ChanceConstrained ShrunkenPrimalDual Subgradient (CCSPDS) Approach for Decentralized Electric Vehicle Charging Control
Abstract
In this paper, we develop a chanceconstrained decentralized electric vehicle (EV) charging control framework to achieve “valleyfilling” meanwhile meeting individual charging requirements and satisfying distribution network constraints. The control design is formulated as an optimization problem with a stochastic nonseparable objective function and globally coupled chance constraints. We propose a novel chanceconstrained shrunkenprimaldual subgradient (CCSPDS) algorithm to support the chanceconstrained decentralized control scheme and verify its efficacy and convergence with a representative distribution network model.
I Introduction
Electric vehicles (EVs) have shown their effectiveness in increasing energy conversion efficiency and reducing GHG emissions [1]. Besides, with proper control, the charging process of a large fleet of EVs can be leveraged for provisioning gridlevel services [2]. Extensive studies have been conducted on the modeling and control problems of EVs for “valleyfilling” [3], load balancing [4], and frequency regulation [5]. The potential of EV charging control for facilitating renewable energy integration and providing additional services is discussed in [6].
Apart from the benefits, charging a large fleet of EVs without proper control would have negative impacts on distribution networks. Lopes et al. [7] surveyed the challenges of integrating EVs into the mid and lowvoltage distribution networks, including voltage sag, transformer overloading, network congestion, and increased power loss. The negative impacts caused by nonnetworkaware controlled charging were verified in [8]. Hence, an impactfree EV integration necessitates designing EV charging control in the context of distribution network constraints.
Recently, an increasing amount of attention is being paid to controlling EV charging for distribution network impact alleviation and grid service provisioning. Considering network impacts, Richardson et al. [9] optimized EV charging profiles to minimize the total power consumption. In [10], a realtime control for power loss minimization was studied. Authors in [11] and [12] separately developed decentralized EV charging controllers to achieve valleyfilling under voltage constraints. Studies in a large amount of literature have shown that one of the most effective services EVs could provide is to fill the load valley at night [13]. In our paper, we target at valleyfilling meanwhile meeting both local charging needs and distribution network constraints on voltage.
Most existing EV charging control algorithms assume a powerful central controller that can handle large computational loads, [14, 9, 10, 15]. As the EV deployment in the distribution network keeps increasing, centralized approaches however do not scale well and become unrealistic. In this paper, we aim to design an optimizationbased decentralized control scheme that does not require a communication network among EV chargers. Achieving this is notably challenging, because firstly the valleyfilling objective constructs a coupled and nonseparable function, and secondly the network voltage constraints strongly couple the individual charging powers in form of linear inequalities [12]. For this type of problem, Koshal et al. [16] developed a regularized primaldual subgradient approach, however convergence errors exist due to the regularization of the Lagrangian. In [11], the authors proposed an Alternating Direction Method of Multipliers (ADMM)based decentralized algorithm. Though proved to be practically applicable, the required twolayer communication network complicates the communication and poses computing burdens to all buses. Additionally, ADMMbased algorithms often face the problem of a large number of iterations. In our previous work [12], we proposed a shrunkenprimaldual subgradient (SPDS) approach that can eliminate convergence errors, reduce the number of iterations, and alleviate communication loads.
Unfortunately, the aforementioned algorithms are not capable of handling uncertainties in EV charging (e.g., random baseline load and random customer’s behaviors). Among the limited literature, Liu et al. [17] proposed an SPDSbased eventtriggering decentralized control algorithm to handle EV drivers’ random arrivals and departures, where customers’ baseline loads are assumed perfectly predicted. Hassan et al. [18] developed a chanceconstrained ADMM for decentralized distributed energy resources control considering random solar generation. However, decentralized implementation of chance constraints on voltage was treated deterministically, leaving that problem open. Handling chance constraints on voltage in a decentralized fashion is notably challenging because: (i) Voltage constraints are strongly coupled; (ii) Gradient of the chance constraint function is illconditioned, leading to zero updates in iterationbased algorithms. In this paper, we aim to leverage the outstanding convergency, optimality, and speed of SPDS, to design a chanceconstrained SPDS (CCSPDS) to handle decentralized EV charging control problems that have chance constraints on nodal voltages. To the author’s best knowledge, this has never been attempted before.
The main contribution of this paper is twofold. First, this is the first paper studying decentralized EV charging control for provisioning valleyfilling under chance constraints due to uncertain baseline loads. Second, a novel CCSPDS algorithm is developed to solve a class of optimization problems consisting of nonseparable objective functions and strongly coupled network chance constraints.
Ii Preliminaries & Problem formulation
Iia EV charging model
For a radial distribution network with nodes, let denote the number of EVs connected at the th node and denote the total number of EVs. The individual EV charging dynamics can be represented as [17]
where the subscript denotes the th EV connected at the th node, is the discretetime index, denotes the energy remaining to be charged, , is the charging efficiency, is the sampling interval, is the maximum charging power, and is the control signal.
Let and denote individual plugin time and designated charging deadlines; let denote the valleyfilling service period. In this paper, we assume that individual charging periods are contracted to fully cover the valleyfilling period. This assumption can be easily relaxed. The charging efficiencies, initial state of charge, maximum charging powers, and battery capacities are heterogeneous. Augmenting all connected EVs, we have the system dynamics represented as
(1) 
where , , and is the th column of the diagonal matrix , .
Augmenting the system state and individual control signal in (1) along the valleyfilling period , we have
Herein, , , is the system state at time predicted at time ; , , is the control signal at time predicted at time .
To meet all drivers’ charging requirements, the energy remaining to be charged at the end of valleyfilling, i.e., the th vector block in , must satisfy
(2) 
where .
IiB Distribution network model
In this paper, we consider radial distribution networks. Let denote the set of downstream nodes and let denote the set of all downstream line segments. The feeder head decouples interactions in the downstream distribution system from the rest of the grid and maintains its own voltage magnitude . To better illustrate the EV charging impacts and how controlled EV charging can help alleviate the impacts, we do not consider any distributed energy resources (DERs), voltage regulations, or reactive power supplies.
At time , let denote the voltage magnitude at Node ; let and denote the real and reactive power consumption at Node ; and with a slight abuse of notations, let denote the impedance of the line segment . By omitting line losses, the LinDistFlow model of this distribution network is derived as [19, 20, 12]
(3) 
where
and
where () is the set containing downstream line segments connecting the feeder head and Node (Node ) [20].
At each node and each house, real and reactive loads can be separated into baseline loads and EV charging loads, indicating
where and denote the baseline real and reactive power consumption, respectively, and and denote the real and reactive EV charging power, respectively. As it is unrealistic to assume perfect forecast of the baseline load, a realistic problem formulation must consider and , where and denote the means of the normal distributions, and and denote the standard deviations of the normal distributions. In this paper, we make the following three assumptions:

At each time , power consumption of all houses follows independent normal distributions. The standard derivation is assumed to be a constant at all time.

EVs do not consume reactive power, i.e., .

Power factor at each house is a constant, which implies .
Remark 1: In A1, other distributions can also be used, but normal distribution can better reflect customers’ random behaviors. Constant standard deviation is assumed for the simplicity of presentation. A2 follows as no fourthquadrant charger is considered. A3 follows by assuming power factor correction equipment is available at each node.
Let contain baseline power at all houses, then , where and . Further let denote the nodal aggregated baseline power, where , , is the nodal aggregation vector. Since (3) is linear, the voltage drop caused by the baseline load can be represented as
where , yielding
where , , and with and .
Let denote , denote , and denote , we have
where . The system output augmented along the valleyfilling period is written as
where
and
Herein, denotes the direct matrix sum.
IiC Deterministic valleyfilling
Suppose is deterministic or can be accurately forecasted. According to [12, 17], optimal EV charging sequences for valleyfilling under nodal voltage constraints are obtained by solving
(4)  
s.t.  
where the time index is dropped for simplicity, and
Herein, denotes the aggregated baseline load profile at the feeder head, is the EV charging power aggregation matrix, is for battery state of health protection, and is the lower bound of the transformer service range.
IiD Chanceconstrained valleyfilling
In more realistic situations, where baseline load cannot be accurately forecasted, we have normally distributed . Under this circumstance, the deterministic problem (4) becomes
(5)  
s.t.  
where is the th block of , is the th block of , and is the allowed chance of violating voltage limits. For simplicity of presentation, we use to denote the expectation operator hereinafter. Designing decentralized algorithms for stochastic problem (5) is more complicated than for problem (4) as the coupled objective function becomes an expectation and the coupled inequality constraints become coupled chance constraints.
Iii Chanceconstrained SPDS design
Let the multivariate normal cumulative distribution function (MVNCDF) of be denoted by , then the chance constraints in problem (5) become
(6) 
yielding the unconstrained Lagrangian of problem (5) as
It is proved in [12] that a convex, strongly coupled optimization problem can be efficiently solved by SPDS by following
(7a)  
(7b) 
where is the Euclidean projection, denotes the iteration number, and are the update step sizes for primal and dual variables, respectively, are the shrinking parameters, and is the constraint set for dual variables. The definition of is referred to [12, 17]. It has been shown in [12] that the convergence and optimality are guaranteed if and are appropriately chosen for a designated tuple .
Nevertheless, SPDS in (7) cannot be directly applied to the chanceconstrained problem (5) due to two challenges: (i) The feasible region is nonconvex due to the nonconvexity of ; (ii) The gradient is bounded and in most cases it realizes the value of , implying that updates of the primal variables can barely capture constraints. (This will be illustrated later). In the following, we will present the details of CCSPDS and show via simulations that it can handle the nonconvex chanceconstrained problem (5). Rigorous proof of convergency and optimality gap will be provided in a journal extension of this paper.
The objective function of problem (5) can be rewritten as
and its gradient at can be calculated as
where inside the bracket of the first term on the right hand side is the mean of the total load profile at the th iteration. This is a universal information broadcasted to all EVs in each iteration. To obtain the complete form of we first introduce the following theorem.
Theorem 1[21] Let with some positive definite covariance matrix . Then, the distribution function is continuously differentiable at any and
Here, denotes the 1D normal density of the component , is an D normal random vector distributed according to , results from the vector by deleting component , and results from the matrix by deleting row and column , where refers to column of . Moreover, is positive definite.
By applying Theorem 1, we can write
(8)  
where , denotes the Jacobian matrix, denotes the th component of , and is obtained by removing the th component of .
Though theoretically sound, directly applying in (8) to the primal update equation (7a) is illconditioned. For example in the independent bivariate normal distribution case, the gradient of the cumulative function is .
As illustrated in Figure 1, there are four cases where is extremely small or zero: (i & ii) is far away from to the right and is far away from to the left (and right); (iii & iv) is far away from to the left and is far away from to the left (and right); Translated into (6), these four cases mean: (i & ii) Constraint is satisfied with a high probability; No (or slight) updates w.r.t. the constraint are needed in the iteration; (iii & iv) Constraint is not satisfied with a high probability; Updates w.r.t. the constraint are needed in the iteration. However, in the last two cases, updates cannot be realized as the updating direction (gradient) is . To overcome this illconditioned gradient, we define an adjusted gradient as
where
and , are defined in the same way. The adjusted density function and cumulative function are illustrated in Figure 1 as the dashed curves which only adjust the left half planes. These adjustments guarantee effective updates once chance constraints are violated. Following this idea, we have in (8) adjusted to
(9) 
and consequently the primal update equation (7a) becomes
(10) 
where
The dual update equation (7b) can be readily calculated as
Remark 2: In EV charging control, the adjusted cumulative gradient vector on the right hand side of (9) is broadcasted to all EVs as a public key. Each EV uses its private key to interpret the public key, so that the control sequence update can fully acknowledge the global chance constraints.
Remark 3: CCSDPS, especially the approach of adjusting the gradient of MVNCDFs, can be applied to many other popular distributions, e.g., Poisson distribution and uniform distribution, with and properly defined.
Iv Simulations
The simplified IEEE 13 Node Test Feeder used in [12] is adopted in this paper. In specific, Node 632 and Node 671 are not connected to any load, and each of the other nodes is connected with houses equipped with level2 chargers. Battery capacities are uniformly distributed in kWh. Initial and designated SOCs are uniformly distributed in and , respectively. Primal step size ; dual step size ; in ; shrinking parameters are . The above parameters were empirically chosen to accelerate the convergence speed. Initial values of and are set to zero vectors. The voltage lower bound is chosen to , which is slightly higher than the p.u. in ANSI C84.1 service standard, to compensate for the discarded line losses in LinDistFlow model and possible voltage violations. For the normal distribution of baseline load at each house, the mean is obtained by scaling the data from Southern California Edison [22]; the constant standard deviation W. The contracted valleyfilling service period is from 19:00 to 8:00 next day. The sampling time is 15 minutes. The allowed voltage constraint violation chance is set to .
Figure 2 shows the valleyfilling performance under the proposed CCSPDS decentralized framework, where convergence of the algorithm can be readily revealed.
In total 50 iterations were executed and each iteration took about 0.16 second. Comparing with the purely deterministic case discussed in [12, 23] where the converged total load profile is flat, the converged total load profile in the stochastic case presents slight variations. This is because the objective function only captures the mean of the baseline load profile, i.e., CCSPDS fills the mean baseline load valley, and once random behaviors realized the valleyfilling performance is slightly compromised. Figure 2 also shows the total load profile resulting from directly applying deterministic SPDS to the stochastic case, i.e., solving (4) by using the mean baseline loads in both the objective function and voltage constraints. The difference between CCSPDS and SPDS w.r.t. valleyfilling performance is neglectable, as both of them are using mean baseline load in the objective function.
Figure 3 shows the nodal voltage magnitudes of all nodes during the valleyfilling period.
Under baseline loads, all nodal voltage magnitudes are above p.u. lower bound. Under the proposed CCSPDSbased decentralized EV charging control framework, only few nodes occasionally exceed the bound during 1:00 to 5:00. Under the SPDSbased control, voltage magnitudes are continuously violating the limit during 23:00 to 5:30 and are much lower than those of the CCSPDS control. This observation reveals the effectiveness of CCSPDS.
Figure 4 presents the statistics of voltage violations.
The upper plot reveals that in the total 52 time slots during the valleyfilling period, CCSPDS can suppress 50 (96%) of them to have less than 10% (only one) of all nodes violating 0.954 p.u., and 2 (4%) of them to have 25% (three) of all nodes violating 0.954 p.u. At each time slot, percentage of the total nodes violating the limit can be treated as a measure of the chance constraint. The 25% violation is as expected since the chance constraints only reflect a probability which cannot be fully reflected by a single sample. Comparing with CCSPDS, SPDS has the histogram more distributed to the right, leading to more violation events and greater violation magnitude. In the lower plot, boxes represent the 25th and 75th percentiles and legs with bars represent the extreme values. It can be concluded that under the control of CCSPDS: Nodal voltage magnitudes have less variations; The 25th percentiles are guaranteed to stay above 0.954 p.u.; Number and magnitudes of violations are less than those of SPDS. These observations further reveal the effectiveness of CCSPDS.
V Conclusion
This paper developed a novel decentralized EV charging control framework where distribution network constraints were formulated as chance constraints due to customers’ random behaviors. A new decentralized optimization algorithm – CCSPDS was developed to solve the chanceconstrained valleyfilling problem as well as generic optimization problems that have coupled objective functions and coupled global chance constraints. Convergence and effectiveness of CCSPDS were verified via simulations. CCSPDS can be generalized to other distributions and extended to consider local chance constraints.
References
 [1] C. Liu, K. T. Chau, D. Wu, and S. Gao, “Opportunities and challenges of vehicletohome, vehicletovehicle, and vehicletogrid technologies,” Proceedings of the IEEE, vol. 101, no. 11, pp. 2409–2427, 2013.
 [2] D. Callaway and I. Hiskens, “Achieving controllability of plugin electric vehicles,” Proceedings of the IEEE, vol. 99, no. 1, pp. 184–199, 2011.
 [3] L. Gan, U. Topcu, and S. Low, “Optimal decentralized protocol for electric vehicle charging,” IEEE Transactions on Power Systems, vol. 28, no. 2, pp. 940–951, 2013.
 [4] A. Mercurio, A. Giorgio, and F. Purificato, “Optimal fully electric vehicle load balancing with an ADMM algorithm in smartgrids,” in Proceedings of Mediterranean Conference on Control and Automation, PlataniasChania, Crete, Greece, Jun. 2528 2013, pp. 119–124.
 [5] E. Karfopoulos, K. Panourgias, and N. Hatziargyriou, “Distributed coordination of electric vehicles providing V2G regulation services,” IEEE Transactions on Power Systems, vol. 31, no. 4, pp. 2834–2846, 2016.
 [6] D. Richardson, “Electric vehicles and the electric grid: A review of modeling approaches, impacts, and renewable energy integration,” Renewable & Sustainable Energy Reviews, vol. 19, pp. 247–254, 2013.
 [7] J. Lopes, F. Soares, and P. Almeida, “Integration of electric vehicles in the electric power systems,” Proceedings of the IEEE, vol. 99, no. 1, pp. 168–183, 2011.
 [8] L. F. T. Romn, R. Cossent, C. Domingo, and P. Fras, “Assessment of the imapct of plugin electric vehicles on distribution networks,” IEEE Transactions on Power Systems, vol. 26, no. 1, pp. 206–213, 2011.
 [9] P. Richardson, D. Flynn, and A. Kaene, “Optimal charging of electric vehicles in lowvoltage distribution systems,” IEEE Transactions on Power Systems, vol. 27, no. 1, pp. 268–279, 2012.
 [10] X. Luo and K. Chan, “Realtime scheduling of electric vehicles charging in lowvoltage residential distribution systems to minimise power losses and improve voltage profile,” IET Generation, Transmission & Distribution, vol. 8, no. 3, pp. 516–529, 2013.
 [11] L. Zhang, V. Kekatos, and G. B. Giannakis, “Scalable electric vehicle charging protocols,” IEEE Transactions on Power Systems, vol. 32, no. 2, pp. 1451–1462, 2017.
 [12] M. Liu, P. K. Phanivong, Y. Shi, and D. S. Callaway, “Decentralized charging control of electric vehicles in residential distribution networks,” IEEE Transactions on Control Systems Technology, vol. 27, no. 1, pp. 266–281, 2017.
 [13] Z. Ma, D. Callaway, and I. Hiskens, “Decentralized charging control of large populations of plugin electric vehicles,” IEEE Transactions on Control Systems Technology, vol. 21, no. 1, pp. 67–78, 2013.
 [14] K. ClementNyns, E. Haesen, and J. Driesen, “The impact of charging plugin hybrid electric vehicles on a residential distribution grid,” IEEE Transactions on Power Systems, vol. 25, no. 1, pp. 371–380, 2010.
 [15] J. QuirsTorts, L. Ochoa, S. Alnaser, and T. Butler, “Control of EV charging points for thermal and voltage management of LV network,” IEEE Transactions on Power Systems, vol. 31, no. 4, pp. 3028–3039, 2016.
 [16] J. Koshal, A. Nedi, and U. Shanbhag, “Multiuser optimization: Distributed algorithms and error analysis,” SIAM Journal on Optimization, vol. 21, no. 3, pp. 1046–1081, 2011.
 [17] M. Liu, P. K. Phanivong, and D. S. Callaway, “Electric vehicle charging control in residential distribution network: A decentralized eventdriven realization,” in Proceedings of IEEE Conference on Decision and Control, Melbourne, VIC, Australia, Dec. 1215 2017, pp. 214–219.
 [18] A. Hassen, Y. Dvorkin, D. Deka, and M. Cherkov, “Chanceconstrained ADMM approach for decentralized control of distributed energy resources,” in Proceedings of Power Systems Computation Conference, Jun. 1115 2018, pp. 1–7.
 [19] M. Baran and F. Wu, “Optimal sizing of capacitors placed on a radial distribution system,” IEEE Transactions on Power Delivery, vol. 4, no. 1, pp. 735–743, 1989.
 [20] M. Farivar, L. Chen, and S. Low, “Equilibrium and dynamics of local voltage control in distribution systems,” in Proceedings of IEEE Conference on Decision and Control, Florence, Italy, Dec. 1013 2013, pp. 4329–4334.
 [21] A. Prkopa, Stochastic Programming. Springer Science & Business Media, 2013, vol. 324.
 [22] Southern California Edison. (2011). [Online]. Available: https://www.sce.com/005_regul_info/eca/DOMSM11.DLP
 [23] M. Liu, P. K. Phanivong, and D. S. Callaway, “Customerand networkaware decentralized EV charging control,” in Proceedings of Power Systems Computation Conference, Dublin, Ireland, Jun. 1115 2018, pp. 1–7.