1. Smart grids as Common-pool
Resources
Managing Electrical Vehicle Charging
through Evolving institutions
Amineh Ghorbani
Energy and Industry Group
Faculty of Technology, Policy and Management
2. 2Titel van de presentatie
Smart grids
• Mitigation strategies in cities
• Adaption of renewable energy systems
• Solar panels
• Electric Vehicles
• Electricity grids need to be smart to
• Manage the inflow and outflow of electricity: consumers vs.
prosumers
• Huge overloads: electric cars
3. 3Titel van de presentatie
Managing smart grids
• Centralized vs decentralized
• What happens to my privacy?
• It’s too difficult, can somebody do it for me?
• Multi-agent systems
• Autonomous entities called agents to manage energy usage
• Agents represent households
Maximize profit – self-interested
• Minimizes user interaction but still requires central control
Household information needs to be shared
• Attractive on the consumer side, but what about grid
operators?
4. 4Titel van de presentatie
Smart grids as Common-pool resources
• Grids are CPR because:
• Users cannot be excluded from them
• They are limited in the amount of electricity they can provide
• (very similar to irrigation systems)
Question:
Can viewing smart grids as common-pool resources that are
self-organized by artificial agents allow more efficient
management of the grid?
While minimizing information sharing and user interaction.
5. 5Titel van de presentatie
The grid and the tragedy
Tragedy of the commons
6. 6Titel van de presentatie
A CPR-oriented multi-agent system design
• Assumption: only considering EV owners as a starting point.
• Electricity as resource unit
• Artificial agents come up with behavioral strategies about their
electricity usage patterns
• Shared strategies will eventually turn into institutional rules.
• No cheating no monitoring, no sanctioning
7. 7Titel van de presentatie
Design Elements
1. Agent properties and attributes
2. Structure of behavioral strategies
• Voltage rules vs. Time rules
1. Selection of behavioral strategies
• Learning
• Copying
• Innovating
1. Selection of institutions
• Voting
• One institutions selected at a time
• Structure: ADICO
• Change over time
8. 12Titel van de presentatie
Results – Performance indicators
• Average state of charge (SOC)
• Dependent on the node of the agent (heterogeneous) , distance from
the main transformer
• SOC of all agents
• SOC of the weaker agents
• Number of weak failures: unable to charge fully
• Number of strong failures: unable to charge sufficiently
9. 13Titel van de presentatie
Results – First look
• Comparison with and without software agents
• The average SOC over all agents is increased.
• There is a redistribution of access from advantaged agents to the
disadvantaged.
• The average number of weak and strong failures decreases
• Comparison with and without institutions
• The average SOC is significantly bigger with institutions
• Weaker agents perform better with institutions
• The average number of weak failures is reduced but not always
• The average number of hard failures is reduced.
10. 15Titel van de presentatie
Results – Under what conditions does
the model perform best?
1- What type of
rules?
Voltage Rules work
much better than time
based rules
11. 16Titel van de presentatie
Results – change of institutions over time
The adaption and change of institutions increases the
performance of the model over time.
12. 17Titel van de presentatie
Conclusions
• Viewing Smart grids as CPR systems and allowing self-
organization and endogenous institutions can result in a
significant increase in the performance of grid.
• A mutli-agent system design can minimize user interaction, and
allow decentralized control.
• More adjustments and reconfigurations of the design required to
maximize the optimal conditions.
Thank you for your
attention!
Editor's Notes
Urban population is responsible for 70% of greenhouse gas emissions worldwide. Consequently, citizens hold the key to mitigate these emissions. The mitigation strategy that citizens are increasingly becoming involved in, is the adaption of renewable energy sources (RES) for houses and cars. The production and consumption of renewable energy sources and the use of electric vehicles (EV) is promising, but poses problems to the low voltage grid.
Smart grid technology can be used to manage and thus mitigate the negative effects RES and EVs have on the grid.
In order to minimize interaction with users, intelligent agents can be designed to take over many user responsibilities. Artificial agents can also increase privacy by minimizing information sharing. Nonetheless, with current multi-agent system designs, a centralized management is still needed, which would require private infor
However, most centralized and decentralized approaches for the management of smart grids lack acceptability because they fail to meet social requirements such as respecting privacy and minimizing user interaction.
mation.
Centralized control of an agent-operated grid system can be omitted, when grids are viewed as common pool resources (CPR). Grids are CPRs because their users cannot be excluded from them and they are limited in the amount of electricity they can provide. Smart grids are nested CPR problems. They can be looked at as either a supply side provisional or an assignment problem. Although a CPR management approach can result in the tragedy of the commons, allowing the grid to be operated by endogenous institutions can be a promising solution.
How much information sharing in the end?
The strategies + preference values
There will be 24 homes, base
load of 6kVA, connected to the transformer with 4 lines. Between each house-
hold there will be a line of 100 meters. This way the total base load is about
Sbase = 144kV A and the total length of each cable is 600 meters.
The simulation is run with rational agents each of which will charge as much
as possible. Each electric vehicle arrives at home, gets reconnected to the grid
and its software agent does what is most rational, it charges the battery of the
electric vehicle as long as the battery is not full. Each agent theoretically has
access to the grid, and each agent does what is most rational in this situation.
The result of such a simulation run is given in graph 1. In this graph the state
of charge over time is represented for dierent levels. The level in this case is
an indication of how close or far from the electricity source the electric vehicle
is situated. The higher the level the further the EV is from the source. What
can be seen is that the state of charge for EVs on level 5 and 6 are not able to
recover to full charge after being drained. They sometimes even drop down to
0, meaning completely empty.
In this research, we have designed a multi-agent system (MAS) platform for the management of smart grids by building socially-inspired algorithms based on theories of common-pool resource management. We allow agents who represent the owners of EVs to define their own strategies following the ADICO grammar of institutions based on their preferences and past experience. The agents collectively decide about the institution that they must all comply with through a voting system. The agents are designed not to cheat. Therefore, no monitoring is required. The institution continues to evolve to reach a desired situation for all. To test our MAS system, we implemented a simulation platform that mimics the real operation of the grid.
One tick is 15 minutes, one day is 96 ticks, each day starts at mid-night
The arrival leaving time of agents is gaussian distributed, the battery drain is also gaussian distributed.
Border count voting: each rule in the memory is given a weight, from highest to lowest. The weights are summed up for each rule and the highest one is selected.
Q: when do agents vote? How often does it change?
Each agent goes through the following process, depicted in gure 5. The process
consists of two parts. The rst part takes place every 15 minutes and is done
for one week. The second part is done once every week. The part that takes place every 15 minutes is very simple. For one week in
intervals of 15 minutes, the agent chooses an action based on their internal rule. Each week the agent evaluates the rule it has been living with based on how
high their battery state of charge over this past week was, and commits the rule
and an associated preference value to its memory. When the memory of the agents is not yet lled, the agents develop
the rules individually.
Institutional acceptance threshold: if a rule that has won the vote is not in at least threshold percentage of agents memories, the agent does the individual selected procedure again.
I think the agents assign preference values to the institutional rules as well, so the rule can actually go out of the memory of the agents if it’s not in the top-five preferences.
Q: what is a preference value?
The state of the agents is its current state of charge + the voltage at its node and the current time. The actions tell the agents how much it should charge in the next time step.
Rules are either based on the curent time and SOC, or on the voltage of the grid point and SOC
A time rule has a similar structure. It contains two times, tbegin and tend,
to dene a time interval and a minimum state of charge (socthreshold). If the
agents state of charge is above the minimum, dened in the rule, and the time
is within the specied time interval the agent takes action1, otherwise the agent
takes action2.
A voltage rule has a voltage and a state of charge threshold and two actions
- action1 and action2. If the voltage of the agent is below the voltage threshold
dened in the rule and its state of charge is above the minimum state of charge
dened in the rule, the agent decides to take action1. If this condition is not
true the agent takes action2. Voltage rule: (State=Blue, Condition = Green,
Action = Red)
If
Memory: keeps track of experience: rule + preference value pairs
Preference value defined as the average SOC over the past week
The first five weeks the agents only develop the rules them-
Selves
Voting will be omitted for one week if the institutional rule
is not good enough
Q: what does perfoming poorly mean?
With agents means that there are agents who copye, learn etc and have a memomry. Without agents means that every agent just charges as much as it can.
To evaluate the model on a first glance the results from the first experiment are used. The model will be compared to a model run, in which there is no control and every agent tries to charge as much as possible as long as its battery is not completely full. To achieve this dummy agents are created, which will be called 'greedy agents'. Those greedy agents will simply charge every time, their battery is not full at 5kW.
So far, the number of weak/hard failures and the average soc was looked
at over the whole time. Since the failures are events and not variables that
can be kept track of over time, the number of failures cannot be plotted over
time. Dierent time intervals can be dened and the number of failures within
those time intervals can be compared to each other to see how the performance
changes over time. For this the time, when there is an institutional rule is split
into three segments. The time when there is an institutional rule is split into an
early a middle and late part of the simulation. Figure 22 shows the distribution
of the average SOC for the three dierent time intervals. Time progresses from
left to right. The red bars are runs with mostly voltage rules in them and the
blue bars are runs with mostly time based rules. There is a shift towards higher
values for both kinds of rules, but the shift is stronger with voltage based rules.
In the early phase, left graph, the number of hard failures are
distribute up to 20 failures, per run. In the next two time intervals the number
of hard failures drops to 0 for all the runs, which use voltage rules. This means
that, when using voltage rules the software agents are able to come up with
rules,which solve the common pool resource problem, quickly and reliably. In
the second and third time interval every software agent is able to charge its
electric vehicle suciently.
It should be noted that it looks as if there is a shift from time based rules
towards voltage based rules. But this is only due to the change in the scale of
the y-axis, which is caused by the voltage rules being concentrated in one point.
Once a time based rule is selected as an institutional rule, the institutional will
unfortunately stay a time based rule.