A deterministic mixed integer linear programming formulation 64

In this section, we will describe a deterministic MILP formulation for the decen-tralized charging scenario with V2G in which random events are replaced by their expected values ωh = λh, β_h

Y1⁺= 0; (3.25) Equation (3.24) is the power balance constraint, where PL

l=1CP×z_{h−l⁻ _+1}

>0 rep-resents the electricity discharged from PHEVs at time h. At time h, the PHEVs whose discharging starts at time h are in the first hour of its discharging cycle; while those starting discharging at time h−L + 1 are in the last hour. (3.25) – (3.30) are associated with PHEV charging and similar to (3.3) – (3.8), explained in Section 3.1.1. The only difference is that with V2G, the PHEVs whose discharging starts at time h−L and completes at time h (that is new discharging completions z_{h−L}⁻

>0) need to be counted towards the number of empty batteries in (3.26) and (3.29).

Equation (3.31) and (3.32) define the transition functions for PHEV inventory Y_h⁻ (fully charged batteries ready to be discharged to send power back to the grid), where z_{h−L}⁺

>0 represents the PHEVs whose charging cycle completes at time h. Equation

(3.33) and (3.34) define the decentralized discharging policy as opposed to the cen-tralized charging, ensuring that either all or none of the vehicles will be discharged

at any time. Equation (3.35) and (3.36) enforce that at and after timeH−2×L+ 2, no PHEVs are discharged, otherwise its charging cycle cannot be finished by the end of a day.

3.2.2 An approximate dynamic programming formulation

For stochastic cases, an approximate dynamic programming-based algorithm is designed to find a near-optimal policy for making decisions. We assume linear ap-proximation for the value function around a post-decision PHEV backlog y^+,x_h and that around a post-decision PHEV inventory (full batteries) y_h^−,x, and update asso-ciated value function slope approximation ¯V_h⁺ and ¯V_h⁻ using an iterative updating operation. The initial values for all value function gradient approximations are 0.

Starting from iteration n= 2, at each time h, to obtain an optimal charging decision k_h^+,n and discharging decision k_h^−,n, given a specific state S_hⁿ, we solve ISO’s hour-ahead economic dispatch problem with charging and discharging decisions made by individual consumers as a mixed integer linear programming problem. Since the ex-act value function of being in post-decision statey^+,x_h and y_h^−,x is unknown, the value function slope approximation computed in the previous iteration n−1, ¯V_h^+,n−1 and V¯_h^−,n−1, are used to make decisions. The objective of the MILP is given as follows

maxxh

z_h⁺≤bigM ×k⁺_h, 1≤h≤H−L; (3.42) It is important to recognize that solving the above MILP is equivalent to solving four linear programs, by setting k⁺_h, k_h⁻

to be equal to (0,0), (0,1), (1,0) or (1,1) and finding the solution which yields the greatest objective value.

Once an optimal charging decision k_h^+,n and an optimal discharging decisionk^−,n_h are determined, and a particular realization of new information on the number of new PHEV arrivals at time h, λⁿ_h, becomes available to the system, the following post-decision transition functions are used to move forward to the next pre-decision states at timeh+ 1:

1) the post-decision transition function for empty batteries plugged in and waiting to be charged Y_h⁺+1^,n, written as

2) the post-decision transition function for fully charged batteries ready to be dischargedY_h^−,n+1, given by

wherez^−,n_{h−L}

>0 represents the number of new discharging completions at time h, and

z_{h−L}⁺ _>0 denotes the number of new charging completions at time h.

Our algorithm proceeds till the end of a day to finish iteration n. At the end of the iteration, for each time h, we solve a real-time economic dispatch to obtain a new estimate of the wholesale electricity price,pⁿ_h, and use it to update the wholesale electricity price approximation. The best estimates for wholesale electricity prices calculated so far,P_hⁿ, 1≤h≤H, are then used to obtain a new estimate of marginal value of increasing the post-decision PHEV backlog by one unit, v_h^+,n, and a new estimate of marginal value of increasing the post-decision PHEV inventory by one unit, v_h^−,n. Using v⁺_h^,n and v_h^−,n, we update the post-decision states’ value function gradient approximations to obtain their best estimate so far, ¯V_h^+,n and ¯V_h^−,n. The same procedure is repeated for a number of iterations to return a good policy. The details on how to obtain ¯V_h^+,n and ¯P_hⁿ are the same as that presented in Section 2.3.

In the rest of this section, we will explain how to generate a new estimate of marginal value of increasing the post-decision PHEV inventory at time h by one unit, v^−,n_h , and use it to update the value function slope approximation.

Figure 3.1 illustrates how we obtain a new estimate of marginal value of increasing fully charged batteries at time h, y_h^−,x, by one unit (in thousand), given wholesale electricity price approximations: ¯P_τⁿ, 1 ≤ τ ≤ H. If we discharge one more unit of batteries at time h, post-decision PHEV inventory y^−,x_h will decrease by one unit.

By doing this, two things will happen in the future hours. First, for the following L− 1 hours, CP [kW] of electricity generation would be provided by discharging the PHEVs. CP represents the discharging power rate. The associated savings on electricity generation costs are equal to

h+L−1

Fig. 3.1. Illustrating how to generation a new estimate of marginal value of increasing PHEV inventory by one unit, given wholesale elec-tricity price approximations

Note that the one unit of batteries would become empty at time h+L. The second thing that will happen to the following hours is that we would need to recharge it to its full electricity capacity (at the lowest costs), since it is assumed that all PHEVs need to be fully charged by the end of a daily cycle. The lowest costs to recharge the one unit of batteries can be computed by solving a trivial optimization problem of finding the optimal charging start time to minimize the associated electricity costs in a full charging cycle that lasts for Lhours. The optimization problem can be written as follows

h+L≤τ≤H−L+1min

L

X

l=1

CP ×P¯_τ+l−ⁿ 1. (3.53)

The marginal value of decreasing fully charged batteries by one unit can be estimated by the net reduction on electricity generation costs expected to make, according to

L The marginal value of increasingy_h^−,xby one unit,v^−,n_h , can be estimated using exactly the opposite of what is calculated in (3.54), written as

v_h^−,n = min From (3.55) we can see that when electricity price in the immediate future is low, gains from increasing fully charged batteries (or storage resources) will be relatively large, meaning that more vehicles should be charged to store energy when electricity price is low and sell it back to the grid later when electricity price is relatively high.

By far we have presented our approximate dynamic programming-based models and algorithms for three PHEV charging scenarios. The numerical results for these three schemes will be presented and analyzed in the next section.