s entities in terms of status and provided services by in real time to events and stimuli that arrive from other devices attributes and methods. By a bottom up approach abstract or from the grid: a scattered network of sensors, actuators, entities (objects) are identified as parts of the system. Their communication nodes, systems control and monitoring. In fact properties and the interrelationships between computer pro- the power grid today is no longer designed as a simple network grams are designed by making them out of objects that interact that delivers energy according to one direction: from few large with one another. power generation to many small consumption points at the end Despite of these similarities OOP just breaks software and users. information into functional units. This programming paradigm In particular we address the problem of exploiting the does not deal with providing smartness to objects for enabling increasing availability of distributed renewable energy sources intelligent interaction models [2]. The building block of the such as photo-voltaic (PV) panels to improve the energy IoT is the smart object and the novelty is the pervasive efficiency in a neighborhood. deployment of such embedded systems connected to the In fact, the increasing decentralized production of green Internet, interacting with one another. In fact smart objects energy by photo-voltaic panels has changed the classical are mostly fully functional on their own, but value is added model of power supply from the grid to the households.
Nowadays each house becomes a micro-grid connected to a linear combination or regression procedure that can the network where energy producers and consumer exchange include several similar loads. energy between themselves and with the grid as shown in • Regression methods. Regression is the one of most widely Figure 1. used statistical techniques. For electric load forecasting
As we all know, the electricity demand varies during the regression methods are usually used to model the reladay and over the seasons. The electrical current is designed tionship of load consumption and other factors such as to withstand the maximum level of demand, and then in the weather, day type, and customer class. absence of constant peaks turns out to be over-sized and • Time series. Time series methods are based on the asunderutilized. sumption that the data have an internal structure, such as
The main challenge here is to provide intelligence to autocorrelation, trend, or seasonal variation. Time series
consuming and producing devices like photo-voltaic panels, forecasting methods detect and explore such a structure.
energy storages, appliances, sensors and actuators to let them In this paper we use a similar-load approach for short-term
collaborate in order to agree on the best schedule of energy learning.
consumptions according to monitoring information and pre- Our contribution, and in particular the CoSSMic project
dictions. [
The object oriented paradigm appears to be very useful to negotiation among users’ devices on real power grids, that to represent and manage this type of problem but it is not enough the authors’ knowledge has not been implemented before. The because we need to provide the devices with all advanced framework will be validated on real infrastructures by trials capabilities to collaborate as intelligent actors of the IoT. that involve inhabitants of two different European countries
Intelligent agents will act on behalf of devices learning (Germany and Italy). Both software and hardware will be
energy requirements and predicting energy availability. The integrated and customized ad hoc to be compliant with existing
energy profiles will be then used to negotiate energy exchanges installations. In [
In this paper we present the learning-based approach for the Market that supports the energy negotiation based on XMPP1
distributed energy market. The main issue we focus on is a protocol. Agents can make calls for proposals, accept offers
reasonable compromise between the resolution of the profile and negotiate with other agents. Additional details about how
representation and the performance and real time requirements network of agents have been exploited in other applications
of the system. We discuss some experimental results obtained domains are described in [
We focus here on a learning approach for short-term load forecasting to estimate load flows and to make decisions about task scheduling. The learning approach means each consuming device behave like an adaptive, intelligent agent, gradually adapting to its environment, and gaining more confidence in its predictions. This has the advantage that great part of the smartgrid configuration is leveraged by agents. The big disadvantage is that the agent has to learn from scratch, which means it might take a some time before the agent can make accurate decisions.
Several studies have been conducted regarding the
application of multi-agent system in the energy management
and negotiation. Paper [
Authors derive their methodology empirically by collecting
data from a large variety of loads and showing the significant
commonalities between them. A large variety of statistical and
artificial intelligence techniques have been developed for
shortterm load forecasting [
FP7-SMART CITIES, 2013) is an ICT European project that aims at fostering a higher rate of self-consumption of decentralised renewable energy production by innovative autonomic systems for the management and control of power micro-grids on users’ behalf. This will allow households to optimise consumption and power sales to the network by • Similar-load approach. This approach is based on searching historical data about loads to identify similar characteristics to predict the next load. The forecast can be
1Extensible Messaging and Presence Protocol Fig. 2. CoSSMic Platform - Components View a collaborative strategy within a neighborhood. In fact the increasing of the decentralized production of green energy is affecting the current energy management scheme both at the grid and at the user level. The amount of electricity used and the energy produced by photovoltaic systems varies over the course of the day, from season to season and depending on the weather conditions. On the other hand the users may not have the opportunity to turn on their appliances when the PV is producing because they do not know when or because they are outside or simply because they do not need those appliances at that time. For this reason there is not an alignment during the time between production and consumption. At the user level this limits both the optimal utilization of green energy and affects the budget spent.
The CoSSMic project proposes an intelligent scheduling of user’s consumptions within a micro-grid. A micro-grid is typically confined to a smart home or an office building, and embeds local generation and storage of solar power, and a number of power consuming devices. The main goal is to enable the collaboration among households so that the energy produced by the PVs is consumed by any available consumer in the neighborhood. An autonomic system will able to shift the loads in each neighborhood, according to users’ constraints and preferences, in order to find an optimal match between consumption and production during the day, so that the use of renewable energy is maximized. Appliances (refrigerators, washing machine, drier and dishwasher, water heaters, air conditioners) equipped with intelligent controllers are represented by software agents negotiating energy exchanges according to availability, demand, user’s preferences and constraints.
The energy negotiation between agents will be based on rewards for local producers. To obtain the maximization of self consumption, a predictive approach can be considered, but we need to predict PV production based on the weather forecast and the parameters of the plant. Moreover the power consumed by each device must be known in advance or learned from monitoring information. Also the behaviour of consuming devices can change each day according to a number of parameters (e.g. the external temperature for the refrigerator or the air conditioner, the user’s needs for the electric cars or for the oven), which are often unpredictable. schedule of local devices. In fact the allocation of energy to devices is modeled as task schedule with energy and time constraints.
The algorithm to find the best scheduling of the consuming appliances is designed and implemented as a distributed negotiation among software agents.
Agents are classified according to two categories: • Consumers: they buy energy for consuming devices. E.g.
they will run in houses to manage objects that absorb energy: electric car, computers, ovens, washing machines, etc. • Producers: they can sell energy. In this category there are, for example, power generators, solar panels, wind turbines.
Those devices, which are able both to produce and consume energy such as energy storages or electric cars, will be represented by a couple of agents belonging to the two different classes defined above.
For the energy negotiation there is not a concrete marketplace, but a virtual market is implemented by a negotiation protocol that uses P2P overlay of agents.
The user will define preferences and constraints about the utilization of his appliances. This policy will be used by the smart devices to find the best plan that maximize the energy self-consumption of the neighborhood. Besides the
IV. OVERVIEW OF THE COSSMIC SOLUTION user’s constraint and preferences each agent should know the
Figure 2 shows the main components of our architecture energy/power profile of its device.
playing the CoSSMic scenario. A detailed description is pro- A requirement for producer agents is the knowledge of the
vided in [
The Graphical User Interface (GUI) supports interactive much energy will be produced by PV panels is computed using control and configuration of the system. It allows to plan the weather forecast, properties of the PV plants and historical usage of appliances defining constraints and preferences. It series. also provides to the user real time monitoring information, A requirement for the consumer agents is the knowledge statistics on historical data and predictions. of the consuming profile of the managed device. Of course
Mediator Services are used for storing and management such a profile will depend on different parameters. In the case of smart grid information. Mediator services are accessed by of a washing machine the energy consumed could depend device drivers to store measures and by agents to collect on the operation mode, on the amount of clothes, but in the information about energy production and consumption of the case of an air conditioner or of a freezer the temperature is a household. The same services are used to save data about the relevant feature to take into account. In this case monitoring information will be used to learn the energy profile while the CoSSMic platform is running.
Devices in the home can send information about electricity consumption through wireless interfaces (for example UHF or Zigbee). Mobile devices (e.g. electric cars), instead, send information through the CoSSMic Cloud. In both cases the information, through the Mediator, reach the agent platform.
The mediator integrates also a number of device drivers, which allow to send real time commands to electric devices in the smart house. For example, through device drivers the mediator APIs allow to switch on or switch off devices, when it does not violates any constraints, in order to save energy.
A. Energy profiles
In CoSSMic optimization of task scheduling use device profiles. Profiles include some meta-data and time series, i.e. series of time value pairs, which describe the cumulative energy consumed or produced when the device is running.
Each device may have more profiles. In fact, as we said before dishwashers and washing machines typically have different operation modes, and in that case there will need one device profile for each mode.
Static or synthetic profile can be used just for the first run.
However profiles may change run by run, therefore dynamic profiling is needed. In CoSSMic a learning approach is used to improve dynamically the device profile in real time. For instance, the consumption profile of a heater depends on the ambient temperature. Initial measurements for a certain ambient temperature can be used for the first run. However using monitoring information within the current environment and for the current operation mode will be exploited to update the profile dynamically and to improve the prediction for the next schedule. is below a threshold (e.g. 0.01 kW). In fact in this case we suppose that the washing machine is off and the run is B. Learning model terminated. A new energy increment above the threshold will
Before the start of the trials and until the software is not correspond to the starting time of the next run. available, we used monitoring information to test learning In a second steps we clustered the different runs according algorithm and to evaluate the proposed approach. to their duration and the amount of consumed energy to
In particular during the trial the user will program the device identify different operation modes. In Figure 3 the cumulative by the CoSSMic interface and making the system aware about energy consumption of different runs of a washing machine the starting time and the operation mode on the next run. In are shown. The time series run9 and run13 correspond to the the following example we used raw energy time-series, which same operation mode. have been collected using a PG&E Landys+Gyr smart meter. The problem is the best approximation of the profile using
The problem here is that time-series include different work- two or more time-series by a compact representation that does ing modes which are different for profile and duration and it not affect performance and real time requirements. is necessary to identify the different runs. Another problem In Figure 4 the blue points and the red crosses are the is that the samples come at a different rate, according to the samples from run9 and run13. The black line represent the energy consumed. To identify different runs of a time series polynomial 3rd degree polynomial curve that fits the points of a washing machine, we split the energy time-series using a with a minimal square root error. This representation will supervised approach. introduce of course a residual error, but it is monotone and
Different runs are recognized by looking at sequence of needs only 5 float parameters to be represented (4 coefficients, values whose variation for a certain period (e.g. 10 minutes) duration and residual error), independently from the amount raw samples.
In order to improve the resolution of our representation we also considered the utilization of a b-spline. The b-spline is a piecewise polynomial function of degree k. Polynomial curves meet and are continuous in a number of control points. We can control the resolution increasing the number of control points.
In Figure 5 the black line is a b-spline representation of degree k=3 of the best fitting of run9 and run13. In this case 5 control points have been fixed at regular time intervals and the number of float parameters to be transmitted is 27.
A consumer that wants to schedule his consuming tasks will execute the following steps: • Connects and login to the XMPP server to join the
neighbourhood. • Estimates its own need of energy. • Brokers the producer that can offer the required energy • Send a proposal (SLA template)
On the other hand, the producer will: VI. SLA BASED ENERGY NEGOTIATION • Connect and login to the XMPP server
The main objective of a negotiation is to reach an agreement • Wait for incoming proposals between a vendor and a customer. A Service Level Agreement • Evaluate the proposal accepting or not (SLA) defines an agreement between a provider and a client • Send the related response in the context of a particular service provision and can be Consumers will adopt a ranking mechanism to broker the between two (one-to-one) or more (one-to-many or many-to- producer to be contacted. Every consumer associates a rank many) parties. to each producer, that is an integer that indicates the quality
In our scenario an energy consumer and an energy producer of the producer. Each time the producer accepts the proposal, agree to shift the energy workloads of a device to optimise the this number is incremented by 1. The consumer will continue mapping between production and consumption, but within a to call the producer which has the highest ranking. time period defined by the earliest start time and the latest start Each producer will try to allocate the consuming workload time defined by the user. For this reason the emergent behavior according to the model presented in the previous section of whole multi agents system will implement a distributed choosing the start time for the incoming task that satisfies scheduler that allocates consuming and producing tasks for the user requirements, but minimizing the use of energy from each device within the neighbourhood. grid compared to its own production profile.
The negotiation protocol will be implemented by an overlay The negotiation protocol is started by a consumer agent each of agents which exchange FIPA2 messages using XMPP as time a new execution is planned for the handled device sending transport layer. The XMPP server is used for authentication a proposal. The message body of a proposal is a machine and to support communication across firewalls. Moreover readable SLA template. many concepts of the XMPP protocol has been re-used (friend- Templates used by agents for negotiation define the energy ship, presence, multi-user chat, etc..). Alternative server-less requirements, by the profile discussed above, including user’s solution will be investigated in future works. preferences and constraints (e.g. start date, termination date, cost, etc.). The SLA will complement the energy requirements 2Foundation Intelligent Physical Agents with the negotiation parties and actual start time.
of production change and the negotiation will be restarted by the consumer.
Of course the algorithm may be affected by the problem of local minima that may bring to a not optimal employment of available energy. The final formulation of the optimization algorithm is not available yet, but it is related to a distributed approach where each agent as limited knowledge of the system. We will accept a sub-optimal solution that can be obtained with limited processing and communication resources.
performances of SPADE and JADE agent platforms working with different XMPP implementations. although the difference with JADE on OpenFire and SPADE
SPADE [
JADE is an agent platform fully implemented in Java lan- with a new agent that makes 1.000 floating point operations
guage. It simplifies the implementation of multi-agent systems to simulate the optimization algorithm. We have one producer
through a middle-ware that complies with the FIPA specifica- that always accepts requests and 50 consumers.
tions. It also provides a set of graphical tools that supports the
debugging and deployment phases [
We experimented three different XMPP server
technologies:Tigase, OpenFire and the light implementation provided From figure 8 we noticed that in JADE the time to complete
by SPADE itself. Tigase [
The testbed is one host equipped with a 2.67 GHz i5 use Raspberry Pi to host our software, our choice is SPADE. processor, 4 GB of memory amount and Windows 7 operating Because of the limited amount of resources required by its system. Here we evaluate the performance running a single XMPP server we have also decided to host on the Raspberry consumer and a single producer that always accepts negoti- Pi the SPADE XMPP server. Server to server connection will ation requests. Message exchanged have an empty payload be used for the communication between agents executing in because the purpose is to evaluate the technology. different households.
In Figure 7 the chart shows the average time to close In the last experiment we used two RaspberryPi as testbed. It a transaction. It is clear that with one consumer and one hosts 10 consumers and 1 producer to simulate an household producer, SPADE on OpenFire exhibits the best performance with 10 passive devices and 1 solar panel. Each device is represented by one SPADE agent. Here we evaluate how the resolution of the profile will affect the performance.
Consumer agents chose randomly the producer to be contacted only the first time whereas for the next time each agent, using a learning mechanism, will contact the most reliable producer. In the experiments to implement the learning capacity we use a ranking mechanism. Each agent creates a vector with all active producers and a ranking is associated to each producer. The ranking is an integer initialized to zero incremented by one each time the corresponding producer accepts the proposal. The consumer will contact the producer that has the maximum ranking and it will continue to contact the same producer until a proposal is refused.
In order to simulate a real scenario we introduce a delay between a request and the subsequent that follows a Poisson distribution with an interarrival of 500 ms.
The request message from consumer is divided in two sections: in the first one there are metadata information, in the second one there is the energy profile.
In listing 1 there is a message example where the meaning of the fields is: milliseconds in the case of 10 consumers and 1 producer for each Raspberry, using a delay between requests and adopting a ranking protocol for brokering.
The real parameters represent the coefficients of a 3th This work has been supported by CoSSMic (Collaborating degree polynomial curve. For lake of space we do not show Smart Solar powered Micro grids - FP7 SMARTCITIES 2013 the results when the raw profile is sent, but we can see that - Project ID: 608806). such solution is not feasible with the available resource. The response message contains the Actual Start Time (AST) for REFERENCES the current run, i.e. the time when the producer can provide [1] L. Atzori, A. Iera, and G. Morabito, “The internet of things: A survey,” energy. Comput. Netw., vol. 54, no. 15, pp. 2787–2805, Oct. 2010.
We observed that the mean time to obtain the response [2] Lw.itBhenoebdjeiccetsn.”ti, A“RCeMthinSkIGinSgOsFmTartSoobftjwecatrse: bEunilgdiinnegerairntigfiNcioatleisn,tevlloilg.en2c5e, is of about 5.3 seconds with a standard deviation of 676 no. 3, p. 59, 2000.