The gent software encodes the state defining attributes specific to PV, load, and transmission line, cost parameters, and operating voltage and power limits. It also encodes behavior to coordinate generation-load balance and other operational requirements as well as maintaining security and reliability. Agents communicate using TCP/IP protocol on the basis of a message-oriented and content independent protocol and open-source libraries that comply with the standards of the Foundation for Intelligent Physical Agents.

Agents operate PV distributed resources, loads, and coordinate to ensure balance and stability. PV agents monitor, communicate, and control generation. The inputs to the agent are PV active power, generation of the diesel generators, power flow at the PCC connecting to the main grid power, and status of circuit breakers. The output of the agent is two set-points for the diesel generators and four commands for the PV and grid circuit breakers. Load agents monitor, communicate, and control loads by shedding PV available generation in an islanded mode. The inputs are load active power (block A, B, C, D, E, and F) and the status of circuit breakers. The outputs are shedding commands. A simulated model of communication between MATLAB-Simulink and JADE agents are shown in Figure 1.

PV agents operate PV distributed resources while coordinating with other agents to ensure balance and stability of the microgrid system. Figure 2 presents the

algorithmic sequence of operation of PV agents. In a grid connected mode, PV agents actively monitor the status of faults and enable PV units to produce power to meet demand and operate a two-way power flow with the utility grid. In an islanded mode, PV agents coordinate with diesel generators to produce power in case of PV fault or shortage of power supply. PV agents coordinate with MG agents to maintain stability and reliability. They provide commands for the set points of diesel generators (DG) by implementing a simple linear control of a lower limit of 0% and upper limit of 70%, with an increasing rate of 7/3000, formally defined as follows:

f ( x ) = { 0 x < 50 7 3000 x 50 ≤ x < 450 0.7 x > 450 (1)

where f(x) denotes the set-point percentage at DG 5 of the total capacity and x denotes power produced at DG 6 in kW. At an upper limit of 70% of DG 5 capacity (30% is left for a safety margin), it starts negotiating with load agents to reduce power and prevent MG failure.

Load agents coordinate to provide enough power at load points. In a grid-connected mode, the shortage of power is supplied by the utility grid. In an islanded mode, a scheduling algorithm is implemented in the load agent software as shown in Figure 2. The load agent sheds load (PShed) based on a priority sequence. For GJU case study, critical loads are in blocks B and D. During the shedding process, load agents coordinate with PV agents for enough power generation. The algorithm for this coordination is shown in Figure 3. Each time a PV agent approves the request to connect a 100 KW (Pconnect), the load agent executes its algorithm and starts checking the possibility of connecting the load based on the priority associated with other load agents representing blocks C, E, and F.

This scenario simulates the microgrid under normal operating conditions in a

grid-connected mode. All loads get power from both PV local generation and the utility grid. The Summer and Winter PV generation and load profiles that are used for simulation. In Summer, PV generation peak is at 1980 kW and is greater than the load peak at 1400 kW. In this case, diesel generators are generally turned off and PV circuit breakers connecting to the main grid are on for excess power flow. In Winter, PV generation peak is at 380 kW and less that load peak at 1200 kW. In Summer, diesel generators are turned off and PV circuit breakers connecting to the main grid are on to facilitate power flow. The PV and PCC have on-status and the set points at DGs are 0. There is no load shedding at buildings. As a result, the net power metering at the PCC is negative indicating excess of PV generation at the PCC with main grid and at each of the building blocks B, D, E and F as depicted in Figure 4. The net power metering at block C is positive indicated that there is no excess of power and its power shortage is compensated by the MG. In contrast, for Winter, the diesel generators are turned on and the PV circuit breakers connected to the main grid are on. The

net metering at the PCC becomes positive, which indicates a shortage of PV generation at the building sites. As a result, the power flows from the main grid to the MG.

This scenario simulates the microgrid in an islanded mode for Summer and Winter PV generation and load profiles. The PV circuit breakers connecting to the main grid are off. Diesel generators are turned on to compensate for power shortage when needed. In Summer, the MG load agents run a shedding schedule to maintain generation, load balance. When the total generation reaches above 70% of the total capacity, the PV agents propose a load reduction from the load agents, which in return accepts reduction as long as critical loads are met. Upon acceptance, the load agents begin shedding loads following the algorithm depicted in Figure 3. Power shedding starts in conjunction with maxing out of the diesel generator at 70%. Load agents keep track of shed load for each block. The load shedding implementation is done by numerically subtracting power demand from available power for three building with less priority. The net power metering of an islanded mode for a summer generation and load profiles is shown in Figure 5. The accumulative power of block F is always negative as it has the least priority. The negative power indicates DG producing power at the building block. Load shedding is also evident in block E that it is kept at zero during the afternoon. When instant power is shed from block F and E, the load at block C is reduced. Since the utility grid is disconnected, the power flow at PCC is always zero. Small amounts of power indicate power loss at MG feeders and transmission lines.

In Winter, a similar approach is implemented. The only difference is that the amount of load shedding in Winter is naturally smaller. This enables PV agents to begin coordination using the scheduling algorithm described in Section 3.2. Figure 6 shows the net power metering at PCC with main grid and at each building block (B, C, D, E and F). The amount of Pshed and PConnect play an important role in the number of successful collaborations between agents. A higher value could result in stability issues and lead to decrease the chance of accepting the proposal, for both shedding and reconnecting. While a small value would not be appropriate and leads to increasing agent communication. To avoid communication overhead, a small value (e.g. 20 kW) is chosen to ensure smooth load reduction while taking into account generation capacity. A larger value was found to increase spikes in DG outputs as shown in Figure 7, leading to unstable MG operation. DGs have enough capacity to meet instantaneous critical loads.

This scenario simulates how agents coordinate to handle faults at MG. A single phase-to-ground fault is introduced at the main feeder connecting the utility grid on phase B. Since the fault takes place at the upstream grid, MG immediately disconnects from the utility grid and shifts its operation to islanding mode [18]. A fault incident is time critical and requires agents’ immediate response by disconnecting MG directly after receiving a fault signal. Agents communicate an emergency to switch to an islanded mode. This fault scenario is studied for both Summer and Winter PV generation and load profiles.

In Summer, a fault is simulated during off peak time (non-working hours), and hence MG operates at its minimum load. Figure 8 shows PV agents disconnecting and switching to DG operation during a fault incident. PV disconnection is necessary to prevent any damage that might occur to the inverter or the

lack of adequate inertia in the microgrid to accommodate the amount of installed PV. As can be seen from the Figure, agents briefly misinterpreted the large disturbance of the fault and started to shed the load, but it was shortly reconnected back as the instantaneous load is seen small enough to be fully supplied by the diesel generators. Figure 9 shows net power metering at PCC with main grid and at each building block (B, C, D, E and F). The islanding operation is indicated by the time during which the grid power is zero. Afterwards, PV agents coordinate to reconnect PV generation and the utility grid when the fault is cleared, and the microgrid resumes its normal operation for the rest of the day. In Winter, a fault is simulated during a peak time (working hours) and hence MG operates at its maximum load. PV agents collaborate to meet load and to keep MG stable by switching to DG and disconnecting PV. PV agents negotiate with load agents to reduce load when the DG power production goes above 70% of the total capacity. Load agents reduce their load to half while keeping the critical loads connected. As in Summer, the status of DG, PV, and PCC, load shedding, and DG power output exhibit the same behavior during a fault incident.

Figure 10 shows the voltage fluctuation during 0.012 second of the fault incident when transiting to the islanded mode. PV agents successfully operate DG in block E to follow the temporary slack generator in block F. It is obvious that the output power mirrors the set-point graph generated by PV agents. This fast response enables the system to stay stable and to smoothly transfer the microgrid from grid-connected to islanding mode and vice versa. The entire latency between the moment the signal is distorted when the fault takes place and the recovery of

the signal when the microgrid stabilizes is 0.012 s. The hardware implementation of agents plays an important role in the response time to stabilize MG and restore to normal operation. The agent behavior clearly shows that the multi agent system is a viable solution to operate MG in real-time.