol-1815/paper14.pdf
142
Case Representation and Adaptation for
Short-Term Load Forecasting at a Container
Terminal
Norman Ihle
OFFIS, Institute for Information Technology, Oldenburg, Germany
norman.ihle@offis.de
Abstract. The electricity consumption of a terminal is mainly related
to the number of container movements and the weather of each day. With
the introduction of electric mobility for heavy duty container carriers at
a seaport container terminal short-term load forecasting gains an impor-
tant part in the procurement process. This paper describes a case-based
approach to the forecasting of the electricity consumption time-series of
the following day based on historical consumption load curves. It mainly
focuses on the case representation which is based on a daily view on
the so-called sailing list that is used to plan terminal operations and
the adaptation processes that are applied to the time-series after case
retrieval. The evaluation of the approach shows some promising first re-
sults.
1 Introduction
The introduction of electric mobility in industrial enterprises increases the im-
portance of energy procurement even in sections where energy procurement was
not in focus until today. Especially when some processes are changed from con-
ventional to electric mobility in the division of logistics, economic potentials
arise through the use of flexible electricity supply contracts. One example of
such a logistic system is a seaport container terminal. Tasks like the transport
of containers from the quay crane to the bulk storage area can be automated
and electrified. This can be achieved by using battery-electric powered engines
in the heavy-duty vehicles instead of diesel engines. In opposite to Diesel, that
can easily be stored on terminal grounds, the fluctuating demand of the fuel
’electricity’ has to be procured on a short-term basis. When using the electricity
exchange, the day-ahead market is the last chance to match the procured elec-
tricity load curve with the latest demand forecast values. The forecast of a daily
load curve, a time-series with 15-minute time stamps, is referred to as short-term
load forecasting (STLF).
Over time many methods for short-term load forecasting have been developed
and are now in use. Some of those methods are developed for specific scenar-
ios, others follow a more general approach. Regarding the application of these
methods to the load forecast of a container terminal there is no experience doc-
umented in the scientific literature. Even though the more general methods can
Copyright © 2016 for this paper by its authors. Copying permitted for private and academic purposes.
In Proceedings of the ICCBR 2016 Workshops. Atlanta, Georgia, United States of America
143
be applied to all kind of energy consumers, there is no systematic evaluation
which method yields the best results when being applied to a single industrial
site from the logistic domain. Since the consumption of electricity at a container
terminal is subject to some special features that might not be available for other
load forecasting scenarios, e.g. the consumption relies highly on the number of
containers being handled, a solution based on Case-Based Reasoning (CBR) is
proposed. The idea is that days with similar logistic operation requirements have
similar load consumption patterns. This is due to the fact that only a few han-
dling equipments have a great impact on the load profile, especially the quay
and yard cranes and the reefer container.
In the following a case-based load curve forecasting approach is presented that
is based on the list of expected ship arrivals and departures. In chapter 2 the case
representation is introduced with a focus on how the data for exactly one day is
transformed and represented. Chapter 3 describes shortly how time values are
handled in regard to similarity. If a similar case is retrieved the corresponding
consumption time-series can be adapted to fit it to the forecasting situation.
Two approaches for the adaptation are introduced in chapter 4. Chapter 5 then
presents the first results of the evaluation before chapter 6 discusses some related
work. Chapter 7 closes the paper with the conclusion and an outlook on further
work.
2 Case Representation
In seaport container terminals the number of logistic operations to be planned
is very high. Containers are continuously delivered and picked up via train,
truck or ship, the quay and yard cranes have to be assigned accordingly and
the storage areas have to be administrated. The most important source for the
operation planning is the list of ship arrivals and departures, the so called sail-
ing list. The list is a table that includes the most relevant information for the
expected ship arrivals and departures. The list is continuously updated with the
newest available information that the terminal receives from the ship operators
or freight companies. For example, a ship reports its departure from the previ-
ous terminal and the expected travel time to the next terminal. Due to weather
influences on the traveling speed and other circumstances this information can
vary throughout the route. One day before a ship arrives the expected arrival
time is pretty reliable. Figure 1 shows some examples from the sailing list of the
Container-Terminal Altenwerder (CTA) in Hamburg (Germany) from Septem-
ber 2013. The data from the sailing list can be used to build up the case base
while the time-series data of the electricity consumption is stored in a dedicated
database outside of the case structure due to performance and efficiency reasons.
For simplification reasons one row of the sailing list will be referred to as arrival
in the following.
144
Fig. 1. Example of sailing list data
2.1 Using Sailing List Data for Case Modeling
In a first step data from the columns of the sailing list can be used for case
modeling using an attribute-value representation. Besides the expected arrival
and departure times of each ship, the sailing list includes information about
how many containers have to be loaded and unloaded from each ship during the
berthing time. These data together with the ship type can then be used to find
similar days in the past. Similar in this case means that ships with a comparable
number of containers to be handled have been arriving or departing at almost
the same times of the day. The information is adopted for case modeling:
– JSNR: Attribute CSJSNR
– Ship type: Attribute CSType
– Expected arrival time: Attribute CSArrival
– Expected departure time: Attribute CSDeparture
– Number of containers to unload: Attribute CImport
– Number of containers to load: Attribute CExport
The JSNR is a unique identifier of each ship’s berthing time. The ship type
distinguishes between a barge, a feeder ship and different seagoing vessels, often
represented by the code of the regular tour they are operating on. The expected
arrival and the expected departure times describe the expected berthing time,
which is stored, expressed in minutes, in an additional attribute (CSBerthing ).
2.2 Splitting the Berthing Time to Represent One Day
Some entries in the sailing list represent a berthing that includes two or more
dates according to the calendar, e.g. the ship APL VANDA with JSNR 307896 in
Figure 1. Since the time span for the electricity consumption forecast is one day
it is necessary to build up arrivals that represent the information for exactly one
day from 00:00 to 23:59 o’clock. The available information has to be adapted to
fit into a daily view of the sailing list information. To build up the daily view a
berthing that exceeds the date limit can be split up to fit into multiple (virtual)
arrivals, each representing the data for exactly one date. The information for
JSNR 307896 can be split up to the following three entries:
145
– Arrival 06.09.2013 17:55 - Departure 06.09.2013 23:59
– Arrival 07.09.2013 00:00 - Departure 07.09.2013 23:59
– Arrival 08.09.2013 00:00 - Departure 08.09.2013 14:00
The number of containers to be handled per ship have to split up as well to
fit the new virtual arrival information. For the split up, knowledge about the
crane assignment at the seaside of the terminal can be used since the container
handling is not constant during the berthing time. Using the simulation model
presented in [2] and [3] it can be shown that the container handling numbers for
large ships (more than 250 containers to be handled) decrease over the berthing
time. After some preparation time the handling number starts at a high rate that
is decreasing over time because cranes get allocated to other quay areas due to
little space at the quay or higher priority of other arriving ships. To approximate
this observation, a weighting function is introduced to calculate the number of
containers C that is handled at each full hour hi of the berthing time. Full hours
are chosen here to approximate the distribution of the handling numbers over
the time because more exact information is not available and a break-down to 15
minute values has no further advantage. Based on the average linear handling
rate that is needed to load and unload all designated containers Ctotal of the
ship during all full hours of the berthing time hn it weights the first hours high
and decreases linear over time:
0, 6 Ctotal
Chi = (1, 3 − ∗ (hi − 1)) ∗ with hi = 1, ..., hn (1)
hn − 1 hn
Based on this function the container numbers of a sailing list entry can be
split into handling numbers for each virtual arrival and departure. If a ship
arrives at the exact beginning of an hour, the first full hour reserved for handling
preparations is not regarded for htotal and gets assigned a container handling
number of 0. This is also applied if the ship arrives within a full hour.
2.3 Adding Additional Daily Information
Having adjusted the sailing list so that all available data can be assigned to
exactly one day, daily cases can be built up. For each daily case additional infor-
mation can be added. The values for the average temperature and wind speed
of each day are of special interest because it can be shown that particularly the
temperature value has an impact on the electricity consumption. Besides the
weather information the weekday is added to the case since different repeat-
ing consumption patterns can be observed in the electricity consumption load
curves of different days. While the differences between the weekdays are not as
remarkable, differences between load curves of weekdays and days on the week-
end differ even though the container terminal realizes a 24/7 operation. This
might be explained by the lesser use of the office building and the Sundays driv-
ing prohibition for trucks in Germany. A special kind of weekday are holidays
at which the terminal is actually shut down as well as the day before and the
day after such a holiday. Figure 2 shows some typical load curves on different
146
Fig. 2. Example of typical load curves
days. Two attributes with daily information can be derived from the sailing list
information: the number of arrivals for a day and the overall number of con-
tainers that have to be handled. The second value is a good indicator for the
average electricity consumption level for that day. The more containers have to
be handled the higher the average level. Additionally an attribute saves the date
value. It has the function to link to the corresponding time-series data of the
day.
3 Similarity for Time-Based Attributes
When comparing two cases the arrival and the departure time of each ship plays
an important role to assess the similarity. In the sailing list the information is
available in a date-time format, the date being separated from the time by a
blank. Since the single cases represent a daily view on the terminal operations,
the date information can be disregarded: only the time information is interesting
for similarity calculations. In order to be able to use simple existing similarity
measures the time value is converted to an integer value. This value represents
the number of minutes that have passed since the start of the day. The time
value 00:00 therefore is represented by 0 and the end of the day at 23:59 by the
value of 1439. If the absolute distance between the time values of two different
arrival attributes is less than 60 both arrivals were not more than an hour apart
at the point in time of the day and are therefore quite similar. The similarity
decreases fast if the difference increases. To represent this, a sigmoid similarity
function based on the distance is chosen as suggested by [1]:
1
simCSArrival (dCSArrival ) = d−120 (2)
e 30 +1
The same function can be used for the departure time as well and with slightly
different parameters it can also be used for the attribute CSBerthing .
147
4 Adaptation
The adaptation of the electricity consumption time-series related to similar cases
is twofold. Both approaches are based on data stored in the cases. The first
adaptation approach focuses on the single values of the time series and tries to
adapt the values of each hour of one day based on the difference in the number of
container to be handled in each hour. The second one is based on the differences
in the weather data and is applied to the general level of all time series values
for one day.
4.1 Hourly Container Number Comparison
It can be shown that the number of containers handled in an hour has a remark-
able impact on the characteristics of the electricity consumption time series. The
higher the number of containers handled the higher the electricity consumption.
The idea behind the first adaptation approach is to find hourly differences in
the number of containers handled on the day of the most similar case and the
day of the query. For this purpose the aggregated container moves per hour i are
calculated. For each arrival that is related to the day of CASEsim the container
number is split into hourly values using the weighting function discussed in chap-
ter 2.2. In this context not the exact daily values of the virtual arrivals are used,
but the original entries are used in order to be able to use the weighting function
properly and the hourly values are not stored. The same can be done for each
arrival related to the day of the query (QU ERY ). Having done so, the difference
di between the container handling numbers of CASEsim and QU ERY can be
calculated for each hour. The first approach was to use this hourly difference
directly to apply an adaptation factor to the time series values of that specific
hour, but the evaluation showed that this approach had the shortcoming that
it does not take into account the operations before and after the hour, but they
also have an influence on the electricity consumption of that hour since prepara-
tions or post-processing also have to be executed. It might also happen that the
difference in the container handling number is subject to strong fluctuations, for
example when large seagoing vessels with a huge amount of containers arrive and
depart, but the fluctuation in the electricity consumption is not as high at that
point of time because of smoothing effects that cause some delay in the change
of consumption pattern. For this reason not only difference di in the hour i is
regarded when choosing an adaptation factor, but also difference in the hour
before i − 1 and the difference in the hour after i + 1 are also considered. At the
moment an adaptation factor is assigned to different times of day depending on
the discussed differences using a rule set. Table 1 shows an extract of the rules
as they are applied for the hour 8 (08:00 - 08:59 o’clock) of the time series. The
F actor is added to the time series values of the hour. It can be shown that the
fluctuations are higher in the beginning of the day, so the adaptation factors
are higher than the ones for morning and evening hours, while the adaptation
factors for the hours in the middle of the day are rather low.
148
Hour i di di−1 di+1 F actor
8 > 200 > 150 > 150 0.4
8 > 150 > 120 > 120 0.3
8 > 100 > 70 > 70 0.2
8 > 50 > 20 > 20 0.1
8 < -50 < -20 < -20 -0.1
8 < -100 < -70 < -70 -0.2
8 < -150 < -120 < -120 -0.3
8 < -200 < -150 < -150 -0.4
Table 1. Adaptation factors for hour 8 of a day
4.2 Regarding Seasonal and Weather Influences
The second adaptation approach is based on the available weather information.
The highest influence on the electricity consumption is based on the air temper-
ature. It can be observed that the energy consumption of the container terminal
is on average significantly higher in winter than it is in summer. This is mostly
due to additional heating that is required when having low temperatures and
more lighting is needed on winter days than in summer days. Since the available
temperature values are average temperatures for one day, the differences in the
temperatures can be applied to increase or decrease the level of the electricity
consumption, but they are only supposed to be applied when the case represents
a winter day and the query a summer day or the other way around and the
temperature difference is greater than 5 degrees Celsius, since a difference of 5
degrees on summer days does not show a significant effect on the electricity con-
sumption. Additionally boundaries for the adaptation were defined. This means
that if a consumption value is already very high and the temperature based
adaptation would raise this value beyond the upper boundary, the adaptation
is not applied to avoid a forecast of values that are usually not reached under
normal operating conditions of the terminal. The same applies to very low values
and the lower boundary.
5 Evaluation
For the evaluation the sailing lists and the corresponding electricity consump-
tion time-series of the years 2010 through 2013 of the Container-Terminal Al-
tenwerder are used. Today, forecasts are generated by using the last week’s values
as forecast values (last week approach). While these values fit in some cases, they
also generate some remarkable deviation from the forecast to the metered con-
sumption in other cases. So besides improving the forecasting accuracy, it is also
a goal to decrease the maximum deviations in the forecast compared to the real
consumption. For the CBR approach the adapted electricity consumption val-
ues of the most similar case are used as forecast values. Besides this forecast a
second one based on CBR is calculated by using the mean value of the adapted
149
time-series values of the three most similar cases instead of only the one of the
most similar case. This forecast will be referred to as CBR Top 3 approach.
The first results of the evaluation show that the CBR approach can be success-
fully applied for forecasting the energy consumption of a container terminal.
When forecasting the electricity consumption load curve for every day in June
2013, the last week approach yields an average Mean Absolute Percentage Error
(MAPE) of 12.0 while the CBR approach yields to 10.0 in average. Having a
yearly electricity consumption of more than 65000 MWh, a 2% better forecast
avoids penalty fees that have to be paid when the procured electricity load curve
does not fit the actual metered one can be avoided to a great extent. With the
CBR Top 3 approach the results even yield to a mean of 8.6 which means a
further improvement on the forecast. Figure 3 shows the MAPE values for every
Fig. 3. MAPE values for June 2013
single day in June 2013 for the three approaches. It can be seen that on some
days the MAPE values of the three approaches are close to each other while on
some days the last week approach has very high values. On the 4th and the 8th
of June the mean absolute percentage error of the last week approach is even
higher than 30% while with the CBR Top 3 approach the highest values are
slightly above 15%. This shows that the fluctuations can be decreased and the
risks of over- or under-procurement of the electricity for a day is minimized. The
Mean Squared Error (MSE) values of each day in June confirm these findings
that can be observed when forecasting the following month as well.
Over the year 2013 the monthly average MSE of the CBR Top 3 approach is
never higher than 0.07 while for the last week approach the maximum average
value is double that value. The average MAPE for the year 2013 is 8,87% for the
CBR Top 3 approach, while the currently used last-week approach has a MAPE
of 12,32% in average.
150
6 Related Work
In the energy consumption forecasting domain only a few works have been pub-
lished that present CBR approaches. [6] present the system FUTURA to fore-
cast the medium-term load consumption in Peru. To adjust a calculated start-
ing solution, the CBR module uses the average of the retrieved similar cases
to smooth out fluctuations of the starting solution. The similarity is based on
time attributes like month, weekday, and the time of day. Weather influences
are disregarded. This adjusted solution is then adapted based with the expert
knowledge regarding the long-term trends in environmental development. It is
stated that the expert knowledge can change the load curve up to 20% depend-
ing on the influencing parameters. [7] presents an approach for the forecast of
power peaks in a distribution grid based on CBR methods. As attributes the
author uses weekdays, weekday-types and several weather indicators. The case
base is clustered with methods of Self-Organizing Maps (SOM) to allow for a
highly efficient retrieval. The k-nearest neighbor retrieval uses a weighted sum
of all attributes of the request and a cluster. The most similar case is then re-
trieved out of the cluster with the highest similarity value. The power peak of
this most similar day is then taken as power peak for the forecast day adapted
by a factor that represents the yearly growth rate in the overall load of the grid.
In a case study it is shown that the suggested solution outperforms simple last
week value approaches and even artificial neural networks. [4] introduce a system
for very short-term load forecasting of office buildings based on CBR-methods.
The similarity is assessed with a focus on weather indicators like temperature,
moisture, etc. The weights of the single attributes are dynamically adjusted re-
garding the current season and current weather conditions. Using a subset of
the most similar solutions a forecast for the next three hours is built up. It in-
cludes a prognosis whether a new power peak will occur in this time horizon.
The evaluation of the system shows that it can be applied successfully. The mean
squared percentage error is stated to be around 12 to 14 percent. [5] extend this
approach to a forecast horizon of 6 hours and present an on-line implementation
of the system. They also specify the input attributes that are used. Besides the
weather indicators and the current electricity consumption they also use power
values of air conditioning and heaters currently in use and room temperatures
currently metered within the office building. None of the related works deal with
the daily short-term load forecasting of single industrial consumers based on
operation plans.
7 Conclusion and Outlook
This paper presented an approach to use CBR methods to realize a system for
short-term load forecasting for a single industrial customer, a container terminal,
and focused on the case modeling and the adaptation of the time-series. The
forecast of the electricity consumption for every quarter of an hour of the next
day is based on the logistic operation plan that is set up for the next day. This
151
plan includes data about expected ship arrival and departure times as well as
information about the number of containers to be handled with each ship. This
information can be used to find days in the past with a similar operation plan.
The electricity consumption time-series of a similar day can than be adapted
based on differences in the operations of each day and based on weather data.
The approach has some advantages when compared to other established load
forecasting methods: it is generally applicable to maritime container terminals,
the computational cost is rather low and the results are comprehensible. The
evaluations shows that the approach can be applied successfully.
Further improvements might be reached by integrating information about
the container numbers that are delivered and picked up by truck or train to
approximate the number of container handles and their course more exactly.
At the moment this number is estimated because of lack of the data. For the
adaptation process a yearly factor can be introduced that will represent the
yearly average change in electricity consumption that can be observed over the
previous five years. This is also important because in general the case-based
forecasting approach tends to underestimate the actual values when applied at
the use-case CTA, where the average yearly electricity consumption constantly
increased during the last years. It is to be investigated if a CBR-STLF-approach
for single electricity consumers based on operation plans can be generalized to
further domains than container terminals. Logistic systems can be in focus here
as well as more different systems like industrial production sites.
References
1. Bergmann, R.: Experience Management: Foundations, Development Methodology,
and Internet-based Applications. Springer-Verlag, Berlin, Heidelberg (2002)
2. Grundmeier, N., Hahn, A., Ihle, N., Runge, S., Meyer-Barlag, C.: A simulation
based approach to forecast a demand load curve for a container terminal using
battery powered vehicles. In: Neural Networks (IJCNN), 2014 International Joint
Conference on. pp. 1711–1718. IEEE (2014)
3. Grundmeier, N., Ihle, N., Hahn, A.: A discrete event-driven simulation model to
analyse power consumption processes in container terminals. In: Simulation in Pro-
duction and Logistics 2015. Fraunhofer IRB Verlag, Stuttgart (2015)
4. Monfet, D., Arkhipova, E., Choiniere, D.: Evaluation of a case-based reasoning
energy prediction tool for commercial buildings. International Conference for En-
hanced Building Operations (2013)
5. Platon, R., Dehkordi, V., Martel, J.: Hourly prediction of a building’s electricity
consumption using case-based reasoning, artificial neural networks and principal
component analysis. Energy and Buildings 92, 10 – 18 (2015)
6. Vilcahuaman, R., Melendez, J.: Improvement of electric load forecasting reusing
past instances and contextual based adaptation. In: Transmission and Distribution
Conference and Exposition: Latin America, 2004 IEEE/PES. pp. 271–276 (Nov
2004)
7. Wang, Z.Y.: Developed case-based reasoning system for short-term load forecasting.
In: 2006 IEEE Power Engineering Society General Meeting. pp. 6 pp.– (2006)